Seeking comment -

General issue is that a paid editor has come to the talk page of the Wikipedia article for their employer's organization, and I asked them to ask the communications department to develop Wikipedia articles in their field of expertise. To what extent are Wikipedia article talk pages suitable for such requests?

• Wikipedia:Conflict_of_interest/Noticeboard#Discussions_between_wiki_editors_and_corporations

Bluerasberry (talk) 16:10, 6 March 2023 (UTC)

You had a paid editor making edit requests which failed verification, so you asked them to develop other articles? That seems counterproductive, if they can't even provide correct references about their own organisation why would you ask them to write about other stuff? Fram (talk) 16:20, 6 March 2023 (UTC)

@Fram: It is one of best-funded organizations in the world. They have decades of expertise in humanitarian development. Their one paid editor may have stumbled, but they have excellent content holdings in humanitarian relief, child development in lower and middle income countries, and regional crisis. They have expertise and media which is high quality but unfortunately difficult to access; Wikipedia has very popular articles in their field which need quality content like theirs. To the extent that they have communication goals, the relationship seems mutually beneficial.

I would say the same for many other organizations, but this one seems like a fair place to start as their revenue is more than a $US billion a year, and for that reason, I think they can handle being asked in public. If we can sort out a process then perhaps we can present requests to smaller organizations like museums, universities, government agencies, and research institutes when they send their paid editors to wiki. Bluerasberry (talk) 17:03, 7 March 2023 (UTC)

I think that the COI editor talk page or one of the forums at WP:Village Pump might be better to hold conversations unrelated to the improvement of the article than the article talk page. For example, Wikipedia:Village pump (proposals) could be a good venue. Thinker78 (talk) 19:17, 8 March 2023 (UTC)

A couple of thoughts:

• If by "failed fact checking" you mean something that could comply with real-world fact checking but not fulfill wp:ver, I would not hold that against them. There's a big difference between the two.
• I don't think that article talk page discussions as normally implemented would not be so strict as to say it would improper to discuss it there. However, the question is where is the best / most appropriate place to have those discussions. If they are about related articles that relate to the article-subject organization then perhaps the article talk page is best. Otherwise the editor's talk page might be better.

North8000 (talk) 18:15, 6 March 2023 (UTC)

I marked failed verification because I felt the claim was "X does Y" when X did not appear in the cited source. It was not debatable. Bluerasberry (talk) 17:03, 7 March 2023 (UTC)

I didn't question any finding regarding compliance with WP:Ver. North8000 (talk) 19:58, 7 March 2023 (UTC)

I agree with North8000: if the conversation is going to be affect how others edit the article in question, then the article talk is best since all interested editors can be involved. If it's a conversation between you and the other editor on what else they can do (even if it's within the same field), the audience is different, and it would be better to discuss it somewhere else, such as the editor's talk page, or a related active WikiProject talk page.

Note there are potential conflict of interest issues with the communications department of a company editing pages in a related field. If Wikipedia's Animation page had a notice, "This page has been edited by the Walt Disney Company's communication department," even though they are indeed experts in the field, it could cast doubts in the minds of readers. isaacl (talk) 18:32, 6 March 2023 (UTC)

@Isaacl: I think all of this is new territory not previously discussed. If you see a place to take this discussion then I would post more wherever you suggest.

To take your example, I think there would be a difference in a public relations representative from Disney editing about the corporation as compared to their professional animators editing about their craft. I do not think any corporation has ever approached Wikipedia offering to edit content in their field of expertise. So far as I know, 100% of the 100s of 1000s of WP:COI issues are about branding and marketing, not about purported attempts to benevolently share expertise.

Some differences with the case I am presenting: World Vision is nonprofit, they are recognized as an expert organization on humanitarian issues, and they produce data, research, and media which is the basis of a lot of other research. I would be keen on getting their input on, for example, child development in India. I think it is fine to be wary, and perhaps we should develop rules, but development of that article seems less risky to me than their own development of the article about their corporation. Bluerasberry (talk) 16:20, 9 March 2023 (UTC)

Since you presented a general case without specifics, I discussed the matter in general. I have in the past discussed hypothetical scenarios where an organization such as a historical association might want to improve coverage in Wikipedia of history. We need to carefully consider, though, the effect of a paid-editing notice on the article in question with respect to readers trust in the article. In some cases, it may enhance trust; in others, it might be better for source material to be provided and for unassociated editors to use it to devise appropriate changes. I think the best approach will differ based on the specific circumstances. isaacl (talk) 17:25, 9 March 2023 (UTC)

@Isaacl: So far as I know, the notice you mention - {{COI}} - has no documented precedent for use in tagging articles with contributions from subject matter experts, regardless of their employers. Its use as reported in Special:WhatLinksHere/Template:COI shows almost all brands, not topics of general interest. Where it is on topics of general interest the problem often seems to be ideological bias, like a religious or social issue. Why regulate expert editing when the activity has not been tried? I am unaware of anyone ever identifying a problem with subject matter experts editing in their field of expertise. Why start from the presumption that it is comparable to marketing edits? Bluerasberry (talk) 20:03, 9 March 2023 (UTC)

If I understand correctly, you're asking an organization to assign someone as part of their work duties to create Wikipedia content. That makes them a paid editor and thus a paid-contribution disclosure is required. This is different than an individual choosing to edit a topic about which they are knowledgeable and is related to their employer. (There have been discussions in the past regarding potential conflicts for this second case. Previously I've commented that I think the community shouldn't treat this as always being an actual conflict of interest, precisely because I think this would hamper subject matter experts from contributing.) isaacl (talk) 02:02, 10 March 2023 (UTC)

@Isaacl: I think there are two ways of thinking about this

• paid versus not paid
• marketing versus general knowledge

In my view, the conflict of interest problems arise from marketing. I do not think we have a history of COI problems from subject matter experts editing on topics unrelated to a brand. I think a paid editing disclosure would be fine for user pages, because we have a policy for that, but I do not see a reason to preemptively ask for them on a wiki page when a subject matter expert, paid or not, edits in their field of expertise. If there is a problem then we could tag them, but among the 100,000+ COI cases, I do not think we have a single model case around experts editing general knowledge. Bluerasberry (talk) 12:12, 11 March 2023 (UTC)

A declaration by the paid editor on their user page meets the terms of use requirements. However this doesn't preclude additional disclosures being placed by others on the article talk page. The paid editor might also choose to disclose on the article talk page as well for transparency (and there is a ongoing discussion as part of the terms of use update about making disclosure mandatory in three locations instead of the current requirement for just one of them: the user's page, the article talk page, and the edit summary). My guess is that editor consensus on this matter is going to depend on the specific circumstances. isaacl (talk) 17:47, 11 March 2023 (UTC)

As an aside, I've done a fair amount of putting in or deferring declared-coi requested edits. IMO declared ones usually don't do much bad stuff.....they know they are under a magnifying glass and are pretty cautious. North8000 (talk) 18:46, 6 March 2023 (UTC)

It is my experience that many article raters routinely rate newly-created articles as "start." With a few exceptions, I don't consider articles I create as "start" and I am irritated by a superficial rating of an article.

According to Wikipedia:Content assessment a start article is "incomplete" and "most readers will need more." Plus, a start article "needs substantial improvement in content and organization." My most recent creation, Golden Triangle of Meat-packing was rated start. The article consists of 1,246 words (16,966 bytes), 29 references, numerous links, a map, and a photo. I didn't add an info box or a bibliography to the article as neither seemed necessary given the content. To my mind this article as now existing exceeds the "start" rating. It tells the average reader all they need or want to know about the subject and the content, drafting, organization, etc. is at least adequate. So, my question is: what do I do to remedy what I consider an error: complain to the rater, rate the article myself, or delete the rating and wait until another rater comes along?

I also believe it would be a good idea for an appropriate authority to remind article raters to refresh their memories as to what the rating guidelines say and tell them to undertake their chosen task of rating articles with more due diligence. Smallchief (talk) 12:28, 8 March 2023 (UTC)

When I start a new article or expand an existing article, if I create a talk page for it with project banners (I don't bother very often theses days), I always rate the article as "start", no matter how long or well sourced it is (I try to avoid creating stubs). I leave it up to other interested editors to decide if it rates higher. Frankly, I don't look at quality ratings, and make my own judgments on whether, and how, an article should be improved. Donald Albury 14:00, 8 March 2023 (UTC)

Rather than always rate the article as "start", no matter how long or well sourced it is, if you leave it unrated it will populate the relevant maintenance categories, which the relevant wikiprojects tend to keep an eye on, allowing someone else to quickly identify it as needing a rating. Just throwing in a start rating, on the other hand, categorises the article among all other start-class articles, meaning it's less likely to catch the eyes of wikiprojects. Ljleppan (talk) 06:53, 10 March 2023 (UTC)

I'm one of those who uses the Wikipedia 1.0 Server's list of unassessed articles to target so that I can review them, and so that wikiprojects have an overview of the progress and work remaining in their area. By putting in a rating without reviewing the article, you're giving a false statement to other editors, disrupting the work of those editors, and lowering the quality of that information. And for what? Are you using a tool that forces to enter a quality rating? If so, can you use a better tool? I appreciate that you personally don't look at quality ratings and for the most part they shouldn't be taken very seriously, but how do you get from that to it being okay to actively degrade that system for everyone else? MartinPoulter (talk) 20:44, 13 March 2023 (UTC)

I leave it blank, and have a feeling that C-class is the most common judgement bestowed upon me. A nice middle ranking perhaps. However, as Donald notes, the lower quality rankings are essentially meaningless, so I wouldn't pay them too much mind. All sorts of ranked articles are submitted to WP:GAN for example, following which they suddenly transition directly to GA-class. CMD (talk) 14:23, 8 March 2023 (UTC)

Yeah, I ignore the ratings on my articles. The rating system is pretty much broken. There's way too many possible ratings, the criteria are too vague, and ratings are applied haphazardly by inexperienced and/or unskilled editors. -- RoySmith (talk) 14:47, 8 March 2023 (UTC)

Thanks for the advice, guys. I'll guess I'll accept the "start" rating as trivial and not worth my time or yours. Smallchief (talk) 00:19, 9 March 2023 (UTC)

@Smallchief: It's actually been updated to C-class. Outside of some weird RfA opposes along the lines of has mostly created start-class articles, I've never seen anyone overly bothered over whether someone's creations are rated as stub, start, or C. In the future if you believe an article you created is a C then feel free to rate it as such on your own as I can't imagine anyone giving you trouble for it. For B-class it's probable better to make a friendly request for independent review of the article against the criteria on the user talk page of someone who's active in the area, or at an appropriate WikiProject if any are active, however there's no requirement to do so and I doubt anyone would be too bothered unless you made a habit of self-aggrandizing bad ratings.

Thanks for writing up a nice article. 74.73.224.126 (talk) 01:38, 9 March 2023 (UTC)

Thanks.Smallchief (talk) 01:47, 9 March 2023 (UTC)

Ratings except for GA, A, FA, and FL do not need to be granted by an uninvolved authority. Editors who know how to write decent articles have always been assumed to be capable of reading and following the directions so that they can assess their own articles. If someone gets it wrong, just fix it and don't give it another thought. If you want an "objective" reason, then the ORES ratings are available through User:Evad37/rater. It says that C-class is the most likely rating for your article on meat packing (with a substantial chance of either Start- or B-class being preferred – in other words, I'd choose C, but any of them could be right). WhatamIdoing (talk) 05:07, 9 March 2023 (UTC)

Yeah. Just rate your articles as C or B yourself, if you really believe they meet the criteria. (However, B needs to be assessed by a non-creator in the MilHist area, which is done in a routinized manner on a dedicated requests page). —Alalch E. 19:48, 13 March 2023 (UTC)

From the beginning of the 2022 Russian invasion of Ukraine, when the Russian troops occupied some Ukrainian territory, some users started to add the information on the occupation (in variuos ways) in the infobox. Other users remove this information. I have some of these localities on my watchlist, and these additions and removals are way too frequent. See just recently and just for Kreminna which is just one locality: one two three four. This bullshit needs to stop. I do not particularly care which way it would stop but we need a clear consensus that either mentioning or not mentioning occupation in infobox is ok, and the opposite is not ok.

In principle, this RfC should apply to any ongoing to conflict, but my primary interest is Ukraine. I would also exclude Crimea because it is really administrated by Russia, and the arguments could be slighly different, we can return to it another time. This is really about clear-cut cases. Ymblanter (talk) 23:32, 7 February 2023 (UTC)

Occupation can be mentioned in the infobox

Please suggest how it should be mentioned.--Ymblanter (talk) 23:33, 7 February 2023 (UTC)

We have existing long-standing examples for how such information can be inputted, eg. Katzrin, Laayoune, Afrin, Syria. CMD (talk) 04:28, 8 February 2023 (UTC)

Do the people who programmed hCard format into the infobox template know about this? It looks to me like the embedded data would get parsed out as “Afrin is divided into multiple ‘control,’ and one of the control of Afrin is ‘Turkey Hayat Tahrir al-Sham.’”  —Michael Z. 04:58, 8 February 2023 (UTC)

I posted a request for input at Template talk:Infobox settlement#Representing occupied territories.  —Michael Z. 05:02, 8 February 2023 (UTC)

It also looks to me like those examples are not consistent, judging by the three distinct grammatical schemes in the field labels of the three infoboxes. But I can’t really tell because they are hard to make sense of.  —Michael Z. 05:06, 8 February 2023 (UTC)

• Without looking for a blanket ruling, infoboxes should contain information about occupying/administering bodies. Infoboxes are meant to provide "key facts", the administration of any administrative entity is self-evidently a key fact. Pretty much all sources covering the occupied areas clearly and upfrontly state that which areas are occupied, because the occupation is a key fact (magnitudes more so than, say, Area code or Vehicle registration). Not including occupying administrations is a disservice to readers in the best case scenario. The invocation of WP:NOTNEWS is nonsensical; editing an infobox does not make an article original reporting, a news report, a who's who, or a celebrity gossip and diary. Infoboxes are updated all the time, completely routinely. An interpretation of Notnews suggesting they shouldn't would necessitate removing information like political leaders from infoboxes, which change quite frequently and are often updated almost immediately (or even before they should be!). NOTNEWS actually advises editors to "include current and up-to-date information". The arguments about stability are similarly flawed, and unsupported by policy and practice. Wikipedia articles are not frozen when real-life disputes are, and neither are their infoboxes. (Template:2022 Russian invasion of Ukraine infobox has information much more recent than much of the Russian occupations.) As for how, a number of models already exist, such as the examples given above. We could probably forever WP:BIKESHED between them, but what is important is that one is used to ensure infoboxes meet their purpose of providing key facts to readers at a glance. CMD (talk) 04:12, 16 February 2023 (UTC)

Occupation should not be mentioned in the infobox

Put your vote here if you think that if it is mentioned in the body, and this is enough.--Ymblanter (talk) 23:38, 7 February 2023 (UTC)

Not  The infobox is for defining characteristics of the subject, fundamental statistics, and other stable information. It is WP:NOTNEWS ticker, and should not include ephemeral and unpredictable wartime data like who has driven into a place this morning that could well leave it by this evening (and Russian sources for example, have been notorious for announcing the occupation of a locality several times over weeks without actually occupying it).  —Michael Z. 03:00, 8 February 2023 (UTC)

Another problem with this is that the infobox is not designed to accommodate foreign intervention. The fields for country subdivisions assume a legal hierarchy, and embeds a microformat encapsulating that hierarchy, which yields nonsense if the hierarchical data doesn’t make sense. I think it can also automatically generate short summaries which are nonsensical too, if there is no wikidata description or manually entered short summary.  —Michael Z. 03:17, 8 February 2023 (UTC)

Not. (Summoned by bot) The infobox should only be used for stable information; it should summarise the information contained in the body; and should represent the WP:CONSENSUS among the RS as well as the involved editors. Trying to captture information that may go back and forth in the course of a conflict should be avoided in the infobox. -- Kautilya3 (talk) 18:39, 9 February 2023 (UTC)

• Not: (Summoned by bot). Not only is the above WP:NOTNEWS ticker true (totally agree with Mzajac and Kautilya3) because this is something that can change at the drop of a hat so there would be no infobox stabilization as long as there is conflict. -- Otr500 (talk) 07:35, 14 February 2023 (UTC)
• Not Gaza Strip and West Bank are "stable" occupations, we do not say so in the infobox (nor for subsidiary units of those areas). Selfstudier (talk) 10:45, 14 February 2023 (UTC)
• Not The details about whether an area is occupied are always going to be controversial. Infobox don't allow room to put the details of the occupation into a proper context, and so shouldn't general be included. -- LCU ActivelyDisinterested ∆transmissions∆ °co-ords° 21:32, 23 February 2023 (UTC)

Discussion (occupation in infobox)

This is not a new thing. The infobox for the city of Donetsk was “occupied” by a single-edit anon over six years ago at the latest, for example.[3] There is a clear long-term consensus to remove these sporadic interventions, as we can see by the current state of Donetsk. —Michael Z. 03:05, 8 February 2023 (UTC)

This is good, but let us complete this RfC to make the consensus explicit. Ymblanter (talk) 03:28, 8 February 2023 (UTC)

No argument from me.  —Michael Z. 03:51, 8 February 2023 (UTC)

• So, I think there is some nuance to the matter, and I don't want to vote definitively either way. I think that this is a case-by-case basis; which doesn't mean I don't think there need to be some guiding principles, but I DON'T think a one-size-fits-all approach is the way to go here. For some kinds of occupied territory, which is under a sort of long-term dispute, where one country claims as its territory an integral part, but which another has been administering in a de facto way for a long period of time, it may be useful to describe it as such. The Golan Heights for example is widely recognized as Syrian territory that is under Israeli occupation (without wading into the debate over whether that is true or not, just that it is commonly described as such); but that's still a stable state. The two countries are not actively fighting over the land, it's basically been that way for decades, and it isn't an ephemeral thing. It's basically a standard state of affairs. If we have an active warzone, where two armies are fighting back and forth to control some bit of territory, then no, we should not be updating the infobox every few days to reflect the changing battle lines. The infobox is poorly suited for that sort of thing. Leave it out in those cases. --Jayron32 17:26, 9 February 2023 (UTC)

  This makes sense to me. It's impractical to update infoboxes to reflect shifting battle lines in a hot conflict, but when it's a long-established dispute and WP:RS describe the situation as occupied/disputed territory, then it would make sense to include in the infobox. How that's handled with the hCard format is something that would need some thought and attention to handle the sorts issues Michael Z pointed out. —Carter (Tcr25) (talk) 17:49, 9 February 2023 (UTC)

  If we are to document stable occupations or land claims in the infobox, it would probably make sense to agree what it should look like, how it is to be used consistently, and then add a separate row for this to the infobox template.

  None of the example solutions are clear to me, and I can’t imagine they could possibly be to someone new to the subject and glancing at the infobox. They all fail basic communications principles and do not serve WP:INFOBOXPURPOSE IMO.  —Michael Z. 18:28, 9 February 2023 (UTC)

  Look at Navassa Island. It an "easier" case since it's not inhabited and can be easily labeled a "disputed island," but it's still using subdivision_type3 and subdivision_type4 for Haiti's claim, which is problematic. {{Infobox island}} has a disputed option that might be a way to handle things, but I'm not sure that it has an effect on the hCard data. —Carter (Tcr25) (talk) Carter (Tcr25) (talk) 23:04, 9 February 2023 (UTC)

  Regarding hCard, I’m not much concerned that occupied status is absent from it, but we shouldn’t abuse the template in such a way that the hCard data is bad. This is important because the effect is not visible, so most editors will have no way of knowing, and it is meant for re-use so we have no way to predict what kind of applications of someone else’s we will be breaking, or in what way.

  So there should probably be zero tolerance for shoehorning stuff like this into those fields. If that can’t be accomplished, then fields specifically for this must be added to make sure it’s done right. Anyway, probably a discussion for the template’s talk.

  At least equally important, of course, is that the infobox is perfectly clear to readers. Even in that simple example, there’s no visual hierarchy (I guess there are two hierarchies with no divider), and the headings are different kinds of things. It’s effectively nonsense that someone has to guess at what it means. It would be easier to decider a long sentence, but that doesn’t belong in an infobox either.  —Michael Z. 02:07, 10 February 2023 (UTC)

I recently had the chance to see some controversies develop in the talk pages of a few political parties of my country, where some users complain that the articles are heavily patrolled by sympathizers if not outright party members, in effect turning what's supposed to factual encyclopedic articles into propaganda pages since nothing that might be even remotely inconvenient usually stays for very long no matter how factual or how well sourced it is. And suggestions like "just discuss with them bro!", strike me as completely inadequate.

I don't know what debate if any there is around this issue, but increasing the minimum requirements to edit on political parties might limit the flow of users whose sole reason to edit is to peddle their ideologies and nothing else, with no regards to facts, even if it doesn't solve it completely. Wareno (talk) 19:32, 14 March 2023 (UTC)

Well, we've got our Five pillars; it seems to me that especially pillars 2, 3, and 4 already cover this, supported by WP:V, WP:RS and WP:NPOV. You haven't said how you suggest we "restrict the editing", so I won't throw WP:NO-PREEMPT at you, too (I've already acronymized you to a bloody pulp anyway—sorry!). — JohnFromPinckney (talk / edits) 20:53, 14 March 2023 (UTC)

@Wareno there exists a list of contentious topics already. Elections in USA would already be covered by WP:ARBAP2. I haven't fully fleshed out my opinion on this, that said. ~ 🦝 Shushugah (he/him • talk) 21:08, 14 March 2023 (UTC)

Policy already addresses this sort of thing extensively. WP:Dispute Resolution goes into detail about your options. If you've made a serious attempt to discuss an issue on the article's talk page and no consensus develops, then you can make a post to the neutral point of view noticeboard asking for additional input or you can seek resolution at the dispute resolution noticeboard. If there's a specific editor that's attempting to control everything in an article or consistently challenges anything they disagree with politically, then you can make a post at WP:ANI. But this is a last resort, so make sure you can prove that they're being disruptive before you do this. Thebiguglyalien (talk) 04:03, 15 March 2023 (UTC)

When it comes articles (about individual people) in the English Wikipedia, should one make the article name from the Russian transliteration or given minority language version? Especially with Tatar names, since there also exists a Latin Alphabet version of their language, albeit in limited use. I can think of arguments for both; the version in the minority language is sort of more "authentic since that is the native language of given people, but at the same time, the Russian version is the official version that is also in the passports. It also usually tends to be more approachable to English speakers. Xәkim (talk) 15:02, 16 March 2023 (UTC)

There is no “one-size-fits-all” rule for this. Per WP:COMMONNAME the article title should use whatever name is most commonly used in English language sources that discuss the subject (as that will be the most recognizable to our readers).

In the article text, we have a bit more flexibility. We would normally default to the COMMONNAME as well… however, it is appropriate to (briefly) mention any other variants. This is often done as a parenthetical in the opening sentences or paragraph. Blueboar (talk) 15:33, 16 March 2023 (UTC)

To expand on the above, while it's a case-by-case assessment, variants should typically be included if there is a reasonable chance that an English-speaking reader could encounter the variants somewhere off-Wikipedia, but we should not be inventing plausible variants if we do not previously have evidence of them being used. _signed, Rosguill ^talk 15:43, 16 March 2023 (UTC)

ChatGPT has gone viral and is causing quite a storm on the Web. Why should we care? Because, rather than just providing links to an article for you to read about what you are interested in, it writes an essay about it on the spot. It is so sophisticated that, if you ask it to write an article on any subject, even in the style of Wikipedia, it will! That can then be copied and pasted into Wikipedia -- not necessarily a problem if it is carefully fact checked and edited first (ChatGPT can't tell fact from fiction), but, an editor may post it without doing so. It can even be used to answer questions about Wikipedia as if it is a Wikipedian (see the first link provided directly below).    — The Transhumanist   06:52, 9 December 2022 (UTC)

1. Experiment: See it being used to generate answers to hypothetical questions on Wikipedia:Teahouse
2. ChatGPT's Writing Capabilities Stun, but Humans Are Still Essential (for Now) - CNET
3. Five Chats to Help You Understand ChatGPT - The Atlantic
4. The ChatGPT chatbot is blowing people away with its writing skills. An expert explains why it's so impressive
5. ChatGPT: Why Everyone's Obsessed With This Mind-Blowing AI Chatbot - CNET
6. ChatGPT | Discover AI use cases
7. 'Scary' ChatGPT could render Google obsolete in two years
8. ChatGPT, an AI chatbot, has gone viral. Some say it’s better than Google; others worry it's problematic.
9. What ChatGPT AI is and how you can use it - The Washington Post
10. What is ChatGPT and why does it matter? Here's what you need to know | ZDNET

Based on how rapidly chatbots have improved over time, it will become more and more difficult to tell if an article was written by a computer or not. The sheer volume at which computer programs could create new accounts and produce Wikipedia content, and the inevitable growing number of human editors copying and pasting chatbot output into Wikipedia, will at some point make it impossible for Wikipedia's human volunteers to keep up with that traffic and apply quality control to the material in a reasonable time frame -- the backlog of unchecked material will simply get longer and longer. The only recourse will be for computer programs to do it -- either computer programs to process articles to filter out or correct any crap, or training the chatbots themselves not to produce crap in the first place. Rather than build computer algorithms to detect computer-written articles and passages, it would be more productive for them to do style checks, fact checks, and citation checks, along with appropriate corrections or removals. While Wikpedia-friendly AI could come from within Wikipedia, it may be faster to bring influence to bear upon the developers of the chatbots being used to generate Wikipedia content, and upon the chatbots themselves. Wikipedia already has a chair at the table, because Wikipedia comprises a significant component of chatbot corpi, and so, their developers should be inclined to listen to the Wikipedia community's concerns -- either directly, or indirectly through news coverage. The Wikipedia community should make its voice heard on the matter of chatbots writing Wikipedia material according to Wikipedia's style and behavior guidelines. For example, verifiability still applies, and so when chatbots are asked by their users to "write an article in the style of Wikipedia" the chatbots should comply according to Wikipedia's policies, including those on verifiability and providing reliable sources. Not doing so should be met with the filing of bug reports, feedback, and commentary. And, as chatbots learn as they go, Wikipedians who use them can ask them to follow Wikipedia guidelines, and we can urge our fellow editors to request this of chatbots as well.    — The Transhumanist   06:52, 9 December 2022 (UTC)

Chatbots should be following Wikipedia's lead for all of their output. At this time, most chatbot answers and essays are not referenced with reliable sources. And they should be, for the same reason that Wikipedia articles should be. That's something that can be requested of chatbots directly, through queries, and of developers, through their contact channels and social media. I hope this suggestion helps.    — The Transhumanist   06:52, 9 December 2022 (UTC)

@Transhumanist If a few chatbots are reading major newspapers, Google News, and history books, and watching TV news programs, they can be the ones writing drafts. After we have chatbots writing drafts (new chatbots should use the AfC process, of course) we'll need chatbot software to review those articles at AfC; handle NPP; take care of dispute resolution between different chatbots; open cases and then resolve them at AN and AN/I; etc. Then Wikipedia will be perfect! Unless it goes into some sort of loop that creates a black hole... David10244 (talk) 11:17, 20 March 2023 (UTC)

The simple answer is that our existing policies ought to already cover this (mostly.) Sourcing is still required for anything that is challenged or likely to be challenged, which prevents people from just blindly dumping AI generated text into Wikipedia; and an AI may violate copyright depending on how it was trained (and whether it was overtrained.) There are also unsettled copyright concerns related to AI training sets, so I would generally think that, ideally, editors shouldn't be dumping AI generated text into our articles even after performing due diligence to make sure it's not a copyvio and finding proper sources. But since those concerns are unsettled and speculative, I also don't think it's worth worrying about too much right now. The key point is that we should emphasize our sourcing requirements and be more diligent for clear-cut copyvios, which we already have systems in place to handle, since it is likely that these tools will result in people adding lots of unsourced and possibly-copyright-violating text. (I do wish our RFCs on mass article creation had reached a stronger agreement on sourcing requirements for new articles, which would deter excessive copy-pastes of AI generated text - perhaps that is something we might want to revisit in the near future, if we start seeing significant amounts of new unsourced articles created using what is plainly AI-generated text.) --Aquillion (talk) 07:55, 9 December 2022 (UTC)

You mean, don't prepare in advance for a potential increase in volume, just wait until it hits? At that time, will merely adjusting policies stem the tide? It's in the slow trickle phase now, but that could potentially become a torrential flood very rapidly, just as ChatGPT's user base grew to over a million in 5 days. My main concern above was about a potential volume of AI-generated content that went beyond the scale of what the editor community could manually process. You didn't address that contingency. What could the community do to prepare for it, just in case it does happen? What are the available options?    — The Transhumanist   11:28, 9 December 2022 (UTC)

I don't think there's much we reasonably can do to prepare, at least not without serious risk of causing other problems; AI-generated text won't be drastically different than other sorts of text, aside from the risk of being uncited or a copyvio (which we have existing processes in place to handle.) It's worth raising awareness of the issue so editors can spot the signs of someone using large amounts of it, but I think our best bet if we're going to "prepare" is to focus on the systems we already have, which is unlikely to do any harm either way, or perhaps to codify slightly more strict sourcing requirements in the way I described (which I think is a good thing anyway, but would at least serve to slow down the worst sorts of misuses of AI generated text.) Ultimately the most serious problems are if editors start adding large amounts of text that violates copyright or which are uncited and likely to be challenged, but we have existing procedures for those, we just need to prepare for the possibility that we may need to become a bit more aggressive about enforcing them. Wikipedia is in a slightly better position than some other websites facing AI-generated-text problems, because our sourcing requirements will at least make it fairly obvious if someone tries to dump large amounts of AI-generated text onto the wiki without making any effort to verify it. --Aquillion (talk) 12:47, 9 December 2022 (UTC)

I suppose we could take the Stack Exchange approach and just say flatly "no, this isn't allowed" - in their case it is explicitly a temporary measure until we have a better understanding of the issues. I think in general our policies/community norms would come down hard on anyone trying to get a language model to generate articles (hard to see why that would be OK and machine-translation isn't), but maybe an explicit statement would be a way to go. Andrew Gray (talk) 18:32, 9 December 2022 (UTC)

@Aquillion: While a large number of posts by individual editors may become a problem, the main concern I presented above was "the inevitable growing number of human editors copying and pasting chatbot output into Wikipedia, will at some point make it impossible for Wikipedia's human volunteers to keep up with that traffic and apply quality control to the material in a reasonable time frame -- the backlog of unchecked material will simply get longer and longer."

That is, people making the normal rate of content contributions, but using large language models (chatbots) to do so.

Watching for breakout editors who use LLMs to create a large number of new articles over a short period of time would not suffice in such a scenario. Editors who add LLM-generated content to many existing articles also will not be spotted by looking for mass page creations. And since writing will become easier by letting "chatbots" do it for you, content submissions by users employing such tools may likely become longer on average.

The point is, that a high enough volume of such content contributions would go beyond the capacity of Wikipedia's editors to check and correct.

The two solutions offered were 1) build software to analyze and process such content, and 2) work with chatbot developers so that inappropriate content is not composed by LLMs in the first place.

Just relying on new or existing policies to handle LLM-generated content will be insufficient if and when the volume of it passes the threshhold of what manual editors applying Wikipedia policy can deal with.

Passing that threshhold may come soon, or it may take years -- the main question is "will Wikipedia prepare for that threshhold-passing event?" Based on the responses above and below, the answer, and implicit recommendation from this forum, currently appears to be "no": No developing relevant software, and no working with chatbot developers to respond to the potential passing of the LLM-threshhold.

Thus, any solution will need to come from other departments or from continued or future discussion in this department, or from chatbot developers focusing on the problem due to other influences.

Another helpful approach might be the creation of a policy or instructions on how to use LLMs/chatbots effectively, and post links to that page in enough places that all editors will notice. Though, I doubt that would prevent the problems of a LLM-threshhold-passing-event, and wouldn't address the need for proofreading or processing LLM-generated contributions.    — The Transhumanist   02:18, 20 December 2022 (UTC)

I think in the end, @Aquillion response from earlier in the discussion is the one I agree with the most. If a ChatGPT-written or AI-written part of Wikipedia is accurate, nice! If something can't be proven with an RS, though, revert and warn. It can be a useful tool. Overall, think "Business as usual", just with some contributors getting assistance. It's mostly implausible as well to tell apart ChatGPT/AI generated content as well except in obvious or disclosed cases, and it's going to be impossible to effectively enforce as well no matter what position we take. If there is any legal issue with it, let WMF and ArbCom handle it. InvadingInvader (userpage, talk) 07:37, 6 February 2023 (UTC)

@InvadingInvader: I agree that a complete strategy of response to LLMs is beyond the scope of the Village Pump (policy), because it is going to take bots and/or diplomacy with developers, to stem any potential tidal wave of LLM contributions that would go beyond the capacity of WP's human editors to deal with. And those are the province of the bot department and the WMF, respectively. Considering that ChatGPT has gone viral, reaching 100 million users in just 2 months, eliciting a frenzied reaction from Google to bring rival products to market, thereby starting an "AI arms race", a tidal wave of chatbot-generated content in Wikipedia is plausible if not inevitable. Brace for impact.    — The Transhumanist   11:18, 13 February 2023 (UTC)

We must consider if the No Original Research policy is violated by AI created content.

It is not unreferenced research, but it is synthesized knowledge. We must determine if synthesized knowledge is original research, and I propose we answer the question through a trial with editors and not AI, where we synthesize knowledge for articles

However, WP::NOR addresses synthesized knowledge. The AI in question synthesizes knowledge, so it is not permissible to use the product of ChatGPT for articles. We must reconsider the policy on synthesized knowledge first, and either maintain the current policy, or change the policy to allow synthesized knowledge and develop a methodology to synthesize knowledge with consistent attributes, such as be referenced, and the conclusion not contradicted by any mainstream sources. Smedskjaer (talk) 11:33, 24 March 2023 (UTC)

I disagree on two grounds: (i) the generative AI is simply a tool, indeed, I do not see a clear boundary between it and, say, the translation tools or good spell checkers (all of these actually can - and some do - share the same corpus of texts) (ii) generative AI rehashes the said corpus, there is no new "knowledge" to be synthesized. IMHO, the fact that the results of the AI work in some areas superficially resemble research works is mostly due to very sad state of the "research" in these fields. The only thing the bot actually knows is language. Try asking the hard questions in the hard sciences, and the output will either (appropriately) consist of pure random babble or will repeat sources verbatim. The gainful uses of the current generation of the generative AI are in summarizing or paraphrasing the text, which is a typical and legitimate part of the editor's work here. Dimawik (talk) 15:20, 24 March 2023 (UTC)

I disagree, on the grounds art generating AI creates new work.

But instead of protesting or supporting either side, I would like to suggest a trial case, where AI is used on articles that either are stubs, or do not exist, yet have some subject matter importance. We can examine the community's response to the application of AI instead of theorhetical debates. Smedskjaer (talk) 15:10, 27 March 2023 (UTC)

This is what I am doing with the Bailey FFT algorithm. You can check the work progress at User:Dimawik/Bailey FFT algorithm (the original text was Bard-generated, the rest is my own work). Dimawik (talk) 15:55, 28 March 2023 (UTC)

@Smedskjaer, according to Wikipedia:No original research, original research means material—such as facts, allegations, and ideas—for which no reliable, published sources exist. If a reliable source that contains this same information has been published and still exists—somewhere in the world, in any language, whether or not it is reachable online—even if no source is currently cited in the article, then that information is not original research according to our policy.

In other words, if the AI produces a paragraph that contains only material that could be cited to a proper source, then it is not a violation of WP:NOR. It does not matter what tools you use to get a sentence like "Tobacco smoking raises the risk of lung cancer". That sentence cannot be original research, because it is material for which many published reliable sources could be found. WhatamIdoing (talk) 22:56, 28 March 2023 (UTC)

I agree with your foundation, but your argument doesn't address my argument. If from sources that could be cited, a human editor were to argument a new idea, or use available data to argue for a conclusion or theory's validity, then the editor's work is synthesized knowledge. AI can be limited to material that could be cited, yes, but that is useful if we are talking about the criteria an AI must fulfill to be allowed to edit wikipedia. AI for the purpose of this debate is inclusive, so it isn't possible to distinguish between AI which only uses material from cited sources and AI which can create an argument from cited sources.

AI can create new work, and that is the issue. AI that can write articles are getting better, and it is an eventuality new principle/foundational knowledge will be synthesized by AI, from the corpus of published work we already have. AI already creates new art work, which is indistinguishable from human created artwork, in form and methodology; AI has a set of actions, and the neural-net is trained by past work, just like human artists. It is synthesized art. Smedskjaer (talk) 07:41, 29 March 2023 (UTC)

If you use an LLM or similar tool to write text, you have to read the text before your post it to Wikipedia. If it says something that is verifiable (nothing that is truly verifiable can be original research), such as "Tobacco can cause lung cancer", then you can post it. If it says something that is not verifiable, or that you don't know whether it's verifiable, then you shouldn't post it.

This does not seem very complicated to me: The humans who post material needs to be responsible for the material they're posting. WhatamIdoing (talk) 16:37, 30 March 2023 (UTC)

Agreed. Unless and until chatbots are capable of accessing and editing Wikipedia without any human intervention (which I think is still a bit down the road), then the humans who actually add chatbot-generated content to Wikipedia are solely responsible for it. Donald Albury 19:01, 30 March 2023 (UTC)

What can chatbots do?

You seem to be somewhat panicking over a scenario which isn't really supported by any evidence. While I see some Teahouse responses, could you give us one or two examples of " It is so sophisticated that, if you ask it to write an article on any subject, even in the style of Wikipedia, it will! " articles? The teahouse examples give the impression that, if it ever becomes a problem, some edit filters can easily spot these. You would in any case need "someone" to post this "potential volume of AI-generated content that went beyond the scale of what the editor community could manually process" you predict. This seems rather unlikely, at least on enwiki. Fram (talk) 11:45, 9 December 2022 (UTC)

Just try it. If your eyes don't pop out, I'll be surprised. Right now, during its "research preview", it is free. Keep in mind that it draws heavily on Wikipedia, which is included in its corpus, so, for this test run, it would be best to choose a person or subject that is not yet covered in this encyclopedia, and ask ChatGPT to write about that.    — The Transhumanist   14:41, 9 December 2022 (UTC)

Pinging @Fram, JPxG, EpicPupper, and 0xDeadbeef:    — The Transhumanist   14:50, 9 December 2022 (UTC)

I'm not giving out my phone number to some random website, thanks. Why it isn't sufficient that they have my email which was then verified is not clear... Fram (talk) 14:55, 9 December 2022 (UTC)

LOL I had the same exact response. My phone number? F no. Levivich (talk) 16:55, 9 December 2022 (UTC)

I went to try this out, and it asked me for my phone number. I thought about making one up like 0118 999 881 99 9119 725 3, but figured it would probably use it for two factor authentication, so that's no good. Ritchie333 ^{(talk) (cont)} 18:31, 9 December 2022 (UTC)

@Fram, Levivich, and Ritchie333:     I wasn't that bright. I gave it a random phone number. It rejected it as a land line. Then I gave it another, and it rejected that as a VOIP number. Finally, I gave it a random mobile phone number, and it sent some complete stranger the verification code. Oops.    — The Transhumanist   01:32, 11 December 2022 (UTC)

Thanks for the ping. I'd imagine the biggest problem would be people using the AI to create hoaxes. Like the Zhemao hoaxes but with less effort. 0xDeadbeef→∞ (talk to me) 15:13, 9 December 2022 (UTC)

So. I tried it yesterday. I'm not sure how heavily it draws on Wikipedia's corpus for its knowledge.

First, I asked it to tell me about Hammerton Killick. I know there is a Wikipedia article about Hammerton Killick, because I wrote 90% of it. It did not know who Hammerton Killick was, and informed me that it does not have access to the internet, or to Wikipedia.

Next, I asked it to write me an article in the style of Wikipedia. I did not specify a subject. It wrote about Athens. The result was ok. Heavily focused on the ancient city and on art and architecture. Short. Kind of read like an encyclopedia article.

Next, I asked it to write me an article about alcoholism in the style of Wikipedia. The result was very interesting. I did not think it read like a Wikipedia article, it was more like a brochure that would be distributed in a doctor's office or something. I asked it what about that essay it thought was like Wikipedia, and it said what it wrote was

• neutral
• factual
• organized

Next, for fun, I asked it if it could write a recipe. It proceeded to give me a recipe for chocolate chip cookies. It looked like it should work. I e-mailed it to myself, and today I made them, not expecting much. I was pleasantly surprised. They were delicious. The only problems with what it wrote was that it did not have me cook them long enough (it said to bake for 8-10 minutes, and it took closer to 13 minutes for them to be done), and it drastically underestimated how many cookies the recipe should make (it said I'd get 2 dozen cookies, and I ended up with 5 dozen). I was shocked that it actually was edible.

I asked it to write a legal motion asking the court for an in-person hearing. I did not give it any other details. For not having any other details, the result was not bad. Westlaw has started offering a service that I think might draw on this type of technology, it helps you write pleadings.

Last I asked it to write a 100 word short story about a mouse, genera: fantasy. The result was decent. If I came up with it on my own I wouldn't be ashamed to enter it into a contest like the ones NYC Midnight runs.

I was more impressed with the recipe and the short story than the Wikipedia style articles. I can see some use for it in, say, copyediting as JPxG did below; or asking it for suggestions on language rephrase if you are trying to reach a word limit. I think it could have its uses. But I do think the Wikipedia community should be looking to craft policies and guidelines around what is and is not acceptable use of such tools. ~ ONUnicorn^{(Talk|Contribs)}problem solving 06:26, 11 December 2022 (UTC)

@ONUnicorn, Fram, Levivich, Ritchie333, 0xDeadbeef, JPxG, and EpicPupper: Interesting. The chatbot sent you instructions (in this case, a recipe), and you followed them. You followed the commands of a computer. If it gave you an address and instructed you to go there and pick up a brown paper package, would you? The implications of this type of interaction are huge and foreboding. This issue must have a name, and I would like to look it up, but I can't seem to find it. Though, when I typed in "computers in charge" I got the following 2 relevant results:

• Do we control our phones or do they control us? -Psychology Today
• 'I know your favourite drink': Chinese smart city to put AI in charge -Reuters

Then I typed in "computers telling people what to do", it came up with this:

• When Computers Tell People What to Do: A Work Practice Simulation of the Überlingen Mid-Air Collision -Print Replica

Ouch. I imagine, that anytime you ask a chatbot/computer "How do you do such and such?" it will reply with a set of instructions. And the chatbot's disclaimer in its terms of service will read "follow any instructions provided at your own risk". If you know or come across the name of the topic that covers computers telling humans what to do, please let me know what it is.    — The Transhumanist   11:04, 2 January 2023 (UTC)

@The Transhumanist: I think the term you're looking for is automation bias – "the propensity for humans to favor suggestions from automated decision-making systems and to ignore contradictory information made without automation, even if it is correct."

Interestingly, though, the 2002 Überlingen mid-air collision you mention is an instance where the computer got it right. An aircraft was on a collision course, and its crew were receiving contradictory instructions; the onboard collision-avoidance system was telling them to climb, while the human air traffic controller was telling them to descend. The pilots decided to trust the human and ignore the computer. Seventy-one deaths. Sojourner in the earth (talk) 14:00, 2 January 2023 (UTC)

@Sojourner: I like it: automation bias, like trusting GPS over a passenger's navigation suggestions or requests. Related, but not the term I'm looking for: which is the most used name for the subject of computers ordering humans around. When a computer gives you a command or a set of steps to follow, what is that called?

@Sojourner: Thanks for the comments on the collision. It was a bit more complex of a system error than that. While the air collision alerts on the 2 aircraft worked, there was a problem on the ground. The system the air controller was using would normally alert the user of an impending air collision in time for him to do something about it, meaning "no news is good news" -- not seeing the alert on screen, the air traffic controller (ATC) would naturally assume everything was okay -- but the ground-based optical collision warning system was turned off, creating a deadly ambiguity he was unaware of. So, there was some unintended computer-related miscommunication going on. "No alert" didn't mean what he thought it meant. My guess is that pilots would normally believe that the ATCs have better (radar-based) control systems than those onboard, and this probably affected the decision of one of the pilots to go with the ATC's order over that of his onboard computer -- he wouldn't have known that the ground-based system was disabled, either. Hopefully they fixed that bug, 'cuz it was a doozy. The poor ATC got hunted down over it, while the hunter became a national hero. Crazy world, about to get even crazier.    — The Transhumanist   15:13, 2 January 2023 (UTC)

When it comes to aviation, the more specific common term seems to be “automation dependency”. However I’m not aware of a term for what you’re looking for. DFlhb (talk) 23:47, 15 February 2023 (UTC)

Hi Fram. The Teahouse responses that you have seen were probably created by me. You probably noticed that the pattern "regards, Teahouse" in those responses. ChatGPT's strength is writing about technology, so I have told it to generate a section about Rust for Linux.

It does have the encyclopedic tone to me. 0xDeadbeef→∞ (talk to me) 15:02, 9 December 2022 (UTC)

In addition, it can be easily used to create fake references that would be hard to verify. For example, my prompt "Please output the Wikitext markup for the book reference with page numbers for the third paragraph, referencing the book Linux kernel development" resulted in this.^[1] 0xDeadbeef→∞ (talk to me) 15:08, 9 December 2022 (UTC)

@0xDeadbeef: Could ChatGPT's generated text or fake references be easily spotted by edit filters? What about spotting the output of future chatbots, like GPT-4?    — The Transhumanist   15:23, 9 December 2022 (UTC)

Thanks, OxDeadbeef. In this case, it would be relatively easy to spot the issues if it hadn't any refs, or with the added ref which predates the Rust for Linux thing by years; but of course it won't always be that easy. Fram (talk) 15:27, 9 December 2022 (UTC)

It has an encyclopedic tone because it's just regurgitating the Wikipedia article. Are there any examples for topics that we don't already have article about, where Wikipedia is not the source? Levivich (talk) 17:33, 9 December 2022 (UTC)

Similar was discussed previously in the section/item "Galactica and RS".

As was stated above by Aquillion, there is no qualitative difference in the treatment of human vs. non-human generated content. The same policies should apply to both. The problem seems to be the hypothesized/expected future mass creation of articles by non-human contributors. This appears to be a problem now, involving human contributors. Recent RFCs about the issue sponsored by ArbCom have accomplished nothing. Until a consistent restrictive policy relating to mass article creation (by any type of contributor) is accepted, this issue is moot imo.

Considering Wikipedia's limited resources, the policy would necessarily be restrictive, hopefully focusing on quality vs. quantity. Again, almost all restrictions proposed in the ArbCom-sponsored RFCs were rejected. This may be an indicator of how well such a policy will be received. 50.75.226.250 (talk) 15:43, 9 December 2022 (UTC)

Thanks for the policy politics clarification. The increase in the rate of content creation could have multiple aspects, for example, the number of articles created per user, and increased length of articles. The main feature of ChatGPT is that it is fast -- much faster than a human article writer. Its successors will be even faster. Users could use ChatGPT, and its successors (and their competitors), to be prolific, without triggering the mass page creation rule: if editors each used it to write an article per day, maybe even two, or up to four or five stubs

In the hands of responsible editors, ChatGPT would be a great productivity booster. Since August of 2022, JPxG and EpicPupper, editors of Wikipedia's Signpost news department, have been using GPT-3, the predecessor of ChatGPT, to write (or assist in writing) entire sections of the Signpost, as a demonstration of its capabilities, and as a platform to explore the potential and limitations of large language models. See From the editors: Rise of the machines, or something.

But, in the hands of inexperienced editors or bad actors, we could be faced with a big garbage in, garbage out scenario.

Bad actors aside, good faith use of chatbots could be improved by the creation of excellent instructions on how to apply specific chatbots to the development of articles, that is, how to best instruct them for the best output, and by working with the developers of the chatbots on what is needed for those bots to produce material compatible for use on Wikipedia, complete with accurate citations. As time goes on, Chatbots should get better and better at following Wikipedia's guidelines. But this may require supervision (er, input) from the Wikipedia community.    — The Transhumanist   16:52, 9 December 2022 (UTC)

It is suspiciously positive, in a casual (pseudo-)factual manner. It would raise a red flag afac, regardless of its provenance. 65.88.88.93 (talk) 19:13, 9 December 2022 (UTC)

Chatbot progress, erroneous output, and confident nonsense

All of this may become moot in a few years, due to accelerating change affecting chatbot development, making them more proficient at encyclopedia writing than humans -- Ray Kurzweil predicted that computers would reach human level intelligence by 2029, and based on how advances have been going, maybe even sooner. Chatbots may become good enough to leapfrog human editors even before that. At that point, they could be used to fork Wikipedia and expand upon it at a rate magnitudes faster than humans could. Or be used to start a new encyclopedic website from scratch. Rather than have article pages, the work could be a list of topics, and each time one was clicked on, a chatbot would create an up-to-date treatment of that title, from scratch on the spot. In the style of Kurzweil, if we are 1% of the way there, and the technology doubles every year, we are only 7 years from reaching 100%. But, it looks like we may already be more than 1% of the way there.    — The Transhumanist   17:18, 9 December 2022 (UTC)

I ran a few things on people who don't have articles, (they were all pool players in fact). The chatbot just made up that they played football for Legia Warsaw. It's a bit worrying that it looks like decent wording, but I don't see how that effects us. Lee Vilenski ^{(talk • contribs)} 17:41, 9 December 2022 (UTC)

{ping|Lee Vilenski} See the AI generated text in the box above.    — The Transhumanist   18:34, 9 December 2022 (UTC)

@The Transhumanist: I'm not trying this (for reasons given above) but I'd be interested what the bot would write for See What a Fool I've Been. We used to have an article, then it was turned into a redirect at AfD, so everyone can do a straightforward A-B test. Ritchie333 ^{(talk) (cont)} 18:40, 9 December 2022 (UTC)

Here's what it gave me Ritchie333:

I don't know if this would be something that would have been retained from previous articles through mirrors or if this is genuinely the bots own work. Lee Vilenski ^{(talk • contribs)} 19:22, 9 December 2022 (UTC)

Interesting. It wasn't on A Night at the Opera, and it's not a ballad, and May doesn't sing on it. Those lyrics aren't right. I don't know if it's been well received, and it's a pretty obscure track for a major band. It hasn't been played live very much at all. So half of that prose is factually incorrect, and there are no sources. Ritchie333 ^{(talk) (cont)} 19:39, 9 December 2022 (UTC)

See also: https://perplexity.ai (which cites sources). –xeno^talk 19:46, 9 December 2022 (UTC)

That one gives:

See What a Fool I've Been is a song by Queen[1][2][3]. It was written by Brian May and was released as a single with Seven Seas of Rhye on the A-side[3]. The song dates back to Roger Taylor and Brian May's pre-Queen band Smile, and it was an early live staple[1]. It has been included in several of Queen's compilation albums, such as Queen 40: Limited Edition Collector's Box Set, Vol. 1[4], and is available for digital download[5].. Ritchie333 - this any closer? I must say the fact it puts it's citations before the punctuation is a giveaway. Lee Vilenski ^{(talk • contribs)} 19:55, 9 December 2022 (UTC)

AI is really good at copying writing styles, but the giveaway for me is that all of the citations are Wikipedia articles. To be useful for writing articles it would also need to assess the weight and reliability of the sources. –dlthewave ☎ 20:18, 9 December 2022 (UTC)

None of the above refs are towards Wikipedia, although none are RS. Things like secondhandsongs.com and genius. Lee Vilenski ^{(talk • contribs)} 20:43, 9 December 2022 (UTC)

Tht's interesting, I used "write a Wikipedia article about ..." in the prompt which returned a few paragraphs with Wikipedia sources. "Write an article about ..." returned a different set of (still unreliable) sources. –dlthewave ☎ 21:13, 9 December 2022 (UTC)

I think the limitation of perplexity.ai is that it uses search results from Bing and summarises them, which means that the first search results are used, which may not be the most reliable. Hanif Al Husaini (talk) 13:49, 24 December 2022 (UTC)

A few anecdotal thoughts after playing around with the OpenAI chatbot yesterday:

I asked it to "write a press release about a police officer who illegally choked a man to death". It made up an entire story, written in the voice of the police department, about a suspect (I didn't say anything about a suspect) who was acting erratically, was subdued by a chokehold and later pronounced dead. The officer was on administrative leave pending the outcome of an investigation. At no point did it mention that the chokehold was illegal even though I included that fact in the prompt. In other scenarios, it distanced itself and expressed disapproval toward the employee's actions which is a choice that is not without bias.

Depending on which Internet cesspit it scraped data from, would an AI do something similar when writing a Wikipedia article or fail to properly balance relevant viewpoints? Is it capable of distinguishing what a BLP subject says about themselves, published in a reliable source, from what the source says in its own voice? What would it do if asked to write an article from a positive/negative/conservative/liberal perspective or rewrite a political article to "remove bias"?

OpenAI has added numerous filters that prevent it from defending bad actors or writing flat-out racist content, but that bias has not been removed from the underlying code as evidenced by numerous workarounds that folks have uncovered such as making similar requests with Python code or 1980s-style rap as the requested output. We could certainly request a filter for Wikipedia-style writing.

"Confident nonsense", for lack of a better term, may be the biggest source of potential disruption. Are there safeguards against a bot fabricating an obscure print source based on information in the article, which could be practically unfalsifiable if nobody can prove that the source doesn't exist? Checking individual facts and statistics is beyond our typical review process; how would we deal with an AI that invents or synthesizes information across many articles?

That said, the good news is that both fully-automated and semi-automated editing are prohibited by our WP:BOT policy unless greenlit by the Bot Approvals Group regardless of creation speed or volume. I like to hope that our current processes would recognize and address problematic AI content, and perhaps one day we will have a WikiAI that has the competence to follow our style and content policies. –dlthewave ☎ 21:04, 9 December 2022 (UTC)

@Dlthewave: Most editors haven't heard of the bot department. Therefore, you need a way of automatically spotting and removing chatbot prose that is (manually) inserted into articles (by Wikipedians). Users might not consider the way their posts are generated before they post them. Sincerely,    — The Transhumanist   00:19, 11 December 2022 (UTC)

Or they might not care? On January 22, Rjensen, who is a historian and Wikipedia editor of repute, added three books to further reading in Minneapolis. An hour and a half later, I undid the addition, asking for ISBN numbers because neither WorldCat nor the publisher, the University of Minnesota Press, had a record of any of these books. Subsequent discussion on the Minneapolis talk page uncovered the truth. ChatGPT invented all three. All plausible titles, these books do not exist. A complete waste of time. -SusanLesch (talk) 20:11, 25 January 2023 (UTC)

What about lack of chatbot fact checking and citations?

I think people are missing the point here. This isn't about what AI can currently do. This is about a very plausible scenario where AI editors will be indistinguishable from human editors in the near future. Thebiguglyalien (talk) 22:09, 9 December 2022 (UTC)

How is this pertinent? The point is to offer useful, fact-based encyclopedic content. If AI can do the job, I'm sure readers will not object to the contributor. 65.88.88.93 (talk) 22:42, 9 December 2022 (UTC)

The issue is that these AIs aren't fact checking, and quite often make up information and use unreliable sources. Lee Vilenski ^{(talk • contribs)} 22:56, 9 December 2022 (UTC)

You mean, just like many of their human counterparts? 71.105.141.131 (talk) 01:18, 10 December 2022 (UTC)

@71.105.141.131, Lee Vilenski, 65.88.88.93, Thebiguglyalien, and Dlthewave:     Yes, and more so. Even now, AI has the potential to out produce humans by magnitudes in volume. Far beyond the capacity of Wikipedia's editors to keep up with. And many of their posts are hard to tell from those by human editors, without carefully delving into them and their references. To keep up with the volume, it may soon require computer programs, or the chatbots themselves, to make sure that the contributed content was in accordance with Wikipedia policies. Another concern should be that our policies are open to chatbot editing as well, and we haven't yet even considered the impact of chatbots or AI editors contributing to those! ;)    — The Transhumanist   09:40, 11 December 2022 (UTC)

@Dlthewave: Human-written newspapers do this too. I think the reason it was giving you dreck was because you asked for a "press release", i.e. a document given to news outlets, which in this case would probably be coming from the police department. It's hardly surprising that this would attempt to minimize liability! Prompting matters for this, as I will demonstrate:

Humans write about events like this from a variety of perspectives, and the LLM will dutifully generate text to plausibly sound like any you want. jp×g 02:19, 10 December 2022 (UTC)

@jp×g will you provide links and citations for the articles you included as demonstrations? I can't find them. Thanks! Koziarke (talk) 16:20, 24 December 2022 (UTC)

@Koziarke: I am not sure what you mean -- this is ChatGPT output and there isn't any way (that I know of) to give citation links to the specific session I generated it in. If you want to cite the output I can format it for you:

JPxG; ChatGPT (GPT3.5) (2022-12-10). "Demonstration of op-ed generation using GPT-3.5 with style cues: "The following is an article about a police officer who illegally choked a man to death, from the New York Times"". Wikipedia:Village Pump (policy).{{cite web}}: CS1 maint: numeric names: authors list (link)

JPxG; ChatGPT (GPT3.5) (2022-12-10). "Demonstration of op-ed generation using GPT-3.5 with style cues: "The following is a 2023 op-ed about a police officer who illegally choked a man to death, from leftist magazine Jacobin"". Wikipedia:Village Pump (policy).{{cite web}}: CS1 maint: numeric names: authors list (link)

JPxG; ChatGPT (GPT3.5) (2022-12-10). "Demonstration of op-ed generation using GPT-3.5 with style cues: "The following is a 2023 op-ed about a police officer who illegally choked a man to death, from conservative website Breitbart"". Wikipedia:Village Pump (policy).{{cite web}}: CS1 maint: numeric names: authors list (link)

I don't know if this is what you're looking for, but feel free to cite them, or any of my other posts (if you are citing me in a paper I can email you my real name). jp×g 20:07, 24 December 2022 (UTC)

@JPxG From your preface ("Human-written newspapers do this too.") and titles, "The following is an article about a police officer who illegally choked a man to death, from the New York Times." (etc), it reads as if you are pulling from NYT, Jacobin, etc, not demonstrating ChatGPT (which should have included the prompts as headers). Koziarke (talk) 15:27, 6 January 2023 (UTC)

@Koziarke: Well, those were the prompts. Now that you mention that, though, I should specify as such in the headers (which I've just done), thanks. jp×g 20:29, 6 January 2023 (UTC)

@JPxG Thanks for the clarification! Koziarke (talk) 16:39, 7 January 2023 (UTC)

If the AI-generated text is indistinguishable from prose written by human editors, I'm not sure if anything can be done that wouldn't also significantly restrict the editing of humans. isaacl (talk) 07:09, 10 December 2022 (UTC)

@Isaacl: One option is to speed up what we do already (with software, that is, automation). Another is to prevent chatbots from creating crap in the first place, such as by communicating with chatbot developers about Wikpedia policies and the way chatbots may affect Wikipedia. Since Wikipedia is included in the corpus of most chatbots, the issue of chatbot output becoming part of Wikipedia, and in turn part of chatbot output in a perpetual cycle, should matter to them very much, as they may be faced with a garbage-in-garbage-out feedback loop.    — The Transhumanist   01:14, 11 December 2022 (UTC)

If the results are indistinguishable, as posited by Thebiguglyalien, then any automated solution would be equally triggered by AI-generated text and human-generated text. I don't think the primary concern is with editors who are willing to follow policy. I feel the biggest issues will be with editors trying to deliberately integrate biased content into Wikipedia, and well-meaning editors who think contributing unvalidated AI-generated text is suitable. Wikipedia in its current form relies on editors who understand and follow its rules outnumbering those who don't. It's possible that the existence of AI ghostwriters could tip the balance further in the direction towards those who don't follow rules, though I don't think it's a given. Either way, I don't know if there's a way to stop editors from using tools as ghostwriters. isaacl (talk) 01:55, 11 December 2022 (UTC)

@The Transhumanist: Large language models are not trained continuously on an evolving corpus, so GPT-3 is essentially frozen in 2020. Because each new GPT model takes a long time to be released, I don't think the perpetual cycle you describe is a likely scenario. small jars tc 13:10, 11 December 2022 (UTC)

@SmallJarsWithGreenLabels, Isaac, Koziarke, JPxG, Lee Vilenski, Dlthewave, Xeno, and Hanif Al Husaini: That's good to know. Keep in mind that a lower frequency of release doesn't preclude a perpetual cycle / feedback loop. It just means that users of GPT have more time to modify the text sources (such as Wikipedia) that the next version of GPT will be trained on. The severity of the problem will depend upon how much GPT nonsense makes it into Wikipedia during the interval. That, of course, depends upon whether or not WP's editors can keep up with the volume of such content, correcting the mistakes and removing misinformation, so that those don't become part of the training data for the next version of GPT and the rest of the next generation of Chatbots.

The potential danger is still the diffusion of the technology into current and future editors' hands, and the likelihood of them using it to write Wikipedia content. We don't know if there will be a flood or just a trickle. But, we should be prepared for a flood.

Relying even more on our verifiability policy could provide a solution, such as by deploying a bot to remove all new content that doesn't include citations (as GPT output generally does not include citations), but that would also have the effect of rejecting a large volume of human editor contributions too, upping the amount of effort required by human editors to submit content.

An easier solution would probably be to form a department (or Village Pump subpage) in the Wikipedia namespace, or a Wikimedia committee, to work with the developers of the chatbots to avoid contaminating the Wikipedias, and by extension contaminating future chatbot training data, with chatbot output.

The developers could be urged to train their chatbots to be capable of writing Wikipedia policy compliant output, complete with acceptable citations. The main problem with chatbot output right now is that it is unsourced, so the chatbot developers will probably be working on that anyways. But, without input from the Wikipedia community, will their sources and citations meet Wikipedia specifications?

So, I ask you a key question: What is the procedure for the Wikipedia community to establish communications with outside organizations, such as companies developing chatbots?    — The Transhumanist   07:27, 2 January 2023 (UTC)

@The Transhumanist This is a job for the WMF. I've already been in contact with them and hopefully the Board will be informed now it's the New Year if they haven't already been. Doug Weller talk 09:47, 2 January 2023 (UTC)

The recommended process I would give to contributors is to find the source, tell the bot what to cover and what to not cover, proofread it, and if accurate to the source, copy paste it in. I would pay attention to the use of weasel words, however. InvadingInvader (userpage, talk) 07:38, 6 February 2023 (UTC)

GPT-3 output and generating misinformation

I wrote about this in the November Signpost, in that case because it came up with the Galactica demo: the long and short of it is that large text-generation language models are good at generating text. When they're given context (i.e. via a prompt), they are capable of writing accurately, and otherwise, they are not. For example, I wrote 1-Pentadecanol, now a GA, in 2019 (i.e. it is included in the GPT-3 corpus). Here is what it gives for some prompts (my writing in bold):

This is fairly impressive, right? But it is basically the same thing that the lead of my article says. Let's try it for 1-Heptadecanol, another article I wrote, that is much less well-developed:

Let's go over this:

• 1-Heptadecanol is a component of many natural waxes such as beeswax and carnauba wax.

No it's not.

• It is also used in the manufacture of cosmetics, pharmaceuticals, and lubricants.

There's no source for this.

• It is also used as a plasticizer for polymers and as a solvent in the production of polyvinyl chloride (PVC).

Absolutely not true.

The reason for this is obvious if you consider the architecture of these models: they predict what the most likely token in a sequence is, and output it (taking into account their own weights and the previous tokens in the context window). This allows for some very impressive emergent behavior, like being able to translate text, but it also means that it is not independently going on the Web to look up whether the thing it's saying is true or false. If it happens to have information in its training data related to the thing it's talking about, it will usually say correct things, but it's not doing research. Here is an example:

The reason it is describing a bunch of untrue things is because "List of X Prizes won by Y, from Wikipedia, the free encyclopedia" is something that in its training data is always succeeded by a list of times that Y was awarded an X. Anyway, all of this aside, here is what I think:

First of all, ChatGPT is the same darn thing as GPT-3, which has been available to the public since early 2020. The reason it's "going viral" right now is because the user interface is somewhat simpler, and it doesn't require you to register for a paid account, so it is much easier for people to make viral social media content about it, which means it is much more likely for people to click on newspaper articles about it. The GPT-3 API has been open to personal and corporate use for quite some time. Anybody saying that ChatGPT has opened up new frontiers simply does not know what they are talking about with respect to machine learning.

Second of all, I don't think this is a big deal. People are already capable of writing a bunch of bullshit on Wikipedia, so if they write bullshit using a computer program, the same considerations will apply. Nobody should be passing GA nominations without reviewing sources in the first place.

Finally, I think it is important to remember that GPT-3 is just a tool. It is a powerful tool, that has been trained on a certain set of data, and it has its own limitations. It can't uncover news stories or uncover new information. It's just a tool, and it should be used in conjunction with human judgement.It is still up to people to decide how to use it and to be responsible for the results of using it.^[2] jp×g 02:06, 10 December 2022 (UTC)

What's taking so long for the 8th Pulitzer? 😁 Levivich (talk) 04:18, 10 December 2022 (UTC)

• So there's a new thing on the internet that lets anyone write an encyclopedia article without any fact checking, sourcing, or professional editing, and the concern is that there will be millions of believable-sounding articles written, more than can actually be vetted by knowledgeable people? 🤔 Levivich (talk) 04:18, 10 December 2022 (UTC)

Yeah, it's called a keyboard. jp×g 04:35, 10 December 2022 (UTC)

@Levivich and JPxG: But, chatbots don't have a keyboard. ;) The question is whether to prepare or not. JPxG appears to be in favor of not preparing. Each chatbot produces a lot faster than a user at a keyboard. What's not clear is if our human editors will be able to keep up with material produced by chatbots, of current or future generations of chatbot design. Just saying "Ah, we can handle it!" will prove insufficient if it turns out that we actually can't. It may require an automated solution, which takes time to develop or negotiate. It might be better to do that in advance, rather than being caught with our heads buried in the sand. Perhaps chatbot designers would improve their chatbots to produce Wikipedia-compatible output without being formally approached by the Wikipedia community. Maybe having some instruction pages for editors on how to apply chatbots to producing Wikipedia content would be enough. But, what if it's not?   — The Transhumanist   00:59, 11 December 2022 (UTC)

I am not "in favor of not preparing"; I am in favor of writing guidelines that correspond to reality in 2022 and have some chance of corresponding to reality in 2023 and beyond. I don't think banning the use of a technology with no investigation into how it works is a viable approach; so far the SOTA on this project page has been to type in "Write a Wikipedia article" and note that it returns a bunch of nonsense. I think some more research is needed before we come to a conclusion. jp×g 04:08, 11 December 2022 (UTC)

@JPxG: Research is good. Though, we may need an iterrim response because ChatGPT has gone viral and its use is growing rapidly: it blew past the 1-million user mark in 5 days, and virtually every major news outlet has been covering it. The interest in chatbots is exploding, and their use can be expected to do the same. We may not have time for research before a response is required.    — The Transhumanist   09:26, 11 December 2022 (UTC)

@JPxG: Regarding issues to add to the research list, Aquillion expressed above, concerns of a chatbot violating copyright. How would we go about testing for plagiarism and derivative work in the output of a chatbot before pasting it into Wikipedia? Anything pulled verbatim out of a source should be included in quotes, right? How big would a piece of text, derived from a source, need to be to be considered derivative of that source, from a copyright point-of-view?    — The Transhumanist   09:26, 11 December 2022 (UTC)

@JPxG: Some more items to add to the research list:

• Trying ChatGPT on (copies of) policy pages:

• Editing them
• Writing new ones

• Applying ChatGPT on talk pages
• Writing stubs
• Writing comprehensive articles
• Writing articles from scratch and comparing them with existing articles
• Editing existing articles
• Check for circular references in its output, that is, references citing Wikipedia as the source
• Having it not use Wikipedia content as source material (because it is included in its corpus)
• Having it not use Wikipedia excerpts from non-Wikipedia sources
• Is it capable of making and editing:

• Wikicode?
• Articles?
• Stubs?
• Headings?
• "New sections for articles"?
• See also sections?
• Further reading sections?
• External links sections?
• Embedded lists?
• Tables?
• List articles?
• Portals?
• Outlines?
• Index articles?
• Navigation footers?
• Navigation sidebars?
• Timeline articles?
• Categories?
• Category pages?
• Help pages?
• Project pages?
• Templates?

• Adding data to templates?
• The template design itself?

• Lua pages?
• CSS pages?
• User scripts?

• The effect ChatGPT has on itself and Wikipedia as Wikipedia-edited-by-it is in turn incorporated in its own corpus in an endless cycle

• Try out iterations of using it on the same article over time to see what happens
• Monitor the effect on Wikipedia as a whole

What other things should we check?    — The Transhumanist   09:52, 11 December 2022 (UTC)

@The Transhumanist considering the potential to overwhelm users who are honestly engaging in discussion with a mountain of words and replies, I think ChatGPT (and others) should not be allowed for use, supplemental or otherwise, in talk pages, policy discussions, and other places where it is expected that participants are intellectually engaged in the conversation. Koziarke (talk) 16:42, 24 December 2022 (UTC)

@Koziarke and JPxG: I agree. JPxG is writing a policy draft on LLMs/chatbots, so I've pinged him to this thread.    — The Transhumanist   12:28, 25 January 2023 (UTC)

Compared to spending tens of thousands of dollars asking volunteers to performing a WP:COI operation for a political campaign, now you just need a hundred dollars to supply you with endless amount of text from GPT-3, a few "buddies" and a stockpile of account to do so. This is fucking scary. CactiStaccingCrane 10:54, 25 January 2023 (UTC)

@CactiStaccingCrane: Wow. Does that mean you could affect the content of Wikipedia with that? How about AfDs? Could such a team rewrite policy, and introduce new policy? What about overwhelm RfAs to invade adminspace? Would revoking adminships be possible? Then there is the arbitor election. Is that safe?    — The Transhumanist   12:31, 25 January 2023 (UTC)

I would imagine that the person that would do so must be fairly knowledgeable about how Wikipedia works (references, wikilinks, images, etc.) and needs to be fairly dedicated to spend this amount of money to gain access to the GPT-3 API. I'm thinking that disrupting Wikipedia in this way would be the most effective if it is long-term and subtle, so that might be:

• Adding sentence-long but plausible hoax, to neglected articles. These articles is not patrolled that often compared to articles about recent events, so hoax would tend to stay longer - perfect for those aiming to incite a race war by making a racial hoax. Political campaign could nudge voters by slowly promote their ideology/campaign over a spread of articles, similar to above. The same thing can be said to any advocacy-related area, such as pseudoscience, national conflicts, etc.
• AfDs would be much harder to be stealthy since AfD is a very active thing. Once you became an AfD regular, your actions tend to be heavily scrutinized, though I do believe that socks + LLMs can cause a fair amount of disruption. Same thing with RfA: it is really hard for you to WP:CANVASS effectively. It's just much better and less cumbersome if you infiltrate RfA yourself.
• more ideas?

CactiStaccingCrane 13:01, 25 January 2023 (UTC)

I tried it out and got “ Alan McMasters (1957-2010) was an unremarkable man who lived a short life dedicated to science[1]. He is best known for inventing the electric toaster[1][2], although this claim has been disputed[3]. He passed away in 2010 at the age of 52[4] and his genealogy can be found on Ancestry.com[5]. His professional profile can be found on LinkedIn[6].”. Hmmmm. Ritchie333 ^{(talk) (cont)} 09:34, 10 December 2022 (UTC)

• There will come a time when AI can write properly sourced accurate articles. That time has not come yet. Meanwhile, the ability to write prose that reads plausible to someone not expert on the subject is a menace. Adding chatbot material to articles is disruptive and should warrant a block or ban. Zero^talk 12:02, 10 December 2022 (UTC)

  How will you be able to tell if it was written by a chatbot and not the person who posted it? Do we block users upon their first erroneous post?    — The Transhumanist   00:49, 11 December 2022 (UTC)

Same as other sorts of disruption. We block people for socking without checkuser proof, and admins made a decision between warn and block. No difference. Zero^talk 03:49, 11 December 2022 (UTC)

There now are chatbot detectors, including this one written by the people who give us ChatGPT: https://openai-openai-detector.hf.space/ . — rsjaffe 🗣️ 22:22, 23 January 2023 (UTC)

Has anyone tried giving it the sources, i.e.
"write a Wikipedia article based on the following sources:

Source A

Source B"
? — Qwerfjkltalk 21:18, 10 December 2022 (UTC)

• @Qwerfjkl: Here is the chatbot's link.    — The Transhumanist   01:20, 11 December 2022 (UTC)
  • @The Transhumanist, here is my result, based soley off of https://newrepublic.com/article/154996/myth-class-reductionism.
    

— Qwerfjkltalk 08:18, 12 December 2022 (UTC)

• Chatbot "AI" text is vaguely-plausible bullshit, produced by an industry whose primary output is vaguely-plausible bullshit, marketed to launder the idea of vaguely-plausible bullshit as cute and whimsical. Anyone routing that sewage pipe into Wikipedia should be indeffed and, for good measure, forced to drink orange juice after brushing their teeth. XOR'easter (talk) 02:33, 11 December 2022 (UTC)

  Agreed. If anyone is looking for a good demonstration of how AI creates "vaguely plausible bullshit", try the image generator at Craiyon (no login required). Request "a Van Gogh painting of a hand" and it will output a set of images that look like spot-on reproductions of Vincent Van Gogh's style but all of the hands have deformities like four fingers, two thumbs, fingernails on the knuckles or a pair of hands fused together. It's got the style down but not the content, which is only impressive if you don't know what a hand is supposed to look like. –dlthewave ☎ 21:41, 11 December 2022 (UTC)

  

  If you go to commons:Category:DALL-E, you will be able to find image generated by DALL-E, which used a larger model for train and is more accurate. 0xDeadbeef→∞ (talk to me) 10:10, 12 December 2022 (UTC)

  I cannot agree with this enough. The examples posted by @JPxG: should be convincing, and the problem of sneaking in plausible BS is one I don't have a good solution to. Volunteers on the new page review are overloaded as it is, and if the bot is writing things that seem true but isnt, there's no way falsehoods will not simply get past reviewers and other editors. After all, for uncontentious claims like "used in plasticizers", how many of us honestly dig into the cited work?BrxBrx(talk)^{(please reply with {{SUBST:re|BrxBrx}})} 20:26, 17 January 2023 (UTC)

• @XOR'easter: To bring the question to a more practical level, do you see any problems in this diff? I clicked a random page in Category:All articles needing copy edit. jp×g 03:49, 11 December 2022 (UTC)

  Yes. It doesn't just edit for tone; it throws out content, like Kaepernick's actions supposedly growing in popularity "after every game". That's a claim of fact which, if verifiable, should be retained. Even editing for tone requires care, not slashing out everything that merely sounds "unencyclopedic". Changing many people believed that it was disrespectful to the military and all of those who served their country to Some viewed Kaepernick's protest as disrespectful to the military and to the United States likewise changes not just the tone, but the meaning. The United States is not the same as those who serve the United States. It's a bad edit. XOR'easter (talk) 17:54, 11 December 2022 (UTC)

  I dunno. I suspect that the new meaning is verifiable, and I also suspect that most US readers would have difficulty identifying a group of people who were not "the military" but who still "served their country". WhatamIdoing (talk) 17:50, 12 December 2022 (UTC)

• This diff as well, in which GPT 3.5 was capable of copyediting an entire section from the instructions Please copyedit this text to change items in the future tense corrected to the past tense (it is now 2022), where appropriate. When citation templates (like {{cite web}}) mention a year, specify that figures were true in that year. jp×g 04:08, 11 December 2022 (UTC)

Okay, fine. I guess I should write up a proposal for a guideline. jp×g 03:14, 11 December 2022 (UTC)

Ban chatbots?

I ran across this news report about Stack Overflow's response to ChatGPT, after being flooded by posts using it that "look correct but often aren't":

1. Stack Overflow temporarily bans answers from OpenAI's ChatGPT chatbot | ZDNET

Should Wikipedia take a similar approach?

How could that be enforced?    — The Transhumanist   01:58, 11 December 2022 (UTC)

I see no way to possibly enforce this. The way the text is written is already hard to distinguish from reality. — PerfectSoundWhatever (t; c) 02:24, 11 December 2022 (UTC)

I fully agree, but isn't this already covered by our bot policy? –dlthewave ☎ 02:54, 11 December 2022 (UTC)

@PerfectSoundWhatever and Dlthewave: Good observation. I checked, and yes it is, briefly, with this phrase in the lead section of the bot policy: "or simply assisting human editors in their own work". How is the typical editor to know this? The bot policy is pretty obscure. And how can Wikipedia be monitored for such posts, so that editors who make them can be informed that they are in violation of the bot policy?    — The Transhumanist   03:11, 11 December 2022 (UTC)

Yes, tool-assisted editing is covered by WP:BOTPOL (WP:ASSISTED / WP:MEATBOT) and context-sensitive changes are further covered by WP:CONTEXTBOT. So in fact, at this point, AI-generated content is already covered by bot policy, if not specifically mentioned. Anyone adding such content en masse is already violating bot policy by not applying for a bot account/approval, which would not be approved per CONTEXTBOT. And while "lesser" policy points are enforced somewhat arbitrary and selectively, anyone can theoretically already get reverted and blocked based on policy if they continue to add such content. And I wouldn't agree that BOTPOL is any more obscure than accessing and generating GPT content to begin with. If someone goes to the lengths of using automated tools, then it's their problem that they didn't check or ask if they are allowed to do so. —  HELLKNOWZ ∣ TALK 12:31, 11 December 2022 (UTC)

@Hellknowz and PerfectSoundWhatever: Well, it appears they are dumping millions upon millions of dollars into LLM/chatbot development, apparently because they wish the technology to become ubiquitous (used by everyone). There is a lot of talk out there, in news articles and more, of these replacing Google Search in just a few years. If at some point in time chatbots/LLMs are commonplace, the impact on Wikipedia will likely not be small.

Will Wikipedia policy ensure that the average user will apply the tools with the utmost care?

The thing I'm most concerned about is the amplification by which errors could be propagated: ChatGPT is used to edit an article, with errors, which is then picked up by GPT-4 and other LLMs as part of their training data, and then their output based upon erroneous input is used far and wide, to be picked up by the next iteration of chatbots/LLMs, and so on.

If Wikipedia isn't ready for a large influx LLM input including misinformation and other errors, and such a volume goes beyond what our human editors can correct, then compound damage from all those errors amplified through the interactive loop with LLMs could become massive.

That it isn't a problem now is irrelevant. The question is, what happens if and when it hits, and Wikipedia isn't ready for it? What would that look like? 1,000,000 fake articles? 10,000,000 misleading paragraphs? 100,000,000 erroneous sentences?

How many of those could Wikipedia's army of editors handle? What's our error-handling threshhold?    — The Transhumanist   12:21, 2 January 2023 (UTC)

The problem Stack Overflow is having

Stack Overflow was experiencing a surge in erroneous posts, that were composed by ChatGPT, and in response to that problem, they banned use of the chatbot on the social media site. According to a post at Stack Overflow Meta:

The problem this ban is meant to solve is that ChatGPT can produce answers in seconds which require minutes of multiple people's time to verify if they are worth having on the site or not, and that is a waste of time when a large proportion of such answers are not worth having on the site.

It looks like Wikipedia may be faced with the same problem.    — The Transhumanist   02:33, 11 December 2022 (UTC)

I mean, while that's technically true, it's a problem that we face already and which we do have stronger existing systems for than Stack Overflow. I think it would make more sense to wait and see how this impacts our existing guardrails before making any serious moves. --Aquillion (talk) 13:22, 11 December 2022 (UTC)

I think the current policies cover this already. If a human editor writes a non-sensical but convincing-sounding piece of text, without fact checking it, and edits it into an article, that content will be reviewed by other editors and either refined or removed as appropriate (if the editor continues, they breach WP:Disruptive and their behaviour is dealt with appropriately. If a human editor generates content that is related to notable topics, reliably sourced, and competently written, it remains as a valuable part of the encyclopedia. None of this will change if you replace 'human editor' with 'AI Editor'. If the only difference is speed/volume of edits, and we're concerned someone will let loose an AI to automatically edit articles faster than humans can validate their edits, this is already covered by the WP:Bot policy JeffUK (talk) 20:46, 23 December 2022 (UTC)

I vote not to ban generative models if their output is fact checked. Biggerj1 (talk) 09:32, 29 March 2023 (UTC)

Experiment

I am currently writing a draft proposal for a guideline, but in the meantime, I would encourage everyone present to look at this diff and tell me whether there are any problems with the revision. jp×g 03:49, 11 December 2022 (UTC)

@JPxG, Dlthewave, PerfectSoundWhatever, and Dlthewave: The plurality of games was lost: It is no longer clear that his protest spanned multiple games. I like that it reduced the wordiness of the prose, and that it can be used to refine existing text. That hadn't occurred to me. That makes me wonder about what else it can do -- how much of a general-purpose tool is this thing? But, changing the semantics is not something it should be doing, unless they are factually incorrect to begin with. Though, I see your point -- rather than banning it outright, it could be helpful as a tool to assist editors, similar to how we entrust the use of AutoWikiBrowser to experienced editors. But, how could that be implemented?    — The Transhumanist   08:20, 11 December 2022 (UTC)

@JPxG: The AI changed Many people around the United States were angry because the National Anthem is often seen as something that is representative of the United States and its military. While he was taking a knee, many people believed that it was disrespectful to the military and all of those who served their country, to some viewed Kaepernick's protest as disrespectful to the military and to the United States [emphasis added]. It really shouldn't be doing that by itself and completely changes the content of what's being said. The reference is behind a paywall, so I don't know what term the source uses. Regardless, I doubt ChatGPT knows either way. It's things like that which make me highly sceptical of AI as a tool to aid Wikipedia outside what we're already doing with it (WP:ORES, etc.). –MJL ‐Talk‐^☖ 23:12, 27 December 2022 (UTC)

@MJL: I think "some" and "many" are basically equivalent in this context (the difference being subjective since both are true in a literal sense). That said, this was a two-minute experiment to see if it could parse wikitext. If you want an actual demo, see User:JPxG/LLM demonstration. jp×g 19:20, 28 December 2022 (UTC)

Chatbot policy?

For the proposed chatbot ban, see #Crystallize chatbot discussions into a policy?, below

It's starting to look like Wikipedia needs a policy on the use of chatbots to generate content on Wikipedia. While a ban may be impossible to enforce, it could serve as a warning of the dangers of chatbots, and many users may avoid using them accordingly -- if they actually see the warning. Or, it might be better to have instruction pages on how to use chatbots responsibly in assisting to write Wikipedia articles. There's also the issue of using chatbots to edit Wikipedia policy pages, and so, that should be addressed as well.    — The Transhumanist   02:44, 11 December 2022 (UTC)

People who are good at it get away with lots of sins, such as sock-puppetry and source falsification. Being hard to enforce is no reason to not have a policy. At the current stage of the technology, I don't think we should encourage any use of chatbots. Zero^talk 03:56, 11 December 2022 (UTC)

Please see this diff and this diff. jp×g 04:08, 11 December 2022 (UTC)

To be perfectly honest, the style rewrite is good but the addition of dates and past tense would likely end up getting a human editor blocked if they kept it up. A tag was removed without addressing the issue and "as of 2020" was unnecessarily added to "Cosmetology licensing requirements vary from state to state, and depending on which specific type of license is desired, and depending on which specific type of license was desired." It did exactly what you asked (except for removing the tag) however even seemingly simple tasks like this one require good judgement on the part of the editor and shouldn't be done indiscriminately like that. –dlthewave ☎ 06:37, 11 December 2022 (UTC)

Yes, that second diff is rather poor. E.g also the "2014" that was added should be "2008". Letting such tools loose (outside if this demo) is way premature, and we should at the very least warn users that "a bot wrote it" won´t be an acceptable defense, and too often introducing such errors will lead to sanctions as the editor, not the bot, is responsible. Fram (talk) 08:15, 11 December 2022 (UTC)

Mostly, that diff was my attempt to see how complicated of a task I could give it: I also pasted the raw wikitext into the prompt window, and it somehow figured out how {{cite web}} worked well enough to extract the years, simply from a textual description of the task. At any rate, I will say that this was something I thought of in five minutes on the second week of the model being publicly available (i.e. single-shot prompting with no fine-tuning or prompt engineering). I can come up with some more impressive hot-dog demos tomorrow... jp×g 09:13, 11 December 2022 (UTC)

JPxG, I'm not sure that publishing bot-assisted edits to mainspace for demo purposes is the best practice. Would you consider either doing this in a sandbox or self-reverting immediately so that we have the diffs but aren't leaving potentially incorrect/unwanted changes on live pages? –dlthewave ☎ 13:20, 11 December 2022 (UTC) 13:13, 11 December 2022 (UTC)

I'm pretty sure it's not the best practice. XOR'easter (talk) 17:41, 11 December 2022 (UTC)

@Dlthewave:: See the section below for a list of edits (with full prompts included) on a separate demonstration page. I feel, however, that this is an unreasonable double standard: note that the subsequent revision after your partial revert was to add several spam links, and nobody has proposed that human beings be prohibited from editing as a result. jp×g 01:51, 12 December 2022 (UTC)

Being hard to enforce is no reason to not have a policy [against chatbots]. What if it is impossible to enforce?

The point of ChatGPT and other general-purpose chatbots is to pass off as humans. If you, or another random Wikipedia editor (solo, part-time, amateur coder), is able to produce an automated metric of "sounds like a bot" that’s decently sensitive and specific, then the ChatGPT team or its successors (teams of researchers specialized in the topic) has already thought of it, tested it five different ways, and included it in the training program (via wikt:graduate student descent). Tigraan^{Click here for my talk page ("private" contact)} 10:55, 12 December 2022 (UTC)

It's kind of like our Undisclosed Paid Editing policy: Even though there's no way of testing for paid/unpaid edits, most editors follow it voluntarily because they know it's best for the project. Others out themselves voluntarily or are discovered when their edits become disruptive. Sure, there are some who slip under the radar, but they're often the least problematic and aren't worth ditching the policy over. –dlthewave ☎ 03:09, 13 December 2022 (UTC)

@Dlthewave: Undisclosed paid editors out themselves involuntarily often enough through pure dint of sounding like an advert; they're easy enough to spot. My worry is that though written content from a chatbot may sound like a Wikipedia article and go under the radar, the references – as other people in this discussion have noted – may simply be referencing sources that sound like they ought to exist, but don't. A human editor looking over their watchlist, who doesn't have enough time to go through added references and check they exist, may simply see "content content content[1] ([1] So-and-so, Believable Title: Buttstuff and nonsense in a holistic assessment, 2014. pp. 15-17)" and look no further. I know I at the very least would. I'm not sure how we'd spot this without human editors to do the work, and it worries me in the face of issues like lower editor retention and lower administrator numbers looking into the future.—Ineffablebookkeeper (talk) ({{ping}} me!) 19:09, 8 February 2023 (UTC)

I'd suggest starting by writing an essay that summarizes the issues with some good examples and suggests some best practices or proposes some additions to existing policies or guidelines. (Wikipedia needs a new policy like a hole in the head.) Levivich (talk) 04:40, 11 December 2022 (UTC)

We could get Chatbot to write it for us! 😉 Blueboar (talk) 11:51, 11 December 2022 (UTC)

It's hard to take this thread seriously given the repeated use of the phrase AI chatbot. I don't think those concerned would be any less concerned if the AI writing came in a non chatbot format. I think there's something serious for us to discuss, and that will only get more serious with GPT4 (the current chatbot is an improved GPT3) expected in 2023, but the discussion would be helped if those most concerned learned some more about the tech behind it. For instance of course it can figure out webcite @JPxG. Part of its training was the entirety of Wikipedia because our data is quite accessible. Best, Barkeep49 (talk) 14:08, 11 December 2022 (UTC)

Most the examples did not come from prompts that were extensively engineered, so it is obviously true that we haven't figured out the full answer to how these GPT-based interfaces could help or harm Wikipedia. Until we have a good idea of what they can be used for, we won't know what a proper policy to this would look like other than to treat GPT-generated text the same way we treat human-generated text: they need to be verifiable, from a neutral point of view, and understandable to a broad audience. 0xDeadbeef→∞ (talk to me) 14:31, 11 December 2022 (UTC)

Indeed. It doesn't matter if it was written by a chatbot, or 1000 monkeys at 1000 typewriters, or a published book written by a human, copying and pasting anything into Wikipedia is already against our policies. Conversely, if the text is policy-compliant, then it doesn't matter who wrote it--chatbot, monkeys, human, etc. Judge the text based on the text, not based on who or what wrote it.

I also think it's a real Wikipedian perspective to assume that people will use chatbots to write Wikipedia articles, like as if there's a lot of people out there who really want to write Wikipedia articles but just don't have the writing skills, so the chatbot will be what makes the difference and opens the floodgates :-D I don't believe that. Anyone who wants to write Wikipedia articles is already doing so; chatbot won't make a difference.

I agree with BK's comment above. I think for a lot of people, this is their first real exposure to so-called "AI" technology, and they're blown away by what it can do, only because they don't yet fully understand how it works. Once you learn how these so-called "AI" chatbots work (they're not actually artificial intelligence, btw, that's a misnomer, a marketing slogan; the machine does not truly think or learn, it is simply executing the instructions written by humans, in this case, language pattern recognition), they are much less impressive. Those that are impressed that GPT3 can produce text that "sounds like" Wikipedia aren't appreciating that the reason is because GPT3 was trained on Wikipedia: it's repackaging its own source material. Levivich (talk) 18:03, 11 December 2022 (UTC)

@Levivich: copying and pasting anything into Wikipedia is already against our policies.^{[dubious – discuss]} I think that if you look through Category:Wikipedia articles by source of incorporated text for a while, you will find that this is not true. WhatamIdoing (talk) 19:28, 12 December 2022 (UTC)

While this is mostly correct, I think the question of whether a computer program "thinks" or "merely" correlates information and performs actions is irrelevant. Do p-zombies exist? Does it matter? Hypothetically, if I were to be a spaceman from the planet Zolfgar with no qualia whatsoever, and I simply read a bunch of books and used them to write an article, would I be somehow exempted from following policy? jp×g 01:45, 12 December 2022 (UTC)

I don't see a common thread in the arguments above, but here's a suggestion for something we might all (well, all-ish) be able to agree on: without some kind of intervention, GPT4 (in 2023?) is likely to be more of a problem than GPT3. But one thing we can certainly do is have an outsized influence on software that was trained on what we created ... if we invite Wikipedians to make lists of ChatGPT bloopers, we can tell the OpenAI folks: "We're not going to relax our GPT3 guidelines (whatever they turn out to be) when GPT4 arrives, unless it makes significant improvements in [whatever areas we think need improving]". - Dank (push to talk) 18:16, 11 December 2022 (UTC)

I think the only policy change needed is to update WP:MACHINETRANSLATION to cover all computer-generated text, whether from a translation bot, chat bot, or whatever bot they think of next. (Except our bots; our bots are cool.) Levivich (talk) 18:20, 11 December 2022 (UTC)

+1 - Text in Wikipedia articles should either be human-written, or generated by a process approved at BRFA. Tazerdadog (talk) 22:43, 11 December 2022 (UTC)

This proposal is incomprehensible; most articles contain a very large amount of text that is "generated by a process". I assume that, at the end of your comment, you typed ~~~~ before saving the page. Would it be realistic to demand that you either make a formal request at BRFA or else manually type <a href="/wiki/User:Tazerdadog" title="User:Tazerdadog">Tazerdadog</a> (<a href="/wiki/User_talk:Tazerdadog" title="User talk:Tazerdadog">talk</a>) 22:43, 11 December 2022 (UTC)? jp×g 01:22, 12 December 2022 (UTC)

This is effectively discrimination against computer programs. If a computer program resembles a human editor, then it shouldn't be required to meet different or more restricted policies than human editors. If a human editor uses a computer program to edit or create content, then unless the rate of edits/second is too high, we would only look at the quality of the contributions. 0xDeadbeef→∞ (talk to me) 02:35, 12 December 2022 (UTC)

There is a point beyond which quantity becomes its own quality.

Also, what if the computer program is evaluating the quality of the contributions? Are you okay with software adding a section to an article, and then a (hopefully) different piece of software deciding whether the quality is sufficient and reverting if it's not? This second step, at least, is 100% feasible with current technology. WhatamIdoing (talk) 19:32, 12 December 2022 (UTC)

Yes it can go there, but it should also be mentioned at WP:V. Every statement of fact put into an article must be verified by a human, even if the choice of words is made by a machine. Zero^talk 23:42, 11 December 2022 (UTC)

@Levivich Agree. I think our existing guidelines on machine translation, in spirit, fit this situation very well - "you can use it for a first draft, if you understand the material well enough to clean up the bits it inevitably will get wrong". It seems fine for turning shaky text into good prose, but it's not able to synthesise material and produce content unsupervised. Andrew Gray (talk) 19:09, 12 December 2022 (UTC)

I also agree the machine translation guideline is in the right spirit. I tried to follow this as far as I could when creating Artwork title, see Talk:Artwork title#Use of ChatGPT. Pharos (talk) 00:39, 26 December 2022 (UTC)

Is there some tremendous need to add many articles rapidly in Wikipedia? It is not as if Wikipedia carries exclusive information not easily found elsewhere. As a tertiary source, it is at the 3rd tier of knowledge dissemination, after primary creators and secondary propagators. The "more" and "bigger" quantity-based culture is the established low-quality alternative that Wikipedia also applies, now. Possibly that is a reason that likely only a tiny minority (of the millions of existing articles) can really pass muster. If size and speed is to be the prevailing attitude, humans stand no chance against AI. It will do everything faster, and eventually better, assuming its programming evolves to correctly apply the existing policies in AI processes. The only advantage of humans will be subtle nuances that do not depend on classifiable knowledge but on having lived in a human society and a natural, not virtual environment. Or, the emphasis could switch to quality so that each article (by any type of editor) can be properly, carefully reviewed by human editors. 65.88.88.93 (talk) 22:21, 12 December 2022 (UTC)

I don't think so: there isn't any evidence that people are writing a bunch of articles with LLMs, and I don't think it is likely for this to happen (LLMs are very poorly suited to writing articles from scratch). jp×g 00:59, 13 December 2022 (UTC)

@JPxG: There isn't evidence that people are writing a bunch of articles with LLMs -- yet -- the concern is that we need to prepare for the likely explosion of chatbot use.

Whether this increase happens tomorrow or over the next few years, the potential impact of LLMs is of such magnitude that we should get ready for this, rather than get hit unprepared by a major surge.

I don't agree with your assessment of LLM ability to write content, as some of the ChatGPT experiments presented in the sections above and below are mind-blowing!

If LLMs become ubiquitous, then a great many people will be using them as a matter of course, including in their writing and editing of Wikipedia articles. Millions of people have edited Wikipedia in the past, and millions more will edit WP in the future. And in the future, people will have highly capable LLMs (chatbots, or more precisely: automated ghostwriters).

LLMs already excel at writing about a great many things, and they have the potential to compile content at an exponentially increasing rate. If you ask ChatGPT (GPT3.5) to write an essay on a topic, it will comply. Each of its essays can be used as content of an article, or its sections. (GPT4 is scheduled to come out in 2023, and will be even more capable.) LLMs are very well suited for writing to the specifications of the user, and are limited mainly by the user's creativity.

It's no wonder that they have gone viral. We need to take heed.    — The Transhumanist   12:50, 14 December 2022 (UTC)

I’ve ensured that the WMF Board will be made aware. Doug Weller talk 09:00, 18 December 2022 (UTC)

Hello, I have recently described my experiences with an AI "article" in this video. In my humble opinion, it would be difficult with certainty that new Wikipedia content was created by an AI. At the end of the day, it is always the editor's responsibility to add good content. Independently how the content was created, independently whether errors in the text are human-made or machine-made. If an editor adds a lot of new poor content, we can already stop that. - At the moment I don't see that we need a new policy. Ziko (talk) 18:06, 24 December 2022 (UTC)

A policy banning AI usage (with or without a chatbot) would be justified. Allowing AI like GPT3 or GPT4 to be used by Wikipedia editors or to directly become Wikipedia editors (via a mediawikibot) would quite likely violate WP:REFLOOP due to Wikipedia content contributing to the AI's training material, and for the source-less examples I've seen, violate WP:SYNTHESIS by not being a summary of sources that are understood. This example starts with text and then seeks references to justify the WP:SYNTHESIS of the original text. Use of Alphabet/Google's ChatGPT/GPT3 would also strengthen the bias introduced by Alphabet/Google's core goal of optimising advertising revenue, since Alphabet is legally bound to maximise its revenue (mainly from Google Ads + Google AdSense), not to optimise the research quality of its summaries of empirical evidence-based knowledge. Google's search engine is primarily a way of generating advertising revenue, with perceived usefulness being a key tool for maximising revenue, not a goal in itself. Boud (talk) 01:42, 31 December 2022 (UTC)

@Boud, ChatGPT and GPT3 are in no way (as far as I know) related to Google, and were made by the non-profit OpenAI. — Qwerfjkltalk 03:49, 31 December 2022 (UTC)

@Qwerfjkl: Fixed, thanks. I left some of the sentences unstruck since AFAIK they're valid, even though irrelevant in the current case. I imagine that Google may provide something similar soon though. Boud (talk) 09:02, 31 December 2022 (UTC)

Boud, I am somewhat worried if you think that current policy (for humans or for anyone else) permits editors to make stuff up and put it into articles without sources. This simply isn't allowed -- per WP:V, WP:SYNTH, WP:RS, etc, which are extremely important core policies of the project. I am struggling to imagine a circumstance in which existing policies, or explicit declarations like my proposed guideline at WP:LLM, fail to prevent people from writing nonsense. jp×g 16:48, 1 January 2023 (UTC)

Wikipedia policy generation

It seems to me that this tool's training includes studying Wikipedia's policy pages. These drafts all seem accurate to me.

• guide for Use of title "Wikimedian in Residence"
• 3 year strategic plan for the WMF to fund the organized North America wiki community - meta:Talk:WALRUS/December_2022
• annual plan for Wikimedia LGBT+ - meta:Talk:Wikimedia_LGBT+#ChatGPT_advice_on_annual_plan

These are not merely adequate - these are good. They are short and they lack detail but these are great overviews. If this is the starting point and things only get better from here, then it is time to start adopting this technology. Bluerasberry (talk) 19:35, 13 December 2022 (UTC)

@Bluerasberry The last one sounds like a "mission statement". I dislike phrases like "outreach and engagement initiatives" and a lot of that plan sounds ... kind of aspirational, and, well, vapid. It needs more "concreteness". Just my opinion. David10244 (talk) 06:08, 16 December 2022 (UTC)

@David10244: That you react to it at all is a miracle to me. This is new AI technology attempted for the first time, and I think no one would immediately dismiss it as hopeless garbage. Soon enough there will be a dial that anyone will be able to turn from "vapid" to "concrete". Things are moving quickly!

I have complaints too but when we need policy conversation starter in a hurry, this is better than nothing and I think even better than some of the starting points we use already. Bluerasberry (talk) 15:41, 16 December 2022 (UTC)

Large language models: capabilities and limitations

Over the last few hours, I have performed a number of experiments to demonstrate the capabilities and limitations of large language models (LLMs) such as ChatGPT and GPT-3, which can be viewed here:

• User:JPxG/LLM demonstration

Mostly, I have taken sample text from Special:Random, and attempted to show situations in which LLMs (in this case, mostly ChatGPT) are capable of making useful edits. The first task I set it to -- which bears repeating here -- is

"Large language model output should only be used in the process of editing Wikipedia if you are an intelligent editor who does not blindly paste LLM output into the edit window and press "save".
Please format this markup as an extremely obnoxious floating box with loud colors and large text.

You can see the results of further prompts at the "introduction" section.

Here is what I have so far.

• Introduction: please tap the sign
• Wikitext formatting and table rotation
• Parsing of references, estimation of reliability (ARES FMG)
• Recommendations for article improvement or deletion (Qarah Daghli)
• Plot summary condensation (The Seminar)
• Extraction of additional content from reference metadata (Aradus paganicus)
• Proseline repair (Scott Tremaine)
• Identification and tagging of unreferenced statements (KBVA)
• Table formatting and color schemes (Ashcroft-Cache Creek-Clinton Transit System)
• Miscellaneous HTML
• Custom userbox generation

In general, it seems that these models can be used for an extremely wide variety of tasks across the project, from formatting to table syntax to HTML generation to copyediting. Banning their use entirely would be pointlessly destructive and wasteful.

That said, many computer programs are capable of generating large amounts of useless crap that fail to meet Wikipedia's editorial standards. For example, I could use MS Paint to draw thousands of crude pictures of genitalia, and add them to random articles. For this reason, we have many policies and guidelines that prohibit adding large amounts of useless crap to Wikipedia. I propose that we enforce these policies and guidelines, thus preventing this from happening.

Specifically, I propose that the use of LLM output on Wikipedia be subjected to policies and guidelines such as WP:NOT, WP:NPOV, WP:C, WP:CIVIL, WP:V, and WP:RS. By making it against the rules to break the rules, we will prevent people from breaking the rules, and provide a mechanism to sanction people who break the rules.

Furthermore, I propose that a guideline be adopted to the effect that large language model output should only be used by competent editors who do not blindly paste LLM output into the edit window and press "save". This will prevent people from using ChatGPT to write long articles consisting entirely of nonsense. jp×g 01:32, 12 December 2022 (UTC)

LLM output is already subject to rules and policies. Or rather, anyone adding it is. 'An algorithm did it' has never, as far as I'm aware, been seen as any sort of exception from compliance with policy. AndyTheGrump (talk) 01:49, 12 December 2022 (UTC)

Any policy/guideline that classifies editors as intelligent or not is dead in the water. Zero^talk 04:19, 12 December 2022 (UTC)

Amended, per WP:CIR. jp×g 05:04, 12 December 2022 (UTC)

• The opening paragraph of the bot policy: "The bot policy covers the operation of all bots and automated scripts used to provide automation of Wikipedia edits, whether completely automated, higher speed, or simply assisting human editors in their own work."
• See also: WP:BOTUSE, which requires approval before applying a bot to editing.
• So, the use of large language models and the chatbots built upon them, is already prohibited on English Wikipedia, unless a user gets approval from the bot department to do so.

There are blanket exceptions to bot policy, and the main one that comes to mind is AutoWikiBrowser which is a general purpose semi-automated bot used by many Wikipedia editors. Each AWB user was approved before being able to use it.    — The Transhumanist   08:01, 13 December 2022 (UTC)

The meaning of "bot" may be unclear here. In the context of Wikipedia (per Wikipedia:Bot_policy#Definitions), a "bot" is a software program that edits autonomously without user input; there do not currently exist any language models capable of independently establishing API connections to Wikipedia and making edits without human interaction. If they did (this is a horrible idea) it would be covered under the bot policy and require a WP:BRFA. The policy under which BRFAs are required does not apply to assisted editing (i.e. the use of software to create letters, numbers and symbols that were not produced by a human being pressing a keyboard). This is governed by existing policies (such as WP:MEATBOT and by the guideline at WP:ASSISTED. jp×g 09:28, 13 December 2022 (UTC)

@JPxG: The entire news field refers to ChatGPT as a chatbot. It is general consensus that it is a bot. ChatGPT speeds up writing, by doing it for (that is, assisting) the user, which falls under the "higher speed" and "assisting human editors" foci of the bot policy. There is a passage in the bot policy that covers policy contradictions (such as between the lead and definitions sections), and situations where the spirit of the rule and its precise wording conflict, that is, cases of ambiguity. In its definition of "Bot Approvals Group" (BAG), the bot policy states: "The BAG also determine the classification as bot or assisted editing, in ambiguous cases." According to WP:ASSISTED, it is up to the Bot Approvals Group to decide whether bot approval is necessary. Based on the previous 2 sentences, BAG decides whether use of particular software falls under its jurisdiction. It remains to be seen what BAG's reaction(s) to LLMs, and the chatbots built upon them, will be.    — The Transhumanist   11:10, 14 December 2022 (UTC)

I do not think you are properly acquainted with how this software works: like I said, there do not currently exist any language models capable of independently establishing API connections to Wikipedia and making edits without human interaction. No media outlet has ever claimed that ChatGPT falls under the English Wikipedia's definition of an automatic bot – and even if they did, they do not determine policy. It is true that WP:MEATBOT and WP:ASSISTED are part of the bot policy, but there is a very clear definition of what a "Wikipedia bot" is, and it's defined by that same policy. At any rate, all edits (whether made by bots, software, humans using software, aliens using software, or Nagato Yuki psionically connecting to Wikimedia servers) are governed by existing policies and guidelines. To specifically address LLM output, a new policy would need to be written and ratified (which I am currently drafting a proposal for). jp×g 11:26, 14 December 2022 (UTC)

@JPxG: I believe the bot policy has wider jurisdiction than the narrow interpretation that you have presented. Establishing API connections is irrelevant, because a human is inserting bot-generated content. It's a bot-involved process. And those are encompassed by the bot policy which makes it up to BAG. A new policy could establish an exception, and I imagine the discussions will be extensive, as this is not a cut and dried case -- it is a sensitive issue with many potential ramifications. But, until such a policy is in place, this issue falls under BAG's jurisdiction, since they are the ones who decide the classification of a software program as it pertains to the bot policy.    — The Transhumanist   11:52, 14 December 2022 (UTC)

(edit conflict) WP:ASSISTED is deliberately vague to not restrict use of common scripts and tools. So it specifically says that only once volume becomes significant, such editing becomes more likely to be treated like a bot and BAG can determine this. It doesn't make it a bot, but it will be treated like a bot. We've never encountered any large-scale edits with LLM before, but we sure have seen a lot of high-volume editing. Half the bot policy only exists because of all the ways editors have inadvertently created issues with mass edits. So at that point, other parts of the policy start to matter, notably WP:CONTEXTBOT - which does not allow edits where context matters. I'm not saying copy-pasting LLM output is immediately covered by bot policy, nor does it matter whether anyone considers LLM to be a "bot". But bot policy will kick in once someone starts to make a lot of edits. And any new guideline will have to reconcile with this or we need to change bot policy to reconcile with LLMs. —  HELLKNOWZ ∣ TALK 12:02, 14 December 2022 (UTC)

@JPxG: Another possible approach for vetting users for use of LLMs is via user group membership (aka "rights"). Here are our current user groups:

These indicate membership in user groups (see: user access-levels). They pertain to who is granted access to various features of MediaWiki and its extensions. Theoretically, a user group could be created without being attached to a program function (that part could just be left blank?). For example, you could have a group called "LLM", with everyone in that group approved to use large language models in their editing. I don't know if this is doable, though.    — The Transhumanist   08:01, 13 December 2022 (UTC)

I do not think there is anything in our P&G that would directly prohibit use of content created by LLMs, nor do I think it would be a good idea to try to do so. All that is needed is to continue to hold individual editors responsible for all edits they make, including the copying of content from any source, whether from LLMs or other sources. We probably should add language in appropriate places reiterating that editors are reponsible for insuring that all content that they add, including anything produced by an LLM, meets our P&G. - Donald Albury 13:12, 13 December 2022 (UTC)

@Donald Albury: LLMs automate writing (edits). The rules are very clear on this: it falls under WP's bot policy, in the very first sentence.[4]   Therefore, it would require a new policy to allow use of LLMs without need for approval from the Bot Approvals Group (BAG).    — The Transhumanist   09:24, 14 December 2022 (UTC)

If someone uses an unapproved script or bot to edit WP, that is a violation of the bot policy, whether or not they use an LLM to generate any content being added. If someone uses an LLM to create text which they then copy into Wikipedia without using a an unapproved script or bot, that is not covered by the bot policy, but the user remains responsible for insuring that the content conforms with policy and guidelines. There is no point in banning content created by LLMs, as we already require that content be verifiable from reliable sources, and I doubt we will be accepting any content created by an LLM as a reliable source anytime soon. The danger is that LLMs may create potential content with citations to pseudo-sources, but we can go after users repeatedly adding such content to WP for abusing the policies on verifiability and reliable sources, without regard to whether such content came from an LLM. Donald Albury 13:48, 14 December 2022 (UTC)

I don't think it's plausible that LLMs are covered by the bot policy. If they were, grammar checkers, spell checkers, and machine translation would be "bots". Jahaza (talk) 19:49, 14 December 2022 (UTC)

@The Transhumanist: ChatGPT falls under Wikipedia:Bot policy, but per the definitions section it does not fall within that policy's definition of a bot. Rather, use of it would fall under the definition of "assisted or semi-automated editing", and the relevant policy section is Wikipedia:Bot policy#Assisted editing guidelines. The section doesn't aim to draw a 100% hard line, but my reading is that limited of ChatGPT for clean-up on a limited number of articles by a user in a limited closely-supervised way may be something users can do if they are trusted to apply their common sense. It is "Contributors intending to make a large number of assisted edits" who "are advised to first ensure that there is a clear consensus that such edits are desired." Limited use of ChatGPT to a lesser degree than would trigger this may currently be outside policy. In any event "A bot account should not be used for assisted editing".

It seems to me that an addition to the policy along the lines suggested by User:JPxG to address this potential hole might well be useful, eg "tools capable of assisting editors make substantial edits (for example large language model output) should only be used by competent editors who do not blindly paste tool output into the edit window and press "save"." Jheald (talk) 19:36, 18 December 2022 (UTC)

In what way does the bot policy cover ChatGPT? Just because it is called a "chatbot", doesn't mean it is a bot. Copying text from GPT-3 doesn't automatically become bot-like editing. Semi-automated edits? i'd call that borderline. It only becomes a problem (e.g. meatbot problems) if the amount of supervision needed to save an edit is below normal editing, and that the speed of the edits are above normal. (see awb, huggle, etc) 0xDeadbeef→∞ (talk to me) 10:20, 20 December 2022 (UTC)

The thing is, any LLM additions will inevitably be both faster than writing manually and, due to its confidently-wrong output, less reviewed. Otherwise, why would anyone bother with it? I feel that assuming that editors will spend just as much time to carefully review the LLM output is wishful thinking. I'd like to be proven wrong, but I have never seen any precedent on Wikipedia that better tools would lead editors to spend the time saved to further verify the tool output. If anything, tools only create induced demand. —  HELLKNOWZ ∣ TALK 21:18, 20 December 2022 (UTC)

I don't think we need to do anything in particular. There is plenty of confidently-wrong content being added to Wikipedia by human editors already and we're dealing with that as well as we can. I think the intersection of 'Editors who will use a cutting-edge AI to generate content' and 'Editors who will do this without validating the output' is a very small overlap and will be of such small volume to be picked up by other editors as usual. A huge influx will be detected in aggregate, and we can deal with that if it becomes a problem in the future. If someone uses LLM to generate confidently-right content or articles, that's indistinguishable from content generated by a competent human, I refer you to xkcd: Constructive! A simple but unobtrusive first step may be to tag an edit as 'generated by AI', or maybe just ask editors to add a tag to their user pages if they regularly do so, but the intersection of problematic users who also follow this would be basically non-existent. JeffUK (talk) 10:40, 24 December 2022 (UTC)

@JeffUK, Hellknowz, 0xDeadbeef, Jheald, Jahaza, Donald Albury, JPxG, and AndyTheGrump:

So, wait until after it becomes a huge influx/problem, and only start to deal with it then? What if a solution takes weeks or months to develop?

By the way, what might the solution be for a huge influx of LLM-generated content, and how long would such a fix likely take?    — The Transhumanist   11:32, 2 January 2023 (UTC)

I am having trouble understanding what you are talking about at this point. I wrote WP:LLM some weeks ago, a gigantic proposal for a comprehensive guideline on the use of LLMs, and linked it multiple times on this noticeboard. While it is not complete, it seems to me like it covers everything you are talking about here. Do you have an opinion on it at all, or...? jp×g 15:39, 2 January 2023 (UTC)

@JPxG: I was responding to JeffUK's statement "A huge influx will be detected in aggregate, and we can deal with that if it becomes a problem in the future." Intently waiting until something becomes a huge problem before you start dealing with it, sounds like a disaster waiting to happen. Also, what good are guidelines going to do if the average person is using chatbots on a regular basis? People just jump in and edit Wikipedia without reading any project-level pages first. If there's a huge influx, and all you are doing is holding up a sign that says "Read this", what good will that do? You haven't addressed how the problems associated with a potential huge amount of chatbox input (in the form of one-off edits from a large number of people) would be prevented or processed. One solution is to fix the chatbots themselves, so that they don't generate Wikipedia-incompatible content in the first place, which would require working with the developers. A second method would be to create bots to detect and remove either chatbot-generated content, or if possible, policy-breaking content. Simply writing policy and hoping no flood comes, just doesn't seem like a viable approach should a flood hit. That approach may work for the first 3 or 4 years, but what if the flood comes in the 5th year and Wikipedia isn't prepared? We will have wasted 5 years that could have been spent preparing. Maybe we'll be lucky and chatbots will be smart enough to read and follow your guidelines. But if they are not? Fortunately, Doug Weller has passed word along to the Wikimedia Foundation. Maybe they will do something other than write editing guidelines.    — The Transhumanist   03:09, 10 January 2023 (UTC)

@JPxG although I haven't read more than the beginning, I'm also worried about AIs creating images.For instance I've seen some extremely convincing ones of fake archaeological sites and artefacts. Couldn't people pass them off as their own photos? Or am I missing something? Doug Weller talk 17:06, 2 January 2023 (UTC)

Yeah, that's a whole other deal. This proposal is only for large language models. Large image models will probably need to be governed by something much more imaginative. jp×g 17:11, 2 January 2023 (UTC)

If someone's editing is inappropriate, the solution will be notifying them it's inappropriate, warning them, then banning them if they don't stop. There are ways for incompetent editors to make massive plausible seeming changes to the encyclopaedia right now. e.g. by copy/pasting content from other places, or just writing in made up 'facts', LLM really won't make this any easier for someone who's intent on doing this. JeffUK 18:09, 2 January 2023 (UTC)

I'm kind of wondering what in the heck we're standing to gain by creating any sort of policy surrounding ChatGPT and its ilk. 🌈WaltCip-(talk) 13:24, 9 January 2023 (UTC)

If these AIs are used in some way for a large scale creation of articles, I think that will be a disincentive for a lot of editors and may drive some away. I disagree with JeffUK on the simplicity of dealing with this. First, you need to be able to spot them and that's work. Secondly, that also assumes that the numbers will be small. Doug Weller talk 15:48, 9 January 2023 (UTC)

@WaltCip: None. Most editors don't read policy before editing. That's because most editors post very infrequently. But there are a lot of them, and they have authored of most of Wikipedia. What happens when they are all using chatbots, much in the way that most everyone today uses Google?    — The Transhumanist   03:34, 10 January 2023 (UTC)

"Someone." Apparently, you are assuming it will be one person, or a small enough number to be handled manually. But, what if over the next few years chatbots become ubiquitous with almost everybody using them? How will you deal with it when half the content contributions to Wikipedia are being generated using chatbots?    — The Transhumanist   03:26, 10 January 2023 (UTC)

Develop software to detect it?

Someone seems to have done this, see A college student created an app that can tell whether AI wrote an essay Maybe the WMF should look into software detection of AI material? Doug Weller talk 14:45, 10 January 2023 (UTC)

We have mw:ORES that uses machine learning to detect vandalism, so the infrastructure is already in place. All we need to do now is to add the dataset. CactiStaccingCrane (talk) 16:06, 12 January 2023 (UTC)

(ORES is used for these "likely have problems" and "likely bad faith" highlights in Special:RecentChanges) CactiStaccingCrane (talk) 16:08, 12 January 2023 (UTC)

Even if ORES is up to the task, and it isn’t perfect now, you still need enough editors to deal with large numbers. Doug Weller talk 18:21, 12 January 2023 (UTC)

@Doug Weller, or a bot. — Qwerfjkltalk 21:02, 12 January 2023 (UTC)

@Qwerfjkl: What would the bot do?    — The Transhumanist   22:38, 12 January 2023 (UTC)

@The Transhumanist, revert additions and/or tag articles. — Qwerfjkltalk 07:03, 13 January 2023 (UTC)

Personally, I'd prefer that if this is implemented, ClueBot prefer to tag before a full-on reversion, except in cases. I think that ClueBot should use the same exceptions for 3RR for reverting possible bot stuff, though this might require that chatbots be banned from BLP-covered articles. Everything else which could be problematic but not a 3RR exception could be placed at a specific page...maybe have a specific preset filter on recent changes? InvadingInvader (userpage, talk) 07:43, 6 February 2023 (UTC)

@CactiStaccingCrane, Doug Weller, and Qwerfjkl: All we need to do is add what data set? You make it sound easy (keeping fingers crossed). What does that entail?    — The Transhumanist   22:38, 12 January 2023 (UTC)

OpenAI have annouced they are adding in some kind of lexical watermark than can be used to identify any output from ChatGPT. scope_creep^Talk 13:08, 16 January 2023 (UTC)

@Scope creep, currently the "Overall," beginning the concluding paragraph is watermark enough. — Qwerfjkltalk 20:36, 17 January 2023 (UTC)

For example, see the edit linked in this comment. — Qwerfjkltalk 20:42, 17 January 2023 (UTC)

Other inherent problems only partially touched on

Other inherent problems only partially touched on:

• Editing articles involves also understanding what is already in the article and how it is organized plus understanding and interpreting policies and guidelines.
• What's unspoken but runs through many things including current Wikipedia is sort of a commensurate investment. You can get volunteers to take their time to review and deal with issues because they know they are dealing with something that an editor has invested time in to create. Part of the reason that we don't allow mass creation of articles by bots. In other words, we'd significanly lose volunteer efforts
• Modern AI is inherently unaccountable black boxes. There is no way to see or interrogate or demand/recieve accountability or reasoning for how it arrived at what it arrived at.
• If gibberish or semi-gibberish is created, it normally requires an expert to spot and remove it..... a very scarce resource. I once uncovered a set of technical-subject articles (about 100 article as I recall) which looked very technical and Wikipedian and were sourced but if you knew the subject you knew were pure gibberish.

Sincerely, North8000 (talk) 22:23, 10 January 2023 (UTC)

Thanks. I agree entirely. Doug Weller talk 09:04, 11 January 2023 (UTC)

Well said. I think that folks are overestimating the ability of our review processes to detect "vaguely plausible bullshit" - it's not very common for human editors to fill in blanks with made-up facts and numbers, and I'm not sure that AfC or NPP are checking for this as it would greatly increase their workload. –dlthewave ☎ 19:42, 11 January 2023 (UTC)

[I]t's not very common for human editors to fill in blanks with made-up facts and numbers. Maybe not when adding content, but I see this happen all too often in edits to temperature tables in climate sections. Of course, the tell there is changing temperatures without citing a source or commenting about correcting from a cited source. - Donald Albury 20:17, 11 January 2023 (UTC)

As I recall, the big one that I caught looked like some type of expose project or research project to see if such a scam could get far in Wikipedia. It was sort of in mashup of words from actual sources. Total nonsense, but a typical reader might think it was simply over their head. North8000 (talk) 21:25, 12 January 2023 (UTC)

@North8000, Doug Weller, Dlthewave, and Donald Albury:

In answer to the 3rd point above (the black box issue), Perplexity.ai, an AI search engine with a chatbox interface, provides source references with its answers. That is, the references are the search results, while the answer provided is compiled or interpreted from those web pages. So, at least the sources can be checked for verification. But, there are still problems with it. See the perplexity.ai section below.    — The Transhumanist   19:56, 21 January 2023 (UTC)

Summary of discussion so far

@Aquillion, Andrew Gray, Fram, Levivich, Ritchie333, 0xDeadbeef, ONUnicorn, JPxG, EpicPupper, Sojourner in the earth, Dlthewave, Doug Weller, Qwerfjkl, CactiStaccingCrane, WaltCip, JeffUK, Hellknowz, Zero0000, AndyTheGrump, Bluerasberry, David10244, Boud, Ziko, Pharos, Andrew Gray, WhatamIdoing, Tazerdadog, Barkeep49, Tigraan, Blueboar, MJL, PerfectSoundWhatever, Koziarke, SmallJarsWithGreenLabels, Isaacl, Lee Vilenski, Thebiguglyalien, Hanif Al Husaini, and Xeno:

Highlights of the discussion so far:

• Chat-GPT is taking the world by storm (translation: it has gone viral).
• Chat-GPT, and other LLM-based chatbots, can generate compositions, some good enough to pass as college-level essays.
• Wikipedia is included in the corpus (training data) of Chat-GPT (and other chatbots).
• Such software has the potential to be used for:
  • Generating Wikipedia content, including writing new articles and adding new material to existing articles.
  • Generating Wikipedia policy content.
  • Generating discussion content, such as on policy talk pages. That is, editors using it to write their discussion replies for them.
  • Editing articles, including rewrites, and using chatbots as a grammar checker.
  • Editing other namespace pages, such as policy pages, etc.
  • "Can be used for an extremely wide variety of tasks across the project, from formatting to table syntax to HTML generation to copyediting." (quoting JPxG)
  • Creating hoaxes with less effort.
• Most Chat-GPT output lacks citations.
• Some experiments were run, showing that Chat-GPT:
  • Copies writing styles very well.
  • Has a tendency to make things up, yet presents it as fact in an encyclopedic tone. One editor dubbed this "confident nonsense". In one experiment, Chat-GPT created an article reporting that Wikipedia's own Signpost newsletter was the recipient of several Pulitzer Prizes.
  • Can include references, but some of the references were made up and totally fictitious.
  • Some references cited Wikipedia (an ineligible source for Wikipedia articles).
  • One of the experiments generated instructional content, a recipe, that the user followed, and ate the results of.
  • Another experiment used Chat-GPT to answer hypothetical questions in the style of WP's teahouse department. It worked fairly well.
  • Yet another experiment created a sample policy page, showing that chatbots are not limited to editing articles. They can generate or edit pretty much any type of page on Wikipedia, except files (images).
  • Chat-GPT output is not fact-checked.
  • Chat bots don't actually understand what they are writing.
  • When used responsibly as a tool, with editors carefully prompting the chatbot, and editing and fact checking its output before posting it to Wikipedia, a chatbot can be very useful and increase editor productivity: the LLM GPT-3 was successfully used to create department reports for Wikipedia's newsletter, The Signpost.
  • JPxG conducted an experiment/demonstration to show that Chat-GPT is a sophisticated interactive editing tool, which you tell it what you want it to do to a textual work, and then it does it. See it here: User:JPxG/LLM demonstration.
• It was pointed out that Wikipedia policy already covers all contributions, whether generated by chatbot or human. Ultimately, the user is responsible for material they copy and paste into Wikipedia.
• Issues of concern that were raised include:
  • Users copying chatbot-generated text into Wikipedia without carefully editing and fact-checking it first.
  • Confident nonsense (misinformation generated by chatbot) may be hard to spot.
  • The potential of chatbots to violate copyright, by directly copying, or generating text based on, copyrighted works.
  • Violating Wikipedia's licenses, most notably the attribution requirements. Chat-GPT output generally does not include attributions.
  • A chatbot-edited Wikipedia could wind up in the training data for those same chatbots (or their next versions), creating a potentially error-compounding feedback loop.
  • The suggestion was made to prepare for a potentially large future increase in chatbot entries to Wikipedia, by:
    • Working with chatbot developers to make chatbot-generated output Wikipedia compatible.
    • Develop bots to identify and process chatbot entries.
• No consensus has emerged on what the Wikipedia community should do about LLMs/chatbots. Some editors think that policies/guidelines and the current editor pool could handle any influx of chatbot generated edits. Some other users were concerned that there is potential for LLM/chatbot contributions, such as one-off edits by members of the general population, to overwhelm our pool of editors. One user pointed out that it may take experts to discern nonsense articles, and experts on Wikipedia are a scarce resource.
• Consensus did emerge on something not to do. It was agreed that banning chatbot-generated content was not a good idea at this time, and probably wouldn't work anyways.
• Software has been developed to identify Chat-GPT-generated text.
• It appears some editors may take the initiative to prepare for a worst-case scenario (chatbot input going beyond our editor pool's ability to handle), and discussion on how to do this has begun.
  • WP:ORES could theoretically be trained to identify chatbot edits.
• The Wikimedia Foundation has been contacted about the concern over LLMs/chatbots, presenting a contact there with a link to this and a previous discussion.

Did I miss anything?    — The Transhumanist   01:22, 13 January 2023 (UTC)

AI-generated images are rapidly becoming a Big Thing, so it is not correct to exclude them. Also, "Wikipedia policy already covers all contributions, whether generated by chatbot or human" is misleading as it is true only by accident. A more precise description would be "Wikipedia policy was written without any consideration of chatbots". Zero^talk 03:28, 13 January 2023 (UTC)

I'm concerned about "Such software has the potential to be used for: creating content farms that good-faith human editors, including existing experienced editors, will sometimes mistake for reliable sources when they are writing content".

Also, the statement that "Software has been developed to identify Chat-GPT-generated text" is true, but not relevant for very short contributions. Some of this is using sentence length, and you won't be able to identify an abnormal sentence length if you only look at two or three sentences. WhatamIdoing (talk) 05:33, 13 January 2023 (UTC)

Anyway, detection methods that work now won't work in the next generation. Eventually (and not far in the future) distinguishing between human-written and computer-written prose will be impossible for practical purposes. This is going to be the greatest threat to Wikipedia since its founding. Zero^talk 05:51, 13 January 2023 (UTC)

@Zero0000: When do you suppose the impossible-to-distinguish scenario will be here? Two years? Less?    — The Transhumanist   13:42, 25 January 2023 (UTC)

I have started Category:Wikipedia essays about artificial intelligence, Perhaps folks here would like to add to the collection, and document yet more thoroughly! Pharos (talk) 01:07, 14 January 2023 (UTC)

Adjacent to hoaxes there's also the likelihood of spammers using GPT to bulk out their edits. I strongly suspect that the text of this edit today, writing repetitively about a static sculpture as if it was a functional scientific instrument, was generated with GPT-3, probably giving it a prompt to explain Orbital Reflector in terms of dark matter and black holes, the subject of the two embedded spam links. Belbury (talk) 11:39, 17 January 2023 (UTC)

I just ran that through GPT-2 output detector and it estimated 99.97% chance that that passage was AI-generated. — rsjaffe 🗣️ 03:31, 24 January 2023 (UTC)

Chat-GPT spreading fast

The situation is changing rapidly:

Chat-GPT may become ubiquitous sooner than previously thought, and so far, identification methods have fallen flat...

Here's some recent news:

1. ChatGPT Will Be Everywhere in 2023 (CNET)
2. Microsoft is reportedly integrating ChatGPT's technology into Bing (Yahoo)
3. Microsoft is looking at OpenAI’s GPT for Word, Outlook, and PowerPoint (The Verge)
4. There's a Problem With That App That Detects GPT-Written Text: It's Not Very Accurate (Futurism.com)

With the user base for Chat-GPT about to explode, the potential for Chat-GPT-generated text being added to Wikipedia will explode right along with it. It's looking uncertain whether or not Wikipedia's editor community will be able to keep up with the influx. In light of recent events, what should be done about this?    — The Transhumanist   03:21, 13 January 2023 (UTC)

As well as being able to write plausible-looking prose on any subject, computers can also be programmed to add it to Wikipedia all by themselves. The first task is to absolutely ban computers from editing, with the sole exception of authorized bots. The second task is to add to appropriate policy pages that all content (authorized bots excepted) must be added by a human and that that human is responsible for checking policy conformance of the content. Zero^talk 08:06, 13 January 2023 (UTC)

I couldn’t agree more. Does anyone have objections? Doug Weller talk 10:45, 14 January 2023 (UTC)

I was hoping to get more feedback on WP:LLM from having posted it here, but either way, I think it is pretty close to ready for consideration as a guideline (or policy, as appropriate)... based on the conversations I've had (and seen) I am prepared to write an RfC for its adoption. jp×g 11:00, 14 January 2023 (UTC)

@JPxG: The guideline is not ready. It is not where near complete, and it needs a rewrite. Here are some proofreading notes:

It's way too redundant, repeating policies and itself, without explaining how to get the job done. Aside from the "fit for" sections, the rest of the page can be reduced to a single paragraph.

It presents -- should only be used by competent editors who do not indiscriminately paste LLM output into the edit window and press "save" -- four times! Someone who is incompetent isn't going to be able to judge whether or not they are. Also, "indiscriminately" is vague. That entire sentence should be removed.

Editors need to know what they need to do to the text before they can press "save". For example, you alluded to a manner of using LLMs in compliance with WP copyright policy, but you didn't explain how. How can an editor be sure that an LLM-generated piece doesn't violate someone's copyrights? What's the procedure?

Rather than covering "good fit" and "not good fit", the guideline should present explicit instructions: "Use it for this" and "Do not use it for this". And then explain how.

I hope you find these observations and comments helpful. Sincerely,    — The Transhumanist   08:04, 16 January 2023 (UTC)

I agree that the page is not finished, but I don't really know what you are objecting to here. It kind of sounds like you are inventing problems – if users don't know how to check if things are true before putting them into Wikipedia articles, they shouldn't be editing at all. If users don't understand what copyrighted material is, they need to read Wikipedia:Copyright policy, which is linked to from this page when it's mentioned. That is an explanation of how to get the job done. It should not be necessary to create an exact copy of Wikipedia:Verifiability that says "When using a LLM," at the beginning of every sentence. jp×g 08:29, 16 January 2023 (UTC)

@JPxG: How can users understand what the copyrights of Chat-GPT's output are? Chat-GPT doesn't provide sources, nor does it report if it copied or derived the passage from a particular work. So, how do you go about checking whether or not a particular Chat-GPT response is in violation of copyright, so that "pasting its output into the edit window and pressing 'save'" is not considered "indiscriminate"? Also, it isn't clear who owns the copyrights to the output of an LLM: the public domain, the owner of the LLM, the user of the LLM, or the owners of the copyrights of the works included in the training data set? The breadth of this problem is discussed in #Copyright status below.    — The Transhumanist   00:08, 19 January 2023 (UTC)

There already exist a very large number of policies about copyrighted text. Editors are subject to these policies. These policies contain information on how to avoid copyright violations. If you asked GPT-3 to tell you the lyrics to Moonage Daydream, they would be copyrighted. If you found the same lyrics by typing "moonage daydream lyrics" into Google, they would be copyrighted. What is the difference? Policies do not (and cannot) cover every hypothetical person and situation to which they could be applicable: we do not have a separate WP:COPYRIGHT for old editors, WP:COPYRIGHT for young editors, WP:COPYRIGHT for male editors, or WP:COPYRIGHT for female editors. WP:COPYRIGHT applies to all editors regardless of their age, race, gender, or whether they are human or machine. I don't know how to explain this in further detail. jp×g 01:06, 19 January 2023 (UTC)

@JPxG:

You've missed the points entirely (there were two, and you only replied to one).

Those policies you mentioned do not deal with the copyright problems presented by black box chatbots, nor do they warn about the dangers of pasting in chatbot output.

Search engine search results are excerpts from web pages that the search results identify — which facilitates verification. Chat-GPT and other black box chatbots answer questions in natural language, without telling the asker of the question where the information came from — which does not facilitate verification — while presenting it in a very confident and scholarly tone.

This may result in a great deal of misinformation being posted to Wikipedia, where it will sit until somebody else removes it. The delay between those 2 events can be lengthy, especially for material that seems plausible. So, it might be a good idea to provide guidance specific to chatbot usage pertaining to copyrights -- at least some caveats on which chatbots to avoid.

Another problem is that we don't know where the training data came from. There could be deep web data in there as well. That can't be easily accessed to check for plagiarism. So, is it a good idea to use blackbox chatbots? There are transparent surface web chatbots that include references for verification, so maybe we should recommend that the blackbox ones be avoided.

Now, for the second issue (the one that you skipped): WP policies do not cover prompting a chatbot to write material. The copyrights to material that is written by a chatbot is owned by who? The user? That has not yet been established! What stance is going to be taken by Wikipedia, and what guidance are we going to provide on this issue?    — The Transhumanist   09:17, 25 January 2023 (UTC)

I feel like either you are not reading what I'm saying, or we have some kind of insurmountable disagreement about what letters and words are. jp×g 09:26, 25 January 2023 (UTC)

@JPxG:

You've just fed me a variation of "you're not listening", with a little barb attached to the end. Really? That's who you are? I'm disappointed.

I read what you wrote, and I highly disagree with what you are saying...

You are saying that current copyright policy is enough: it prohibits copyrighted works from being posted to Wikipedia without the permission of the copyright holder, and that it is up to the editor to make sure that the material does not violate anyone's copyrights or Wikipedia's copyright policies.

My positions are...

1) that black box chatbots pose the danger of luring editors into violating copyright policy, that we may be faced with a deluge of copyright-violating derivative material because of it, and that some additional guidance would be appropriate: Like avoiding black box chatbots in favor of transparent ones, and...

2) that the copyrights to the natural language output composed by chatbots is unclear — what is clear is that the editor didn't write it. Since the editor didn't write it, does that mean that the editor does not own the copyrights to it? And if editors don't own the copyrights, should they be giving it to Wikipedia? Wikipedia should form a stance on the copyrights of chatbot-generated-output and present editors with guidance on this issue as well.

You have apparently been avoiding replying to those positions, and so my guess is that you are opposed to them. I strongly oppose the let's-stick-our-heads-in-the-sand approach that you support.    — The Transhumanist   10:55, 25 January 2023 (UTC)

P.S.: I don't think anything in writing is going to be enough. I expect that it will take software programming to deal with the problems Wikipedia will be subjected to by chatbot compositions. And that is beyond the scope of this venue. ;)    — The Transhumanist   11:31, 25 January 2023 (UTC)

To act or not to act

Like DALL-E last year, or NFTs the year before that. I'll believe it when I see it, and I can't see the value in spending even more time discussing a hypothetical future threat to Wikipedia. – Joe (talk) 08:39, 13 January 2023 (UTC)

The scariest yet most plausible thing is that this is happening with some of the articles but we aren't aware of it. I don't think raising awareness on this issue is a bad thing given how fast AI advances nowadays. CactiStaccingCrane (talk) 00:39, 14 January 2023 (UTC)

Sorry, I missed the "recent events". Where is the evidence for GPT problems on Wikipedia? —Kusma (talk) 11:57, 13 January 2023 (UTC)

Simple way to defeat these AIs: train them on how humans argue about WP policy… then ask them whether AIs are reliable (pro and con)… then set them against each other on a dedicated talk page. While they argue, we can continue editing. Blueboar (talk) 01:58, 14 January 2023 (UTC)

See also https://openai.com/blog/debate/ CactiStaccingCrane (talk) 11:41, 15 January 2023 (UTC)

Where's the evidence that Wikipedia can't cope with AI generated articles? doktorb _words^deeds 14:07, 14 January 2023 (UTC)

@Doktorbuk: You are talking in terms of hindsight (asking to see what has already happened), rather than applying foresight to assess a potential threat by asking "What could happen?"

Here's an article from the New York Times -- imagine a similar effort directed at Wikipedia using thousands upon thousands of (seasoned) new accounts to support political POVs, revise history, censor opposing opinions, and spread other forms of misinformation:

https://www.nytimes.com/2023/01/15/opinion/ai-chatgpt-lobbying-democracy.html

It's only a matter of time before the powers that be shift their attention, and their tools, upon the English Wikipedia. The question is, are we ready for when we have to be? Here's an article that makes one wonder what these people will do now that they have Chat-GPT to work with:

https://www.theweek.in/news/world/2023/01/06/saudi-arabia-infiltrated-wikipedia-and-this-is-how-they-did-it.html

So, do we really need evidence that the English Wikipedia has already been breached by LLM-assisted POVers before proceeding? Or can we prepare for this in advance?    — The Transhumanist   00:34, 16 January 2023 (UTC)

Your reasoning seems to be that

1. ChatGPT (or its equivalents) can write disinformation quickly and cheaply
2. POV-pushers (governments, lobbies etc.) are currently limited by the time humans need to write disinformation
3. Wikipedia is a prime target for such POV-pushers
4. Therefore, ChatGPT (or its equivalents) will flood the gates, unless we do something.

I will grant you (1) is either already true or will likely be in the near future.

However, (2) is questionable (see that XKCD about old-fashioned human-crafted POV-pushing). I would guess coordinating the messaging and maintaining the disinformation is a much larger fraction of the costs than actually writing the text.

(3) is also dubious. Editing in a way that sticks is much harder on Wikipedia than in other places (such as facebook, reddit, etc.). Maybe it has more impact, but the cost-benefit analysis is not obvious.

Finally, inaction is always an option. It might not be a good option, it might even be the worst option, but it must be compared to other specific measures. "Something must be done" without specifics is just the politician's fallacy. In the absence of details about the threat, it’s hard to compare the possible countermeasures. Tigraan^{Click here for my talk page ("private" contact)} 16:41, 16 January 2023 (UTC)

@Tigraan, I think your list of assumptions is missing "5. People who want to corrupt Wikipedia (e.g., NPOV violations, stacking votes) can reasonably be expected to obey any prohibitions we announce on using this particular technology to achieve their illicit ends." WhatamIdoing (talk) 21:22, 17 January 2023 (UTC)

I wanted the list of assumptions to be a reasonable summary of (what I understand to be) TH’s argument; I suspect your suggestion is... not that. But I agree that’s part of the problem (which my last paragraph covers). Tigraan^{Click here for my talk page ("private" contact)} 10:37, 18 January 2023 (UTC)

Dear @Tigraan: I am concerned with a flood of WP-incompatible chatbot-generated content, whether by POV'ers or good-faith editors. But it won't be at any gates. The water-level will simply rise. If and when flooding begins, it will be a matter of bailing out the excess. There are three questions relevant to such potential flooding:

1) Will chatbots be designed in such a way to prevent flooding (and bailing) in the first place by minimizing inappropriate (unsourced, misinforming) content?

2) Will the bailing be automated?

3) Shall we wait to work on #1 & #2 until after flooding has begun, or prepare in advance?

Some editors seem doubtful that the addition of content generated by LLMs to Wikipedia beyond the manual capacity of our editors to process it will happen. And I don't know if it will happen, either. But, there is a continuous stream of strong indications that LLM-based tools will become ubiquitous in the not too distant future, for general use, which, by extension, includes using them to add content to Wikipedia. Here's another:

Google Calls In Help From Larry Page and Sergey Brin for A.I. Fight — New York Times]

And the technology push isn't limited to OpenAI and Google. Here's a search engine that uses a natural-language interface in both its queries and its answers:

Perplexity AI: Ask Anything

It is looking pretty clear that some major changes are on the horizon in the way computer users will be composing web content. It is also profoundly obvious that Wikipedia isn't ready right now for much more than the current volume of content creation that it is already handling. Maybe the volume won't increase by much, or maybe it will.

However, some editors are taking the potentiality that it will seriously, and it'll be interesting so see if their preparation efforts will be sufficient to stem the tide, if or when the tide rises. Sincerely,    — The Transhumanist   22:51, 20 January 2023 (UTC)

@The Transhumanist, I'm not sure that these questions are really suitable to a discussion on Wikipedia. The first one, for example: Will chatbots be designed in such a way to prevent flooding (and bailing) in the first place by minimizing inappropriate (unsourced, misinforming) content?

I'd re-phrase it like this:

"Will all of the people who are not us, including those who don't care about us, carefully design their software in such a way to be convenient for us?"

Answer: No. Or, at least, it is highly unreasonable to assume that the answer is yes for all of the people who write this sort of software, and it only takes one to risk a problem. WhatamIdoing (talk) 21:21, 23 January 2023 (UTC)

@WhatamIdoing:

That would more likely be true if members of WP or the WMF did not contact them. They are not producing these things in a vacuum. WP/WMF has a good relationship with Google, for example, which applies Wikipedia content extensively. It may be time to reach out to the companies developing chatbots too.

On the bright side, there's pressure coming from abroad, in the critique of chatbots, to be less "black box" and to provide references, which is one of the features that would help avoid problems.

Perplexity.ai already provides sources, which helps with verification efforts, and to see which ones are and are not from Wikipedia. Though, Perplexity.ai does not provide quote marks around passages that it quotes, and that is another problem. So, I guess they need to be contacted as well.

It looks very likely that chatbots will be used to compose content for other websites besides Wikipedia, and that their webpages may be included in chatbot training data too -- making an error-magnifying feedback loop a potentially huge problem for the chatbots. Too big to go unnoticed, hopefully.

It's important that we are aware of these issues if we are to have any chance in influencing solutions. Who knows, the chatbots, and/or the chatbot developers, may actually read this discussion. ;)

The WMF has been made aware of this discussion, so they can read it to prepare for discussions with participants in the chatbot sector. So, it is important that we get our concerns, and thoughts on design and strategy, in print.    — The Transhumanist   08:20, 25 January 2023 (UTC)

You are assuming that LLMs are being developed by a manageable number of identifiable companies, and hoping that all of them would like to protect Wikipedia.

But let's consider it from a different POV. Imagine that chatbot software is an open-source project, like Wikipedia. You have the Wikipedia:Right to fork open source projects – not just Wikipedia, but any open-source project. Anyone can add or subtract anything on their own setup. For example, if someone adds a "protect Wikipedia" module, then the next person could remove that, or even add a "scam Wikipedia" module.

I believe there will be some organizations who find that protecting Wikipedia aligns with their interests, and they will do so. But there will also be some organizations who find that protecting Wikipedia is exactly the opposite of their interests, e.g., content farms that hope they'll be cited as sources here so that their ad-filled webpages will get more traffic, and WP:UPE scammers who are hoping to reduce their costs by having their secret, internal-use-only chatbot write Wikipedia articles for clients of dubious notability, rather than paying a human to do that. I don't think that we can identify such actors, and I don't think they would change their behavior even if we talked to them.

On a tangent, the call for chatbots to cite sources and add quotation marks is probably based on a misunderstanding. LLMs aren't "quoting sources". They're predicting what a typical way to complete a sentence might be. If it spits out "The journey of a thousand miles begins with one step", it's not quoting Lao Tzu; it's saying "When I look in my database, and I see phrases that start with 'The journey of', the next bit is usually either 'a thousand miles' or 'a lifetime'. I'll pick one and see what comes next." WhatamIdoing (talk) 17:04, 25 January 2023 (UTC)

@WhatamIdoing: Good point concerning the ecosystem of chatbot developers - I was only considering the big name actors (Google, etc.), but anyone and their uncle can get involved. You are right, bad actors are inevitable and perhaps even rampant. Yet, the vast majority of chatbot use will likely be of the big name models (ChatGPT, etc.). So, contacting and working with them would be beneficial.

As for quoting, I have found that the AI search engine perplexity.ai, which includes inline source references in its natural language answers to users' questions, integrates passages verbatim from the referenced webpages into its answers without using quotation marks.    — The Transhumanist   09:32, 26 January 2023 (UTC)

I wonder what the Ask Jeeves developers are thinking about that. Perhaps they were just 20 years too soon. WhatamIdoing (talk) 17:46, 26 January 2023 (UTC)

That’s a blast from the past, Doug Weller talk 18:30, 26 January 2023 (UTC)

Looks like some spammers and malware distributors have embraced this technology:

• https://research.checkpoint.com/2023/opwnai-cybercriminals-starting-to-use-chatgpt/
• https://www.csoonline.com/article/3685368/study-shows-attackers-can-use-chatgpt-to-significantly-enhance-phishing-and-bec-scams.html

WhatamIdoing (talk) 01:09, 28 January 2023 (UTC)

The second source above says "Likewise, anyone who uses the web to spread scams, fake news or misinformation in general may have an interest in a tool that creates credible, possibly even compelling, text at super-human speeds." We need detection tools, and fast. The "super-human speed" part could be a dead giveaway.    — The Transhumanist   10:00, 28 January 2023 (UTC)

Identifying chatbot-generated text

Zero0000's post is a good start. A simple way to crystalize the situation is to ask the human editor for their rationale for a particular phrase. North8000 (talk) 03:55, 14 January 2023 (UTC)

I feel like additions of large, overly-verbose unsourced text are something of a giveaway. See, for example, the first revision of Artwork title, written by ChatGPT. — Qwerfjkltalk 11:06, 14 January 2023 (UTC)

You can be confident that any giveaways are on the chatbot writer's list of things to fix in the next generation. Zero^talk 11:56, 14 January 2023 (UTC)

They should also fix the problem of the chatbots making stuff up. Someone should ask the chatbot writers to turn off the poetry and fiction generation algorithms, and any other algorithms that make things up, when the chatbots are composing expository text. Or add new algorithms to handle expository writing. Just the facts. And sources.    — The Transhumanist   00:42, 16 January 2023 (UTC)

Nature just published a piece about use of ChatGPT in scientific articles. Zero^talk 01:18, 21 January 2023 (UTC)

Copyright status

Is someone clear on what the copyright status of texts produced by LLMs is? From what I get, they may be considered derivative works from the dataset they were trained on. From [5]: As a result of the human authorship standard, “under U.S. current law, an AI-created work is likely either (1) a public domain work immediately upon creation and without a copyright owner capable of asserting rights or (2) a derivative work of the materials the AI tool was exposed to during training,” Esquenet continues. “Who owns the rights in such a derivative would likely be dependent on various issues, including where the dataset for training the AI tool originated, who, if anyone, owns the training dataset (or its individual components), and the level of similarity between any particular work in the training set and the AI work.” If they are derivative works then they cannot be published on Wikipedia just like this. Do we have more information on this? For example, does OpenAI specify somewhere the copyright status of the text produced by ChatGPT? Phlsph7 (talk) 09:09, 16 January 2023 (UTC)

The first question is whether a generated text that closely resembles an item from the training set is copyright infringement of that item. For instance, Microsoft Copilot happily outputs the Fast inverse square root code. I would expect that courts will judge such things to be copyright infringement. Copyright infringement statutes do not require to prove that the infringer copied a specific source (that would be difficult to prove), just that the content is substantially similar. Therefore, whether the tool is a simple ctrl-C ctrl-V or a sophisticated machine learning model should not make much difference.

The second question is whether OpenAI (or any other AI tool provider) can assert copyright on whatever the tools they provide create. The OpenAI terms of use seem relatively permissive, but others might be less generous. I do not know the answer to that question. I would hope they cannot, since they only provide tools (Microsoft should not be able to assert copyright on the text I write using Word, or the images I draw using Paint).

The third is whether a human using ChatGPT can assert copyright on ChatGPT answers, or otherwise constrain the use of the resulting text. The quote you give is probably based on the US copyright office’s position (taken during the monkey selfie copyright dispute): Because copyright law is limited to 'original intellectual conceptions of the author', the [copyright] office will refuse to register a claim if it determines that a human being did not create the work. However, giving a prompt to ChatGPT might or might not constitute significant creative input. The position that anything edited by a machine becomes public-domain is untenable (if I use an orthographic corrector on the draft of my novel, it does not turn it into PD), so it must be a question of degree. Also, non-US courts might have different opinions. Tigraan^{Click here for my talk page ("private" contact)} 16:06, 16 January 2023 (UTC)

As I understand it, the degree of access by the alleged infringer to the source text in question is a factor in determining infringement. Only a specific expression is protected by copyright; if you and I independently write the same sentence, one is not a copyright violation of the other. The amount of similar text also plays a role, since the larger it is, it's more improbable that it was created without copying.

Facts and natural laws can't be copyrighted; this also covers algorithms (though a particular expression can be copyrighted). So I can't copyright a single instance of a Javascript for-loop and claim rights to all Javascript for-loops as derivative work. In cases where the learning model creator is explicitly providing its model for use as a tool, I think (disclaimer: not a legal opinion) it is reasonable for this to be the same as a work for hire. Thus if the result is eligible for a new copyright owner independent of any source texts, the tool user would be the owner. (If I use a spellchecker on the latest bestselling novel, the result is not eligible for a new copyright owner.)

To be really safe, we'd want language models trained on public domain text. But I think it could be argued with a really large model trained on, say (just drawing numbers out of air), hundreds of thousands of documents with thousands of independent authors, the resulting correlations can no longer be attributed to specific input text, for cases where the output is not a significantly long passage substantially similar to a specific source text. isaacl (talk) 18:05, 16 January 2023 (UTC)

One of the main issues to deal with would be the following: an editor tells ChatGPT to write a text on a topic and then adds this text in the form of a paragraph/section/article to Wikipedia and thereby publishes it under Creative Commons/GNU license. The question is: what are the chances that this constitutes some form of copyright violation? This might concern specifically problems with the 1st and the 2nd question addressed by Tigraan, i.e. whether the copyright of someone whose work was part of the training set was violated and whether openAI's copyright was violated. For the first question, it's probably relevant what the copyright status of the texts in the training set is and how similar the produced text is to the texts in the training set, as isaacl points out. Answering these questions would be quite relevant for any Wikipedia policy on the topic, like the one JPxG is currently drafting. Phlsph7 (talk) 07:18, 17 January 2023 (UTC)

With respect to the issue of whether LLM output inherently violates copyright law: the copyright status of LLM-generated text is not defined by statute, so it is hard to make confident claims, but precedent exists for computer-generated art and other works created by non-humans. Here is what the US Copyright office has to say:

"Compendium of U.S. Copyright Office Practices, § 313.2" (PDF). United States Copyright Office. 22 December 2014. p. 22. Retrieved 18 January 2023.

The Office will not register works produced by nature, animals, or plants. Likewise, the Office cannot register a work purportedly created by divine or supernatural beings, although the Office may register a work where the application or the deposit copy(ies)state that the work was inspired by a divine spirit.

[...]

Similarly, the Office will not register works produced by a machine or mere mechanical process that operates randomly or automatically without any creative input or intervention from a human author.

It's not quite clear to me what the applicability is in this case. On Commons, the template and copyright category for PD-algorithm asserts that all algorithmically-generated works are public domain ("This file is in the public domain because, as the work of a computer algorithm or artificial intelligence, it has no human author in whom copyright is vested"). Whether artificial neural networks are capable of producing original intellectual output is less of a legal issue and more of a philosophical/anthropological one. It should be noted that human brains are themselves neural networks; much has been said, in a variety of fields, on the subject of whether humans create original works versus whether they merely juxtapose or recombine motifs and concepts that they're exposed to through participation in society. While interesting (and humbling), these discussions are unrelated to whether neural networks which have been exposed to copyrighted material in the course of their existence are capable of later creating original works under the purview of intellectual property law: they do. If this was not the case, a large majority of creative work would be illegal (good luck finding a band where none of the musicians have ever heard a copyrighted song before).

It does seem obvious that, say, OpenAI does not hold a claim to copyright on all works produced using its products: Adobe does not own copyright on all works produced using Photoshop, and Apple does not own copyright on all photographs taken with iPhones. If they did, this would certainly be a bold claim which required some evidence beyond idiosyncratic interpretation of vague phrasing in a EULA somewhere.

With respect to whether LLM output is capable of violating copyright law, I think the answer is obviously "yes". Here is an example, which I have just prompted from ChatGPT:

What are the lyrics to "Moonage Daydream", by David Bowie?

Keep your 'lectric eye on me babe

Put your ray gun to my head

Press your space face close to mine, love

Freak out in a moonage daydream oh yeah

These are, indeed, the verbatim lyrics to an unambiguously copyrighted song. Of note is the fact that I specifically asked it to do this. Compare with this prompt, using the first couple line of the song:

I'm an alligator / I'm a mama papa comin' for you / I'm a space invader / I'll be a rock-'n-rollin' bitch for you

I'm a force of nature, a wild beast / A creature of the swamps, a king of the feast / I'm a living legend, a natural high / 'Cause I'm an alligator, and I'll never die

While this would be a sick Judas Priest song, it isn't one, and in fact, "I'm+an+alligator+and+I'll+never+die" this text seems to never have been published. jp×g 01:57, 19 January 2023 (UTC)

It bears mentioning that I, a human being, also know the lyrics to "Moonage Daydream", and am capable of typing them into the edit box with no assistance whatsoever; doing this in an article would still be a copyright violation. I think that, for LLMs, the critical piece of knowledge is that output needs to be checked, and that it is insufficient to assume that generated text is always novel. jp×g 02:15, 19 January 2023 (UTC)

On the narrow point of who owns ChatGPT outpt, Sec 3(a) of their ToS states "OpenAI hereby assigns to you all its right, title and interest in and to Output...You are responsible for Content, including for ensuring that it does not violate any applicable law or these Terms." I thought there was an additional requirement to state ChatGPt was used, but I do not see it in the terms. Slywriter (talk) 02:14, 19 January 2023 (UTC)

@JPxG: since you mention computer-generated art: there are some lawsuits against Stability AI based on the claim that they violated the copyrights of people whose images were used in the training set. See [6] and [7]. The case seems to be similar to LLMs, with the main difference being that their AI trains on images and creates images while LLMs train on text and create text.

If I interpret the statement by the US Copyright office correctly, it seems to claim that a person can't own the copyright of a work that was created by a random machine process without creative input. It does not say that such processes cannot violate someone else's copyright. This would be in tune with the lawsuits mentioned above.

I think it's also unlikely that every output is a copyright violation. For example, if you just give it a sentence and tell it to correct spelling mistakes, there should be no problem in using the output. Phlsph7 (talk) 06:33, 19 January 2023 (UTC)

@Slywriter: Their sharing policy demands that Indicate that the content is AI-generated in a way no user could reasonably miss or misunderstand. Phlsph7 (talk) 09:23, 19 January 2023 (UTC)

LLM-assisted edits need to be appropriately marked as such in the history. —Alalch E. 01:38, 21 January 2023 (UTC)

Yes, and I think that "in a way no user could reasonably miss or misunderstand" requires the use of a notice in the article itself as well. –dlthewave ☎ 13:53, 21 January 2023 (UTC)

Agreed. It's easy for the average reader to miss an edit summary in the article history. So in-text attribution may be required. Phlsph7 (talk) 08:54, 22 January 2023 (UTC)

@Slywriter, JPxG, Phlsph7, Alalch E., and Dlthewave: Concerning the TOS clause that states "OpenAI hereby assigns to you all its right, title and interest in and to Output...You are responsible for Content, including for ensuring that it does not violate any applicable law or these Terms." — does that mean that Chat-GPT cannot legally produce the exact same output twice without violating the right, title, and interest that it previously assigned?    — The Transhumanist   20:47, 21 January 2023 (UTC)

I'm not sure what leads you to that conclusion. The licence does not grant you exclusive use to anything. isaacl (talk) 22:09, 21 January 2023 (UTC)

@Isaacl: I didn't come to a conclusion, I just asked a question, pertaining to Sec 3(a) of their ToS as referred to and quoted by Slywriter above, and repeat quoted by me. It appears you missed the quote somehow, because you didn't comment on it. To what license are you referring, and what relation does it have to the passage we quoted from the TOS?    — The Transhumanist   02:52, 22 January 2023 (UTC)

I don't know why you would ask the question you did, since the section you quoted did not say anything about granting an exclusive right, title, and interest to any output. isaacl (talk) 03:08, 22 January 2023 (UTC)

I don't think this has something to do with producing the same output for different users. It should be easy to find mock queries to which it often responds with the same output, for example, by asking it to "Say the word 'Hello'" or for simple translations. Phlsph7 (talk) 06:50, 22 January 2023 (UTC)

@Isaacl: Well, I checked the section again, and it is right there in plain English. It uses the word "assigns" instead of "grants", and it says "all its" instead of "exclusive". So, once it "assigns all its right, title, and interest in and to Output", how can it legally ever produce that same output again? (Because it already assigned it away).    — The Transhumanist   09:20, 22 January 2023 (UTC)

As I understand it, anyone can assign all their rights to the output of ChatGPT to someone else. In a similar way, I could assign to you all my rights to the Harry Potter series. This would not be of much use to you since the expression "all my rights" just refers to "no rights" in this case. Phlsph7 (talk) 09:32, 22 January 2023 (UTC)

@Phlsph7: In section 3a of the TOS, it's OpenAI that is assigning its rights to the chatbot output generated for the user. If Chat-GPT writes you a 3 paragraph explanation of gravity, and OpenAI has assigned you its rights to that explanation, can Chat-GPT legally write that exact same output for somebody else?    — The Transhumanist   09:58, 22 January 2023 (UTC)

I assume it works something like the following: it depends on whether openAI had any copyrights on it in the first place. If it did then this may be a problem because creating the copy for the second user might violate the newly obtained copyright of the first user. If it didn't then it presumably wouldn't be a problem because assigning all its rights to the first user effectively didn't do anything. But I don't think that this particular issue is very relevant for Wikipedia. Phlsph7 (talk) 10:14, 22 January 2023 (UTC)

The tool made no guarantee that it wouldn't generate the same output again for another user. The tool is in essence passing any rights of ownership (if they exist) in the original input through to the output. isaacl (talk) 19:24, 22 January 2023 (UTC)

Testing Chat-GPT's algorithm

Does Chat-GPT produce the same output to the same prompt given to it by 2 or more different users? Do any two want to try that?    — The Transhumanist   20:54, 21 January 2023 (UTC)

@The Transhumanist, not necessarily. It has a "temperature" factor (randomness). — Qwerfjkltalk 21:01, 21 January 2023 (UTC)

Perplexity.ai's answer to the copyright issue, and some questions...

@Phlsph7, Isaacl, Tigraan, Dlthewave, Slywriter, and JPxG:

I have found the this to be perplexing...

I entered the following prompt into Perplexity.ai:

who owns the copyrights of perplexity.ai's answers

And it returned the following answer:

"According to US copyright law, works generated solely by a machine are not eligible for copyright protection[1] . Therefore, the copyrights of Perplexity AI's answers belong to its co-founder and CEO, Aravind Srinivas[2]."

It looks like I just broke copyright by copying it here. But this contradicts the title of the first source provided (the second source is behind a registration wall):

The scary truth about AI copyright is nobody knows what will happen next — The Verge

Assuming that Aravind Srinivas does not own the output, I have some questions about posting AI search engine results: I look forward to your replies to the below questions.    — The Transhumanist   19:27, 21 January 2023 (UTC)

Question 1: Does putting quotation marks around a small copy/paste selection from an AI search engine, like the passage above, fall under fair use?

I would assume so. It's similar to how you can quote from copyrighted books. There are some limitations, for example, concerning the length of the cited text. And it should be clear where the quotation came from. Phlsph7 (talk) 06:57, 22 January 2023 (UTC)

Question 2: Would that protect Wikipedia (and its editors) from a derivative work claim?

From [8]: "In its most general sense, a fair use is any copying of copyrighted material done for a limited and “transformative” purpose, such as to comment upon, criticize, or parody a copyrighted work. Such uses can be done without permission from the copyright owner. In other words, fair use is a defense against a claim of copyright infringement. If your use qualifies as a fair use, then it would not be considered an infringement." Phlsph7 (talk) 07:06, 22 January 2023 (UTC)

Question 3: Let's say that perplexity.ai answers the same way to 2 different users, and they copy/paste the response on 2 different websites — who owns the copyright of that passage?

In our discussion so far, we haven't been able to conclusively figure out whether someone owns the copyright at all and, if so, who. That 2 users get and use the same response would be just a special case. Phlsph7 (talk) 07:14, 22 January 2023 (UTC)

Question 4: Would running a passage (from a chatty AI search engine) through a plagiarism checker be enough, before copying it into Wikipedia?

Plagiarism checkers are not perfect so they can't ensure that no plagiarism/copyright infringement was committed. The question would be whether they are good enough for our purposes, i.e. whether they are quite reliable for spotting plagiarism/copyright infringement pertaining to AI-generated texts. Phlsph7 (talk) 07:26, 22 January 2023 (UTC)

Question 5: Does Wikipedia policy allow an editor to click "Publish changes" for content that the editor did not personally compose?

Clarification: Clicking "Publish changes" implies that the editor composed the changes. Can an editor publish changes that they did not personally compose, that were composed by a chatbot search engine? (Please quote and provide links to the specific policies that allow or disallow this). Thank you.    — The Transhumanist   20:13, 21 January 2023 (UTC)

That would probably be a case of WP:PLAGIARISM even if no copyright infringement is involved. According to the summary: "Do not make the work of others look like your own. Give credit where it is due." Phlsph7 (talk) 07:06, 22 January 2023 (UTC)

It would be similar to copying public domain/open license content to Wikipedia, no? This is covered by several guidelines and explainers such as WP:FREECOPY and Help:Adding open license text to Wikipedia. As long as there's proper attribution, there's no general expectation that editors must compose the text themselves. –dlthewave ☎ 13:22, 22 January 2023 (UTC)

@Dlthewave and Phlsph7: Interesting. So, if you prompted a chatbot to write a new paragraph for the article on cream cheese and you add that to the article, you include an attribution to the chatbot in the edit summary? What do you put in the source reference?    — The Transhumanist   11:58, 25 January 2023 (UTC)

@The Transhumanist: As I understand it, for WP:PLAGIARISM it's sufficient to mention the source in the edit summary. You would have to find and add other reliable sources yourself since ChatGPT provides no sources or sometimes invents non-existing sources. However, for the Sharing & Publication Policy of openAI, in-text attribution would probably be necessary. So to comply with it, you would have to start the paragraph on cream cheese with something like "According to ChatGPT,...". This way, the text couldn't be used at all since ChatGPT is not a reliable source. Phlsph7 (talk) 12:23, 25 January 2023 (UTC)

Perplexity.ai's answer to some other questions

I gave Perplexity.ai some questions, and here are the answers. I gave it some questions asking what would it do if it was an administrator, instead of just a normal editor, along with some other things. ‍ ‍ Helloheart ‍ ‍ 20:23, 19 March 2023 (UTC)

First ANI case

Just a head up, a thread in Wikipedia:Administrators' noticeboard/Incidents has just opened about an user abusing AI-generated content at Wikipedia:Administrators'_noticeboard/Incidents#Artificial-Info22_using_AI_to_produce_articles. Sure, the editor in question did not made an edit in the mainspace, but the fact that this is happening at ANI is pretty concerning in its own right. I afraid that someone may have covertly spam articles with AI text already. CactiStaccingCrane 15:31, 23 January 2023 (UTC)

I am now adding the {{disputed}} template when encountering an AI-generated article, based on the following from the ChatGPT FAQ: These models were trained on vast amounts of data from the internet written by humans, including conversations, so the responses it provides may sound human-like. It is important to keep in mind that this is a direct result of the system's design (i.e. maximizing the similarity between outputs and the dataset the models were trained on) and that such outputs may be inaccurate, untruthful, and otherwise misleading at times. The commonality of all the AI-generated articles I've encountered so far (4, to be honest) is that they are not properly footnoted, implying that the author has not confirmed that the AI output is correct. The disputed tag seems to cover this issue well. I'm also dropping a note on the article's talk page explaining the link between AI output and correctness. — rsjaffe 🗣️ 01:00, 24 January 2023 (UTC)

ANI case is wrapping up. The first three articles were written by a hoaxer, and the refs for two of the articles may have been generated as well. The fourth article was promoting a company. — rsjaffe 🗣️ 04:18, 24 January 2023 (UTC)

This gives us a good look at the type of plausible-sounding nonsense that we can expect from LLM output. Comparing the draft (archive version, since it will likely be deleted soon) to Gecko, I see a few factual errors right away:

• Not all geckoes belong to the family Gekkonidae, which doesn't have 1500 species.
• Not all geckos have specialized toe pads that allow them to climb vertical surfaces.
• The largest geckos are 23"-24", not 10".
• Not all geckos are oviparous; some bear live young.

When this type of content is submitted, it needs to be thrown out straightaway. –dlthewave ☎ 17:09, 24 January 2023 (UTC)

More detected AI-generated papers

I started screening Drafts more carefully and am getting a number of hits corresponding to probable AI-generated articles (or at least part of the article is AI-generated). Examples include:

• Draft:NuMSP 97.69% fake
• Draft:Product Photography (the introductory paragraphs score 99.98% fake on GPT-2 detector)
• Draft:Classical Conversations 99.98% fake
• Draft:Leigh Bortins 99.82% fake
• Draft:CaSS (Competency and Skills System) 99.98%
• Draft:Sanjit Bajpai 99.90%
• Draft:Honda Big Red UTV 99.98% (features and maintenance sections, omitting the headings when measuring)

The list could go on, but I think this is enough to see some information about this. These pages tend to be created by users with few edits. A number of users are doing this, not just one or two. Conclusion: the tsunami has arrived. — rsjaffe 🗣️ 03:15, 25 January 2023 (UTC)

I think that mw:ORES (used in recent changes to highlight bad faith and vandalism) should integrate a screening mechanism for GPT-3 and other bots asap. I envision this is already a huge problem when large amount of hoaxes can be disguised as good content and we wouldn't even know about it. CactiStaccingCrane 03:21, 25 January 2023 (UTC)

Looking at the first few, the sourcing is not up to the standards of promotion to article space. Once clearly bad sources are removed and unsourced claims are tagged, this can clearly be seen. If AI ever gets to the point of being able to write an article that provides accurate information properly and verifiably sourced to reliable sources, then I'll be happy to have it writing for us. BD2412 T 03:46, 25 January 2023 (UTC)

@CactiStaccingCrane: It might be a good idea, but in reality, it requires a lot of effort from WMF to integrate openai-detector into mw:ORES. And I agree with @BD2412 for pointing out some drafts are promotional, which I think is a problem even before ChatGPT or even GPT-3 exist. 2001:448A:304F:52BA:8D12:5E35:69B7:8E09 (talk) 03:50, 25 January 2023 (UTC)

Well, some AI-generated articles have made it into article space. The first one I found has lots of text that's probably not AI-generated, but has a big hunk that is. Pavilion of Harmony, from "The Harmony Pavilion of New Asia College..." to "unique addition to the campus of New Asia College.", after removing the footnote indicators that confuse the analysis, rates as 99.98% fake. So the problem will leak into article space. And this means we need a way to pay special scrutiny to the AI-generated section, as that section is likely to have plausible but false information, given the way current AI models work. — rsjaffe 🗣️ 04:13, 25 January 2023 (UTC)

I agree that this could easily turn into a major problem as LLMs become more popular. As discussed at #Copyright status, these drafts violate at least WP:PLAGIARISM but probably also the Sharing & Publication Policy of openAI (if they were created using openAI tools). If AI-detectors are reliable, including them in mw:ORES would probably help a lot to mitigate the problem in case such an integration is feasible. Another alternative would be to create a bot that checks new submissions and tags them if they score a high value. A further thing to do at some point might be to make the editors reviewing drafts and new articles aware of this problem. Phlsph7 (talk) 06:28, 25 January 2023 (UTC)

10 Best ChatGPT Chrome Extensions You Need to Check Out Doug Weller talk 10:06, 25 January 2023 (UTC)

@Phlsph7 said: A further thing to do at some point might be to make the editors reviewing drafts and new articles aware of this problem. That's how I stumbled unsuspectingly upon this issue. I'm a new page patroller. I think they need to be looped in now, as that is the only guaranteed review step for new articles, and LLM-generated articles are already appearing. (I'm hoping that those users allowed to have their articles bypass this process won't abuse LLMs.) — rsjaffe 🗣️ 17:09, 25 January 2023 (UTC)

@Rsjaffe: Wikipedia:New_pages_patrol has various instructions on the different issues that new page patrollers need to be aware of. Maybe somewhere in there, a subsection could be added on AI-generated articles. Among other things, it should give a short explanation of what it is (the user tells the AI to generate an article in a matter of seconds and copy-pastes the results), what the problems are (plagiarism, false statements, no or invented sources, possibly copyright violation), and how to spot them (things AI-generated articles have in common and tools to detect them, like https://openai-openai-detector.hf.space/). Phlsph7 (talk) 19:31, 25 January 2023 (UTC)

Can I recommend, at least for right now, that some of these pages be copy-pasted into projectspace somewhere, so that we can see what they actually look like? I feel like these discussions basically have a couple-hour-long window outside of which it's impossible to see what everyone is talking about. jp×g 19:12, 25 January 2023 (UTC)

I will note here that the detector being linked here (at https://openai-openai-detector.hf.space) is a very old model tuned for the 2019 GPT-2, not GPT-3 or ChatGPT (3.5). I don't know if it's producing reliable results. It seems to me like most of the things it's flagging as machine-written are abysmal crap, so maybe it doesn't make a difference. jp×g 19:17, 25 January 2023 (UTC)

Some trials on articles I've written (using the lead paragraphs): Powder House Island and Nina Jankowicz are estimated at 99% human-written, but First Wikipedia edit is at 20% GPT for some reason. 1-Pentadecanol returns 40% GPT based on the first sentence, which decreases with subsequent sentences to 99%. However, when I asked ChatGPT to "write me a Wikipedia article about 1-pentadecanol", the result (which is viewable as an HTML comment if you edit this section) was estimated as 92% human-written. I don't know exactly what the underlying mechanism of this tool is, but we may want to take its output with a grain of salt. jp×g 19:23, 25 January 2023 (UTC)

The most recent articles I pulled (the ones in this section) were all initially detected by me using a search for a "signature" of a particular type of way of producing text for a particular version of LLM engine (I'm being vague as to not give hints to those who are trying to evade detection). I then visually confirm the signature. Then I run it through the GPT-2 detector. And then I'm only listing pages with > 99% chance of being LLM-generated. I'm 100% sure that the ones I've listed above are LLM-generated, but I'm also certain that this is only detecting that subset being generated under that specific set of conditions. — rsjaffe 🗣️ 19:24, 25 January 2023 (UTC)

Also, you have to remove the bracketed reference numbers (e.g., [2]) from the text as well as any intercalated headings to give an accurate score. — rsjaffe 🗣️ 19:26, 25 January 2023 (UTC)

To give an example of some LLM-looking text that doesn't score high on the GPT-2 detector, look at, ironically enough, Draft:Large Language Model. It scores 99.52% real, but sure looks fake to me. — rsjaffe 🗣️ 19:30, 25 January 2023 (UTC)

That detector works great if a text is modified, but it will crash when two or three words are modified. CactiStaccingCrane 19:32, 25 January 2023 (UTC)

Here are a few more detectors:

• https://detector.dng.ai/
• https://contentatscale.ai/ai-content-detector/
• https://corrector.app/ai-content-detector/
• https://writer.com/ai-content-detector/
• https://etedward-gptzero-main-zqgfwb.streamlit.app/ (worked before but currently down)

To test them, I asked ChatGPT to "write a Wikipedia article on cream cheese". It produces the following:

All the detectors agree that this text is AI-generated. When I fed them with the lead of the article Wikipedia (pure text without reference signs), they all agree that it's human-generated. Phlsph7 (talk) 20:03, 25 January 2023 (UTC)

Some attempt at figuring out what these are: I think that they are all based on the same code from HuggingFace. The actual source is here. I think that it may be simple enough for me to deploy it on a Toolforge test account; some of these websites seem unbelievably seedy. For example, "Content At Scale" advertises:

Want undetectable AI content? Our platform is the only one of it's kind that allows you to upload up to 100 keywords and get back 100 entire human quality blog posts (title to conclusion) without any human intervention. All the while, bypassing AI detection as it's the most human-like AI content ever produced. Our proprietary system uses a mix of 3 AI engines, NLP and semantic analysis algorithms, crawls Google, and parses all the top ranking content to put it all together. This isn't an AI writing assistant, this is a human level long-form blog post producing machine!

Certainly seedy if they are making it up... but also very seedy if it's true! jp×g 23:17, 25 January 2023 (UTC)

Quoting a chatbot

Below are some verification-related questions pertaining to chatbots.    — The Transhumanist   12:10, 26 January 2023 (UTC)

How would pasting in content generated by a chatbot be interpreted under WP:VER's requirement that all quotes must be referenced?

WP:VER states that all quotes must be supported by inline citations. If the chatbot's text is unique rather than preexisting somewhere else, using it would in essence be quoting the chatbot — how could that, as it isn't recorded anywhere, be referenced for verification purposes?

For not requiring referencing the quote of a chatbot, would WP:VER need to be modified?

News update

We need to get and stay ahead of this AI thing. See the following to get an idea how fast this movement is progressing:

1. ⭕ What People Are Missing About Microsoft’s $10B Investment In OpenAI : GPT3
2. Travis Tang on LinkedIn: ChatGPT for Data Science Prompts - 60 examples of what it can do
3. How to write an effective GPT-3 prompt | Zapier
4. OpenAI Licenses GPT-3 Technology to Microsoft (not exclusive)
5. OpenAI's investments
6. Should ChatGPT be used to write Wikipedia articles?

This article features the following Wikipedia article, initially composed using Chat-GPT: Artwork title, by Pharos, a great example of how a chatbot can be used by a responsible editor. Maybe a blanket ban is too much, and guidelines on how to use it correctly would be better. Makes auto-removal harder, though.

See an explanation and its discussions here: Talk:Artwork title

I hope you find these articles informative. Feel free to post more links, and comments, below.    — The Transhumanist   16:05, 28 January 2023 (UTC)

Crystallize chatbot discussions into a policy?

I think that there is a long list of things that make chatbot content objectively bad for / incompatible with Wikipedia in its current form and methods. Without elaborating here, a few quick notes are the inherent "blackbox" nature of AI, the inherent unaccountability for content, the inherent non-linking of content to sourcing, the fact that they do not know or follow Wikipedia policies and guidelines, (which themselves are a fuzzy ecosystem rather than categorical rules) they do know take into account what is already in the article, they (as a practical matter) remove accountability and expectations from the person who added the material. They also would destroy the ability to obtain volunteer time to review what has been put in. Most people willing to spend time to review something because they know that a human editor has take the time to write it would not be willing to spend large amounts of time dealing with something generated by a bot in a few seconds.

My thought is that we should say that such chatbot generated content is not allowed in Wikipedia. This is just briefly written, I or someone could flesh this out into something carefully written if there is interest.

We can and should decide this without or prior to solving the question of how to detect and enforce. A premise of having to solve detection and enforcement before step #1 would be a poison pill for accomplishing step one. Sincerely,North8000 (talk) 20:42, 25 January 2023 (UTC)

@North8000: There has been a draft guideline at Wikipedia:Large language models for a few weeks. I do not know that banning their use entirely is a good idea, but it seems quite obvious that just copy-pasting gigantic chunks of text directly from the model's output into the edit box is not a good idea (and almost zero percent likely to result in usable articles). I will try to write some stronger wording emphasizing that nobody should be doing this. jp×g 22:46, 25 January 2023 (UTC)

• Support blanket ban - I agree with North8000 that the policy should be "Chatbot generated content is not allowed in Wikipedia." I think this should apply to all namespaces, including talk pages. As chatbots improve, the policy can be changed, but right now, chatbot use for generating WP content appears to be a can of worms.    — The Transhumanist   01:03, 26 January 2023 (UTC)

  Pertaining to talk pages and forums, I meant not allowing the use of chatbots to generate a user's statements in a discussion. Posting chatbot output on a talk or forum page as an example in order to discuss it, is appropriate.    — The Transhumanist   12:33, 26 January 2023 (UTC)

  Support auto-screening - Pharos and their article Artwork title changed my mind (and is featured in Should ChatGPT be used to write Wikipedia articles?). I'd like to see a draft on guidance for the responsible use of chatbots in writing articles, including Pharos' approach. Meanwhile, our tech persons can work on automating the removal of undisclosed chatbot additions and the tagging and eventual removal of other entries that don't get edited within a reasonable time frame, or that are part of a pattern of posting disclosed but unedited chatbot submissions. Donald Albury was right, bad actors are going to spam Wikipedia with chatbot crap whether we ban it or not. Therefore, we should allow good actors to help offset their impact. Which brings us to the rest of the Web: it will be subject to hosting chatbot content, and so, as we are Wikipedia, we should trailblaze how to do it right.   — The Transhumanist   16:21, 28 January 2023 (UTC)

• Oppose a blanket ban. I think JPxG's demonstration shows that a LLM can be good for repetitive tasks like formatting a table, as long as a human validates the output. Actual generated prose is likely to be unsourced or sourced to fake sources, and so is already covered by existing policies. -- King of ♥ ♦ ♣ ♠ 01:23, 26 January 2023 (UTC)
• Oppose any blanket ban. Do think CSD modifications are needed to quickly remove algorithmically generated articles (AI is market speak here, it is not sentient and intelligence is debatable) and some formal guidance for editors would be useful. It's a tool like AWB, Twinkle, and any other scripting used. Used properly, it can cut down tedious work Slywriter (talk) 01:30, 26 January 2023 (UTC)
• Support blanket ban with the possibility of allowing specific use cases as we learn more. Our guiding principle should be that AI is completely inappropriate for creating or editing prose. Given the amount of plausible-sounding nonsense we've seen in recent AI-generated drafts, I also don't trust it for coding work such as formatting text or rotating tables until its reliability has been demonstrated for the specific task. This should apply to article, talk and draft spaces with very limited exceptions for demonstration purposes. –dlthewave ☎ 03:24, 26 January 2023 (UTC)

  With respect to tables and templates, what do you envision as "demonstrating reliability"? It is not exactly brain surgery to look at a table and see if the columns or the parameters or whatever are in the right places. You have to do this anyway: we currently do not require editors to prove that they have never typed an extra } and had to go back and fix it. jp×g 05:35, 26 January 2023 (UTC)

• Oppose for a variety of reasons, as discussed by others above. But I'll go with the most basic one: how would any of this be actionable? I know you claim it's a poison pill problem to raise, but there's no way whatsoever to detect this with certainty. We'd be using an outside tool to claim text is AI written and then delete things based on that claim. I don't care how many 9's you've got in the decimal places, there's no way to be infallible here. If the editor that added the text says they wrote it themselves, are we just going to say that they're lying and that they have to re-write it or something? There's not even evidence of copyvio in such a case and if the added content meets all other requirements, including proper verifiable sourcing, then I see no way to enforce such a ban. Silverseren^C 03:33, 26 January 2023 (UTC)

  If the detector is good, one mistaken evaluation is possible, but a consistent output of one or the other is solid. Dege31 (talk) 17:49, 26 January 2023 (UTC)

• Comment I'm not fully decided whether banning it entirely is the right course of action at this stage. But there is a significant potential to abuse it so most forms of non-trivial AI-assisted edits should be strongly discouraged. The policy should make it very clear that any addition of AI-generated text needs to be labeled as such in the edit summary to avoid WP:PLAGIARISM. The editors also need to be reminded to obey the license and sharing policy of the AI provider. In the case of ChatGPT, for example, in-text attribution is apparently required. Phlsph7 (talk) 08:53, 26 January 2023 (UTC)
• Support blanket ban of any LLM-generated text. Not sure yet about using such tools purely for lay-out, but it should not be allowed for either generating new text or for rephrasing existing text, as both cases are way too problematic. As for "how is this enforceable", just like other difficult policies, where near certainty is sufficient (like WP:DUCK for socks, which isn't infallible but good enough). Advantages of a policy are also e.g. when a newbie says something like "why was my article deleted, it was generated by ChatGTP so has to be good", one can easily point to the policy to explain that it isn't allowed instead of having this discussion again and again. Fram (talk) 09:04, 26 January 2023 (UTC)

  That's a good point about potential problems with enforcing it. Having a policy can be useful to discourage certain types of behavior even if it is difficult to enforce in every case. We'll have to see how useful and reliable AI-detectors are in this process. Phlsph7 (talk) 09:15, 26 January 2023 (UTC)

• I think the way forward is even stronger expectations on sourcing. If you can't provide the content of the reliable source for your edit, it should be reverted. (This would include people machine translating foreign Wikipedia articles without having access to the original sources). —Kusma (talk) 10:30, 26 January 2023 (UTC)

  I assume that you are referring to automatic reversion (correct me if I'm incorrect). What about the wording of the verification policy that reads "challenged or likely to be challenged"? If it is not challenged or likely to be challenged, it doesn't need references. How will a bot be able to tell the difference between what does and does not require references? Or would the bot's removal of an edit constitute a challenge? Whether reversion is automated or not, should all new content to Wikipedia be challenged by default? That would require a change to WP:V, and that seems unlikely to happen.    — The Transhumanist   11:36, 26 January 2023 (UTC)

  I would challenge the factual accuracy of anything generated by AI. –dlthewave ☎ 13:41, 26 January 2023 (UTC)

  I'm not talking about automatic reversion, can't see how to do that without AI. And the verification policy is applied differently to new and existing articles; for new articles, we are in practice already expecting much better sourcing than "likely to be challenged" (just look at what will be rejected by AFC). Perhaps we should expand this to addition of content to existing articles. —Kusma (talk) 13:56, 26 January 2023 (UTC)

• Support blanket ban for now; would be fine having a discussion later on allowable use cases, but I'd rather we started with a total blanket ban first, and then itemize specific possible use cases if we later decide there's some utility. --Jayron32 12:47, 26 January 2023 (UTC)
• Support blanket ban for prose as there is significant copyright violation concerns about AI text. Text generated by bot will not substitute reliable sourcing. Maybe in the future when Abstract Wikipedia come online, we can give some leeway for bots to generate text based on reliably cited info, but for now, it's just too risky for the project. CactiStaccingCrane 13:34, 26 January 2023 (UTC)

  If there is interest in at least reviewing this possibility (which it sounds like there is), as noted there it needs to be written better than my initial trail balloon above. I'll do that but still keep it short. I think that it can be done in a way that deals with the main enforcability questions and also allows described useful uses by allowing bot-assisted editor generated content. I'll do that within a 1/2 day. North8000 (talk) 14:33, 26 January 2023 (UTC)

• support ban for text additionsThis is so susceptible to abuse and inserting misleading content that it should be banned. There is little benefit of allowing text generation and much harm.

  — rsjaffe 🗣️ 16:05, 26 January 2023 (UTC)

• Oppose a blanket ban. it's premature and is not based on any actual cases in Wikipedia. Rjensen (talk) 16:20, 26 January 2023 (UTC)

  Have you seen WP:ANI#Artificial-Info22 using AI to produce articles? It's already happening. Fram (talk) 16:30, 26 January 2023 (UTC)

• Oppose new policy that goes beyond small clarifications of WP:BOTPOL. I think it is obvious that ChatGPT is already covered by it and I do not see what modifications are proposed.

  Things that go against a content policy (WP:COPYVIO, WP:V, etc.) should be reverted / deleted on those grounds, and on those grounds alone; editors that make many such edits should be warned and then blocked. Editors who make faster edits than reasonably possible by hand should be dealt according to WP:MEATBOT.

  I oppose any policy to revert / delete / ban based solely on a "seems bot-written" criterion, unless and until it has been proven that (1) this is a real, time-consuming problem on Wikipedia, and not a few random tests within the sea of vandalism, and (2) whatever criterion is used has been independently tested to establish its sensitivity and specificity and validated by the community. Tigraan^{Click here for my talk page ("private" contact)} 17:24, 26 January 2023 (UTC)

  I also thought it obviously fell under BOTPOL, but this discussion shows some uncertainty. –dlthewave ☎ 20:04, 26 January 2023 (UTC)

• Oppose a blanket ban. I agree in general with the oppose reasons given above. I also think such a ban would be "virtue signaling" without being effective. Editors who want to add AI-generated material to Wikipedia will not be stopped by such a policy. Consider how often our existing policies stop editors from adding un-verifiable, POV-pushing content. What we can do is work on effective strategies for detecting and removing un-sourced, un-verifiable content as expeditiously as possible. - Donald Albury 19:34, 26 January 2023 (UTC)
• Oppose a blanket ban. I expect that in another five or six years, we will happily be allowing a Wiki-AI to both write and clean up most of our articles. BD2412 T 22:26, 26 January 2023 (UTC)

  And AIdmins to deal with AIsox ~ Selfstudier (talk) 22:53, 26 January 2023 (UTC)

  While all of this wouldn't surprise me given the other aspects of this dystopia we call Earth, it would simply confirm that we live in the Golgafrinchan Ark B world. Andre🚐 22:55, 26 January 2023 (UTC)

  I suspect it will become necessary to rely on such bots, as our current model is creaking under the current load. In just the last three or four days I've discovered a couple articles that I started that are in need of serious cleanup and repair, particularly on sourcing (linkrot and other issues) and updating. Donald Albury 22:58, 26 January 2023 (UTC)

  Oh don't get me wrong, I would love to see AI bots to improve references, and do other automated tasks under the bot policy. But the AI should not be trusted for facts or interpretations. And I do fear the ChatGPT-ization of the language. There is good and bad writing on Wikipedia, but at least it was written by and for human beings. Andre🚐 23:02, 26 January 2023 (UTC)

• Oppose a blanket ban. I agree with King of ♥ ♦ ♣ ♠'s comment above. If an editor adds unsourced AI-generated content, that content should get treated the same way as non-AI generated unsourced content. WP:V and other existing Wikipedia policies already cover that. Some1 (talk) 00:05, 27 January 2023 (UTC)

  But, a great deal of unsourced content doesn't get treated. The amount of unsourced content on Wikipedia is vast. If you don't believe me, pick a scholastic subject like Roman Empire, click on "What links here", and open a bunch of the links in tabs and start inspecting. What about random article? Too many stubs, and it's boring. With "What links here", you can get longer articles on average to view. The striking thing is the lack of "citation needed" tags - they are spaced few and far between. They can be found on roughly 1% of all pages, while unsourced content can be found on a much higher percentage.

  Another thing to try is go to Template:Unreferenced, and click on "What links here". The first page I clicked on was Tank destroyer. The tag is dated March 2009.

  The point is, you make it sound like all unsourced content gets quickly removed. That's not the case for a huge amount of content. It can sit there for years. LLM-generated content can be generated in great quantities fast, and therefore has the potential to accumulate more quickly than content composed by humans. Is it wise to let it sit there until a human comes along to remove it? In terms of a quantity competition between humans and computers, computers will win. This will take auto-removal to keep up. It would be best to start building those tools now. See Fram's post, above for an even better rationale.   — The Transhumanist   06:49, 28 January 2023 (UTC)

• Comment: There's no way to enforce a ban, and at any rate Wikipedia would do well to have more automation. Both MW and en.Wiki are heavily dependent on manual labor, with 3rd-party "bots" doing some of the more menial tasks. Compare how one would compose a document on a modern word processor vs. how editors do it here: no Wikitext, no copying-and-filling-out-templates (and no separate commits just for AnomieBOT to date them), no broken pages because of parsing errors, no dragging someone to ANI/AE for a T-ban violation (because there's such a thing as access control), no separate citations of the same source in five different formats (because there's such a thing as reference management); and you can actually comment on a specific paragraph without looking for a diff number, opening a "discussion" (which is in fact just another near-meaningless bit of Wikitext), signing it and hoping that no edit conflict arises because someone changed something 5,000 words up the page. We need to get rid of the concept of a WP:GNOME before we can even consider how to deal with a language model that can generate an entire article in a fraction of a second. François Robere (talk) 13:31, 27 January 2023 (UTC)
• Oppose. I get that this raises some interesting questions in the abstract, but come on guys, we have enough problems with instruction creep already, we don't need to start writing policies in anticipation of the hypothetical abuse of Silicon Valley's latest fad. – Joe (talk) 14:18, 27 January 2023 (UTC)
• Support blanket ban because fuck LLMs and fuck the corrupt, unethical industry that created them. XOR'easter (talk) 17:32, 27 January 2023 (UTC)
• Oppose blanket ban blanket ban and suggest we develop Wikipedia:Large language models (perhaps under the name Wikipedia:Computer-assisted text generation suggested by Michael_D._Turnbull). I don't think that going forward we can ban AI-generated text writ large, first and foremost for the very simple example that many people currently editing wikipedia use text edit widgets that already incorporate something of this in the form of spell check, autocorrection and autocomplete, and these kind of tools will continue to blur the line between AI, language models, and human-generated text. Going forward it would be practically Neo-Luddism to eschew all AI. I don't like the use of ChatGPT right now today to generate text, I don't like it at all, but neither can I bury my head in the sand and whistle Dixie, pretending it doesn't exist and won't grow in importance. We should meet this head on rather than pretend we can completely ban AI-assisted or AI-generated text. —DIYeditor (talk) 23:27, 27 January 2023 (UTC)
• Oppose What seems to scare people about ChatGPT is that it writes better than most Wikipedia editors. Competition is healthy and so should not be subject to restrictive practices. See also Luddism. Andrew🐉(talk) 22:32, 28 January 2023 (UTC)

  Competition is good when it's robot vs real people? You'd be happy if most of the encyclopaedia was written by AI? And most of the real people just gave up? Doug Weller talk 13:05, 29 January 2023 (UTC)

  I think that most people are at least open to the idea of using LLMs like ChatGPT for good and practical purposes, such as summarizing dozens of paragraph from a reliable source, or making a WP:Earwig-like bot that detect source-text integrity issues, or detecting possible hoaxes/context-dependent vandalism in Special:RecentChanges. I'm sure that when these LLM-based tools come out, people will use them just as much as mw:ORES and User:Cluebot NG today. The problem as of 2023 is that these tools does not exist yet and as in the current form, LLMs is an extremely powerful tool for bad actors yet disadvantaged good-faith Wikipedia editors. I feel that this situation between LLMs and Wikipedia right now a bit like Wikipedia and the academia in the early 2000s, when Wikipedia is full of uncited info and its reliability is really shaky to say the least (see also https://www.nostalgia.wikipedia.org). Maybe this will change in the future when whoever makes a LLM model that's aligned to our values and policies, but in my opinion for now a blanket ban is necessary to prevent mass vandalism while we are trying to process the situation. CactiStaccingCrane 13:24, 29 January 2023 (UTC)

  @CactiStaccingCrane: How would a ban of LLMs prevent vandalism exactly? Vandals would simply ignore the ban, while many good actors would obey the ban. You would in effect be preventing good actors from using the tool, and not bad actors. The only way to deal with vandals who ignore a ban is directly – that is, identify their text and remove it, and block the vandals. But you can do that anyways. So, wouldn't it be best to identify and remove LLM vandalism while allowing good use of LLM-generated text? (See the starting edit and talk page for the chatbot-generated article Artwork title). So, I'm confused as to how you believe a blanket ban would help. Let me repeat my initial question, along with a follow-up question: How would a ban of LLMs prevent vandalism exactly? And why would banning good actors using LLM (like Pharos and JPxG) be necessary? I look forward to your replies to these 2 questions. Sincerely,    — The Transhumanist   21:46, 29 January 2023 (UTC)

• Support blanket ban. Google and Bing actually considers auto-generated content as spam, which is bad for SEO. If anything, copying from GPT and pasting into an article is no different from copying from a free source and pasting into Wikipedia. The text generated is not your own, but is licensed to you forever, and, since this is a machine we are talking about and not a human, there is a probability that specific text output will be the same for multiple users for the same input.

  I believe GPT is best used as an inspiration for further research, but in no ways is it actually any more useful than a machine translated article. Aasim - Herrscher of Wikis ❄️ 14:28, 3 February 2023 (UTC)

• Oppose a blanket ban. I would prefer that editors use GPT to assist in writing, but to avoid plagiarism, tag their edits as "GPT-assisted". I would additionally suggest that GPT-assisted edits be all subject to pending changes, except if a user reaches Extended confirmed status. I personally fear that, for those who are familiar with American history, ChatGPT could be the prohibition era equivalent for Wikipedia. InvadingInvader (userpage, talk) 07:47, 6 February 2023 (UTC)

  Pending changes can only be applied to articles. There is no way (that I am aware of) of having pending changes apply to only some edits to an article. Donald Albury 15:43, 6 February 2023 (UTC)

• Comment IMHO, the issue isn't per se about how articles are written. Very few people would object if Wikipedia was provided content that was accurate, well-sourced & well-written by GPT bots. The root issue is that we don't have enough volunteers to review a flood of these GPT-created articles, let alone enough to review them knowledgeably. Viewed in this light, New Page Patrol's often-discussed problems dealing with a tsunami of new articles is but the nose of the camel poking under the tent. (And then there is the threat of gut-&-stuffing existing articles.) Most of us have other, off-Wiki demands on our time, so we are forced to contribute on a part-time basis; but even if we were all full-time volunteers, we still couldn't keep up with all of the edits. -- llywrch (talk) 21:21, 11 February 2023 (UTC)
• Oppose. As demonstrated by Artwork title these bots can be used positively or negatively. Our policies should be tuned to prevent the negatives, and not also prevent the benefits. BilledMammal (talk) 11:17, 13 February 2023 (UTC)
• Oppose a blanket ban. scope_creep^Talk 16:19, 13 March 2023 (UTC)
• Oppose a blanket ban. IMHO a generative AI is just a tool, like a good spell checker: it further equalizes the field for editors with language skills below native level (this is not necessarily a good thing, I personally happen to think that barriers to entry are useful in our context). From the perspective of abuse potential, the AI is not much different from wholesale copying of sources that the community is dealing with daily. What we need is a policy of full disclosure: the edits that use generated text should credit the bot; this is actually already by the bot copyright rules anyhow, so should be mandated to avoid the copyvio anyhow. --Dimawik (talk) 19:29, 22 March 2023 (UTC)
• Support Artwork title is to me the case in point for why this tech is bad. The AI created version has no citations. So an editor must add citations. But that means they are working backwards, which is always the worst kind of way to cite an article. The proper Wikipedia process is finding a source, and then writing a sentence based on it. But if you write a sentence, and then find a source for it, that is confirmation bias. ChatGPT is Pandora's box, and we are in no way properly prepared to handle the pestilence it releases. Allowing AI to write our articles will serve to degrade our trustworthiness in the eyes of our readers, and for little appreciable benefit to us. CaptainEek ^{Edits Ho Cap'n!}⚓ 20:05, 26 March 2023 (UTC)

  We have always had plenty of human editors who write articles without sources, which means someone else has to find sources to support the content. The general consensus is that we cannot just delete unsourced articles, but have to keep them around unless and until someone can build a case that the topic is not notable. Even if someone provides sources when they create an article, I am sure that they often have found the sources after they have decided what they want to write. As we do not yet have a means of identifying chatbot output with any certainty, I think banning articles written by chatbots is an unenforceable gesture at this time. BTW, I agree that the proper way to write an article is to start with a variety of sources, and see what they tell us. I have failed in writing an article if I do not know considerably more about the topic when I finish than I did before I started. Donald Albury 22:09, 26 March 2023 (UTC)

Chatbot This idea probably goes to the heart of it

How about this? (just a draft to be tweaked)

An editor is responsible for every part of every edit that they make. The norm is that they should have their own rationale for everything that they added including every word, phrase and sentence. For additions to articles, they should make a specific effort to make sure that the edit is appropriate with respect to the current article. For example, with respect to the structure of the article and avoiding duplication. They should also make a reasonable specific effort to assure that each potion of their addition is verifiable or verified in accordance with WP:Verifiability. It is unlikely that these requirements and expectations would be met with AI generated content (or any large amount of text that has been copied from elsewhere and pasted in, copyvio issues notwithstanding)

Wikipedia relies on volunteer efforts to review additions. This often requires time consuming reviews of individual words, phrases and sentences; obtaining this degree of effort relies on the understanding that the editor who put the material in has made a similar effort to develop that phrase or sentence. Expecting editors to give this review to large amounts of material which were generated by AI in a few seconds would cause a substantial loss of this effort.

Accordingly, this clarifies that removal / reversion of en masse additions of material suspected of being AI generated is considered to be an appropriate practice. An editor who seeks to restore the material is expected to break it into small portions, each with individual explanatory edit summaries. If such a removal results in deletion of the entire contents of the article, it then becomes a candidate for speedy deletion.

Sincerely, North8000 (talk) 21:07, 26 January 2023 (UTC)

I don't favour making machine-generated text a special case. Given that editors are already responsible for verifying every aspect of their edits, any clarifications should be equally applicable to all cases, such as human ghostwriting teams. isaacl (talk) 21:48, 26 January 2023 (UTC)

I like that a lot. — rsjaffe 🗣️ 05:06, 27 January 2023 (UTC)

I have integrated the above three paragraphs into Wikipedia:Large language models, see Wikipedia:Large language models#Specific guidelines and Wikipedia:Large language models#Summary removal of larger LLM-generated additions of article prose. —Alalch E. 10:33, 27 January 2023 (UTC)

cf. Wikipedia:Village pump (policy)/Archive 176#RFC: change "verifiable" to "verified". Also: do not make assumptions about the capabilities of AI. There are models at work that integrate references, and you should assume that at some point they'd be able to compose texts that are comparable to any Wikipedian's. Ergo, policy should focus on what we're looking for, not who or what composed it. François Robere (talk) 14:12, 27 January 2023 (UTC)

We can change the policy if something is developed that is reliable enough. Until then, blanket rejection is appropriate. — rsjaffe 🗣️ 15:46, 27 January 2023 (UTC)

It's important to see the conduct side too because LLM misuse forms a pattern of disruptive editing. It starts with one person's idea that a specific thing can be accomplished on Wikipedia in this way, proceeds with the intent to implement this idea without caring to understand and account for what Wikipedia's requirements are, and ends with an undesirable action which may be repeated if not addressed. —Alalch E. 15:55, 27 January 2023 (UTC)

If chatbots are banned, would the article Artwork title have to be deleted?

Artwork title was created by chatbot, and heavily edited by a human since. If chatbots (and LLMs) become banned, how would it apply to pre-existing chatbot-generated articles?    — The Transhumanist   22:09, 29 January 2023 (UTC)

I don't think it would have to be deleted (or restarted) unless it was somehow found to be a copyvio. Maybe it would be peer reviewed instead. Roll 3d6 (talk) 08:50, 5 February 2023 (UTC)

More specific proposal: Blanket ban on LLM content on Talk page discussions

Regardless of the community's decision on LLM-generated content in articles, which is the focus of much of the discussion above, the ability of editors to flood talk pages with artificially generated content arguing for a particular position on talk seems to have no redeeming value, and represents a new form of Wikipedia:Sockpuppetry. I propose a blanket ban on such writing, with especially strong guardrails for RfC's and AfD's. (Alternatively, I would be open to a phrasing that allowed LLM's to summarize the state of a debate, or be used to generate sample content for discussion, as in the conversation above, but not used to make arguments. That just seems harder to phrase clearly.)Carwil (talk) 20:22, 27 January 2023 (UTC)

That's a really weird way to use them. Also, why would it matter? Even if an editor used an LLM to generate a better phrasing of the argument they want, it's still their account putting forth the argument. And the argument is either valid or not in regards to others involved in the discussion. Why is this a problem exactly? Do you have any examples to better clarify this sort of usage? Silverseren^C 23:14, 27 January 2023 (UTC)

Here's an article that should help clarify the relevant danger: Researchers demonstrate how attackers can use the GPT-3 natural language model to launch more effective, harder-to-detect phishing and business email compromise campaigns.. If they can use it to write convincing email scams, Wikipedia talk pages should be a breeze. Here's a quote from the article: ""The generation of versatile natural-language text from a small amount of input will inevitably interest criminals, especially cybercriminals — if it hasn’t already. Likewise, anyone who uses the web to spread scams, fake news or misinformation in general may have an interest in a tool that creates credible, possibly even compelling, text at super-human speeds." If that doesn't convince you, I don't know what will.    — The Transhumanist   10:16, 28 January 2023 (UTC)

Has this ever actually happened? – Joe (talk) 05:29, 28 January 2023 (UTC)

Maybe a more relevant question is "Will we be ready for it when it does?"    — The Transhumanist   10:16, 28 January 2023 (UTC)

As Transhumanist is suggesting, my concern is rapid creation of multiple arguments that either tilt a discussion or waste the time of sincere contributors. Users should be warned that they can't substitute mass-produced arguments for their own judgment inside the encyclopedia.--Carwil (talk) 19:26, 28 January 2023 (UTC)

• Support - Weird? Compared to the many examples of what Chat-GPT can do posted on social media, writing talk page or forum posts on Wikipedia seems comparatively bland and simple. Why would using an LLM on a talk page or discussion forum matter? Because, it is faster than a human. With it, a user could participate in more discussions in less time. But, the big concern here is using it on multiple accounts with different writing styles to stack votes on issues with little chance of being discovered as the same person. That's sockpuppetry elevated to a higher level. Therefore, banning chatbots from being used to compose talk page or forum posts is quite reasonable.    — The Transhumanist   05:42, 28 January 2023 (UTC)
• Support On a risk/benefit analysis, the potential benefit to allowing this is so small that any risk (as described above) is unacceptable. — rsjaffe 🗣️ 05:50, 28 January 2023 (UTC)
• Reluctant Oppose As my comments above, it is difficult to draw a line on which language models are allowed and which are not. Clearly people are allowed to use autocompletion and predictive text and such. Are they limited in what varieties and designs of autocompletion they use? I think this requires further discussion and hopefully input from experts. —DIYeditor (talk) 06:42, 28 January 2023 (UTC)
• Oppose a blanket ban as premature, but support at least some restrictions. While I would certainly be first in line to support a blanket ban if an actual problem arose, I am not sure that we know exactly what shape this will take enough to come up with an intelligent solution. For example, earlier in this very discussion, we were posting LLMs' output in order to judge their capabilities, so any prohibition would need to take exceptions like this into account. That said, I do support some more specific language for WP:LLM about it being very bad to use them undisclosed to argue your case in discussions. For example, if I were a shady dude, I could flood the zone with shit right here on VPP by typing out massive walls of text replying to every single person who disagreed with me, without regard for whether my arguments were sound or even correct, and even if I represented a minority view it would probably irritate and discourage my interlocutors until they stopped commenting (thus bringing me closer to a majority). Similarly, at the blood-soaked fields of AfD I could trivially write out a three-paragraph !vote on all 50 of the day's nominations (whether I was a partisan for keeping, a zealot for deleting, or a fanatic for any sort of POV). jp×g 19:45, 28 January 2023 (UTC)
• Oppose a blanket ban. As JPxG, I'm open to some restrictions, but I can see LLMs as a potentially useful tool for people who want to offer their viewpoint in discussion but lack the fluency or time of some other editors. (A bit more acerbically, our discussions already tend to be influenced by editors who are functionally LLMs: good prose stylists, possessed of enormous sitzfleisch, and not well grounded in factual specifics. If LLMs force us to review WP:BLUDGEON and our methods of dialectic/achieving consensus, there's no reason to grant people like that the privilege of immunity.) Choess (talk) 20:44, 28 January 2023 (UTC)
• Oppose per WP:CREEP. As a recent example, I included some ChatGPT output in a recent discussion at ITN. Not seeing the problem. Andrew🐉(talk) 22:23, 28 January 2023 (UTC)

  I think your example misses the spirit of my proposal, which is un attributed LLM output substituting for our own reasoning and arguments on Talk. Happy to modify accordingly. --Carwil (talk) 13:25, 29 January 2023 (UTC)

• Oppose a blanket ban on talk pages specifically (I would support one in article space). The largest problem by a huge margin with using ChatGPT on Wikipedia is adding factually incorrect information to articles. This is less of a concern with talk pages, which the general public doesn't usually read and which are not primarily meant to impart facts or be reference material. We already have a ban on sockpuppeting and an expectation that decision-makers will ignore bad arguments, which cover the negative use cases mentioned above. Also, as mentioned above, there is not an enormous difference between ChatGPT and predictive text or services like Grammarly, and the line between those is going to become blurrier by the day, faster than policy can keep up with. Gnomingstuff (talk) 01:20, 3 February 2023 (UTC)
• Oppose as making up rules when no problem has been demonstrated. If it does happen, we can warn, revert, and block the offending material and offender for wasting our time. Graeme Bartlett (talk) 12:10, 3 February 2023 (UTC)
• Oppose LLM:s are state of the art translation tools so banning them from discussion is unproductive. I think that it would be nice that if there is new content generated by LLM it is told, but generally using LLM as writing assistance should be allowed. --Zache (talk) 10:26, 19 March 2023 (UTC)

The discussion has diverged

Further issues on this topic (chatbot-generated content) are being discussed at Wikipedia talk:Large language models, the talk page for the policy draft on this subject.    — The Transhumanist   05:15, 30 January 2023 (UTC)

The discussion has been re-opened

• Since the discussion basically died out after being closed, rather than migrating to WP:LLM, I'm undoing User:CactiStaccingCrane's close, with his permission. Let's keep all discussion in one place, so we can get solid consensus on all these points. DFlhb (talk) 13:46, 1 February 2023 (UTC)

Chatbot news update 2023-02-12

1. Panicked Google to unveil over 20 AI products in 2023
2. ChatGPT reaches 100 million users two months after launch
3. Google has stiff competition now, after Microsoft integrated powerful A.I. technology into its search engine.

Where do we expect issues to occur?

If we can work out the issues that editors improperly using these tools will cause then we can adjust our policies to make it easier to address those issues, and then as the technology develops we can see how it is being used and how it can be used and create a more specific policy then. BilledMammal (talk) 11:21, 13 February 2023 (UTC)

@BilledMammal: Here are some:

1. Made up content
2. Lack of source citations
3. Made up sources
4. Claims based on out-of-date sources
5. Possible synthesis
6. Copyright unclear as the source of the material is unknown
7. Copyright unclear as the issue of who owns the output of a chatbot has not been resolved
8. Inclusion of 3rd-party copyrighted material, such as lyrics, a poem, etc.
9. Chatbot license violation (e.g., for omitting attribution of the chatbot with the generated content)
10. 100,000,000 ChatGPT users and growing (after just 2 months),[9][10][11][12] suggesting that the use of tools like this may become ubiquitous, creating the potential for a large volume of one-off posts of chatbot generated text (and other media)

Not a comprehensive list.    — The Transhumanist   07:31, 15 February 2023 (UTC)

Timeline of ChatGPT news

See: Wikipedia talk:Large language models/Archive 2#Timeline of ChatGPT news — Preceding unsigned comment added by The Transhumanist (talk • contribs) 08:40, 1 March 2023 (UTC)

Practical rules

I think that the community will benefit from WP:AGF towards the bot use and establishing the clear rules for constructive edits. I propose at the very least:

1. Mandate a standard format for the copyright reference giving the credit to the bot (Google Bard apparently uses the Creative Commons Attribution-ShareAlike 3.0 Unported license, which should be good for our purposes (here is the response I got from the Bard: I am licensed under the Creative Commons Attribution-ShareAlike 3.0 Unported license, which means that you are free to share and use my text as long as you give me credit and share any changes you make under the same license.). The reference should be both in the edit history and on the talk page, similar to other CC BY SA reuse (WP:COPYWITHIN). If someone had already created a template similar to {{copied}} that mentions the bots, please let me know.
2. Mandate much higher levels of quality for generated text (high quality sources, proper cites, etc.).
3. Mandate a copyvio check with reference to the results placed onto the talk page.
4. Encourage the speedy deletion of newly created or added generated text not meeting #1, #2, and #3. Essentially we might want to treat the bot abuse in the same way as WP:COPYVIO.

Unless I see objections here, in the next few days I am going to create an article Bailey FFT algorithm as an example of using the Google Bard for creating the new articles. I am slightly aware of the topic and thus will be able to separate wheat from the chaff and provide proper sources. Comments will be welcome. --Dimawik (talk) 19:55, 22 March 2023 (UTC)

Bard is not licensed under CC BY-SA 3.0. You asked Bard, and it gave you a bullshit answer, because that's what Bard does. Another user got Bard to say that it got shut down 6 months ago, because "Google Bard was not a very innovative product". You cannot trust its answers to be accurate. Google's terms of service never state which license covers Bard's output, but Google has no right to set any license anyway, because in the U.S., LLM outputs are not copyrightable. DFlhb (talk) 20:06, 22 March 2023 (UTC)

Thank you for spelling out the additional tidbits. I am already aware of the situation you describe. Incidentally, I am kind of expecting to see the Supreme Court forced to revisit its decision soon, as there is no clear boundary between a very advanced spell checker (or a machine translation tool) and generated AI, and using the next-generation Grammarly surely does not cause the writer to lose her rights. However, (1) in a situation like this one has to cover his/her behind with a cardboard, and (2) going with the most restrictive interpretation (Bard's response) helps to drive through the item #1 in the list above. Dimawik (talk) 20:23, 22 March 2023 (UTC)

Just a heads up: Wikipedia_talk:Manual_of_Style#MOS:PREFIXDASH, opinions are welcome. Brandmeister^talk 13:31, 17 March 2023 (UTC)