Wikipedia

Wikipedia talk:WikiProject AI Cleanup

Main pageDiscussionNoticeboardGuideResourcesPoliciesResearch

No, seriously, what do you have to do to get people to believe you?

[edit]

The discussion was archived but I wasn't done, nor did I get an answer to my question, which is: How do you get people to believe you?

NOTE: This is not a discussion about policy. Please do not derail it into a discussion of whether AI use should be allowed or disallowed. This is about how to get people to see what you are seeing, to trust that it is possible for people to identify AI-generated text with over 90% accuracy if you know what you're looking for, that there is actual research about what to look for, that the signs of AI-generated writing exist as an aggregate pattern over millions of words in thousands of articles, that I have both read said thousands of articles and analyzed the text patterns in them alongside many other people who have done the same, and that whatever article I am tagging falls dead-center within that pattern (because when they don't, I don't tag them). This question would be the same regardless of whether AI is allowed or not.

But it seems like all I ever get is people telling you to shut up, essentially: shooting the messenger by complaining that there is now a template on "their" article, asking you to somehow track down a blocked editor from 2023/2024 to swear they used AI on oath, or the latest charming example: maybe i'll make this a Simple English wiki article. (I guess their calling me stupid is OK since didn't say the dreaded unspeakable F-word?)

What is the sequence of words that will get people to actually listen to you. I would love to know. Gnomingstuff (talk) 12:13, 6 November 2025 (UTC)[reply]

And yes, I am incredibly frustrated because it seems like you can't win. If you tag articles, people complain that you're not pointing out the exact issues (when, for the millionth time, you don't know the exact issues without thorough review; you also don't know what an unsourced statement should be sourced to or whether it's wrong without researching and yet no one complains about pointing out unsourced text). If you point out the issues, people ignore you no matter how clear-cut the evidence provided is. If you fix the issues, people complain that you've destroyed their work. It feels like the only acceptable action to people is to shut up. Gnomingstuff (talk) 12:18, 6 November 2025 (UTC)[reply]
The answer is the same as it was last time: Don't. Ignore the fact that it is or isn't generated by AI and instead focus on why the words on the page would be problematic if they weren't written by AI. For example if there are verification problems, then the problem is not that it was written by AI the problem is that it fails WP:V so explain why it fails verification (and if necessary why failing verification is a problem). If the problem is that the text is waffly and rambly, then the problem is with the writing style so explain that the problem is that the text is waffly and rambly, why this is a problem and rewrite 1-2 sentences yourself to show the difference. And above all, even if you are correct 90% of the time that it was written by AI, just remember that every 10th person you accuse of using AI is telling the truth when they say they didn't. Thryduulf (talk) 12:40, 6 November 2025 (UTC)[reply]

How do you get people to believe you?

I want to know if it is even possible at all. I'm starting to think it isn't. --Gurkubondinn (talk) 15:17, 6 November 2025 (UTC)[reply]
You and I have different approaches. I primarily track the 1325 and 1346 edit filters and have a high threshold for tagging, typically only when I can document multiple content verification failures on the article talk page. I don't think you should change your threshold though - you have a high accuracy rate and are basically single-handedly identifying the magnitude of the problem for the community. My suggestion would be if you run into one of the situations to just ping someone from AIC to come do a deeper analysis of the article for WP:V failures. I'm happy to do that for you. NicheSports (talk) 15:28, 6 November 2025 (UTC)[reply]
there is now a template on "their" article is probably the sum of it unfortunately. I was complaining over coffee and I'm not sure there's a quick solution. If people are getting WP:OWNERSHIPPY over an article, and someone drive-by tags something they missed, they might get defensive. If they can't even see the problem themselves then is it a surprise they revert the tag? Frustrating as all hell, but not surprising. I'd just re-revert with a standard edit summary featuring links to here, the LLM essay, stuff like that. No judgement, just a nice neutral "not all problems are immediately visible, please leave the tag unless/until you've done a thorough review". Until we have an actual AI/LLM/GPT policy they're not actually doing anything technically wrong by removing the tag.
BTW Thryduulf, straight question, do you think it's acceptable to use GPTs to write Wikipedia articles? Please don't waffle over definitions for official policy, I'm not asking your official admin position here, I'm curious what you think as an editor, and just as a person. ~ Argenti Aertheri(Chat?) 22:33, 6 November 2025 (UTC)[reply]
Yes. The reason for this answer is because it is completely and utterly irrelevant how people write Wikipedia articles. If the content they submit meets our copyright, quality and notability standards (WP:V and similar are included under quality here) we should accept it regardless of whether and/or how LLMs or any other technology was used. If the content they submit doesn't meet the copyright and/or notability requirements then we don't want it regardless of any other factors. If they content they submit meets the copyright and notability standards but not the quality standards then it should be cleaned up (if reasonably practical) or deleted (if cleaning up is not reasonably practical). Thryduulf (talk) 00:19, 7 November 2025 (UTC)[reply]
For the avoidance of doubt, I think that acceptance of AI as a tool for writing is inescapable, whether we like it or not. However, unlike the other tools that made writing possible for people who would have had hard time writing otherwise (text editor, spell checker), generative AI is just too fast for humans to check its output. IMHO, part of Wikipedia success is based on a fact that in the past curtailing unacceptable activity here was (marginally) easier than actually engaging in this activity, so a game of whack-a-mole was fun. Generative AI puts us squarely against a machine - a game humans lose - so what we need to think of is either:
  1. limiting the rate of incoming text to the level we, human editors, can handle, or
  2. engaging AI in checking (think of filters on steroids).
I think of an RfC proposal in the context of #1. When thought of as such, alternatives to the outright LLM ban are obvious: just like WP:VPN, institute a special flag that allows some editors to use an LLM, a privilege that shall be hard to earn and easy to lose. Викидим (talk) 20:44, 7 November 2025 (UTC)[reply]
That's not what's actually happening though. In practice generative AI is just too fast for humans to check its output, which is relevant because people are, quite frequently, posting things that look like they do meet our standards until you actually read the linked journal article (or whatever). Then you discover it either doesn't exist at all; makes no mention of the topic; or, my personal favorite, says the exact opposite of what the LLM claimed it said. At which point you have 2-3 options, ignore the code smell, at least tag that there may be other LLM issues, or spend an hour+ verifying everything else that editor added to that article (at minimum). Just deleting the whole article is rarely even an option, and cleaning up may require a subject matter expert. Also, LLMs pretty much by definition violate WP:SYNTH, and the copyright questions are concerning. The technology's just not ready yet, and allowing new/anonymous users to use it, when every clause must be carefully checked for copyright and synthesis, is causing far more issues than its solving. ~ Argenti Aertheri(Chat?) 19:48, 10 November 2025 (UTC)[reply]
I have been experimenting with AI a lot recently, and know the ways to avoid the problems you have pointed to: (1) select the sources manually (2) upload the texts of sources to AI (3) in the prompt, specify explicitly to use "only the sources provided". The results, on many topics, is surprisingly good (still not good enough to be posted "as-is", but requires mostly cosmetic works and checking of the page numbers (I actually have to do the same after creating a new text all by myself). The "Deep Think" option of the Google Gemini even gets most of the pages numbers right. Instead of a blanket ban, I would therefore suggest requiring a "code of conduct" and a special permission, akin to "extended mover". Викидим (talk) 20:08, 10 November 2025 (UTC)[reply]
I would say this 'mitigates' rather than 'avoids' the problems. In my experience, genAI remains lazy and inaccurate in how it uses provided sources, especially as increase the length / number of sources. ChatGPT also seems to go through phases of being more thorough or lazy at finding info from uploaded texts. The best option I've found is NotebookLM, which still tends towards inaccurate / questionable summary in places, but at least links directly to the text chunk the point is based on so can verify and read the point within original source - which can also help surface other issues in the LLM summary, such as the tendency towards overgeneralisation and removal of important contextual information. AbeBRStew (talk) 12:34, 17 December 2025 (UTC)[reply]
No, the best option is to write your Wikipedia contributions yourself, WP:CIR after all. --Gurkubondinn (talk) 12:46, 17 December 2025 (UTC)[reply]
Indeed. I meant more in context of best option when using LLMs with files and not using LLMs to directly write the content to use. Even with NotebookLM having lower hallucinations etc, the text it churns out remains obvious slop. The real strength of NotebookLM is the semantic search and links to relevant sections from the sources. That is something could replicate with just text embedding models and remove LLMs from the process. AbeBRStew (talk) 14:12, 17 December 2025 (UTC)[reply]
A boilerplate, wikilinked claim of authority in the edit summary might help the first impression (WP:AICLEANUPTEAM revert of additions from User:Example), maybe linking to some specific and common telltale signs (strong match on WP:AICR#1, WP:AICR#3, WP:AICR#7). If you don't convey an opening reason or any authority for thinking that it's AI, they'll feel more free to equally strongly claim that it's not, and it'll be harder to convince them from those two starting points.
I've been on the other side of this at Commons, where deletions of unusual-looking image files can go through on a few blunt "this is obviously AI" votes, and it's frustratingly unclear whether those users are providing an experienced eye in a hurry, or are just quick to assume that any image oddities they haven't encountered before can only be, in 2025, a result of AI. Belbury (talk) 09:21, 11 November 2025 (UTC)[reply]
I've been trying this: pointing to WP:AISIGNS (whether in edit summaries or talk pages or both), indicating the specific sections of WP:AISIGNS, indicating specific passages of text which fall cleanly into those sections (which I generally don't like doing because I don't want the takeaway to be that they should just reword those passages and problem solved). Unfortunately it doesn't seem to be helping.
The other frustration to all of this is that it makes tracking difficult if some unknown percentage of the tags are just going to silently disappear, although I could probably make tracking pages for the larger clusters of articles. Gnomingstuff (talk) 07:49, 19 November 2025 (UTC)[reply]
I think if there's any useful cluster at all (i.e. not just a single article), not even necessarily larger ones it's probably worth creating a section on WP:AINB just so it gets archived somewhere. If it's just a single article the decision is harder of course. Alpha3031 (t • c) 08:12, 19 November 2025 (UTC)[reply]

Idea lab: potential additions to G15 criteria

[edit]

Note: this is not an RFC. I'm looking for feedback and suggestions about these, and potentially other, additions to the G15 criteria. I G15 nom a decent number of articles so have been noodling on this for a bit. A few editors have also suggested potential G15 additions this week (@Femke, @Athanelar, @Tamzin) so seemed like the time to kick this off.

Workshop additional criteria

[edit]
  1. Multiple material source-to-text integrity issues + WP:AISIGNS: This may include, but is not limited to: quotations that are not present in the supplied source, unsupported statistics, or material examples of unsupported information. At least two such errors must be present. If more than one source is supplied for a given claim, the claim must fail verification in all supplied sources. The issues must be documented on the article or draft talk page prior to nominating for G15. The article or draft must also show signs of LLM-generated content, but these do not need to be documented.
  2. Author-acknowledged LLM use with evidence of insufficient review: This applies if the author has acknowledged that an LLM was used for the article or draft, and that article or draft demonstrates at least one example of WP:V, WP:NPOV, WP:SYNTH or other content policy violation. The content policy violation may be minor. Both the diff of the author's LLM disclosure and the example of a PAG violation must be documented on the article or draft talk page or supplied in the G15 reason parameter. (Note: potentially contentious section following) The LLM disclosure does not need to specifically refer to the article or draft in question, but the LLM disclosure must plausibly cover the time period of the article or draft's creation. For example, if an editor created four drafts in a short time period, all of which display signs of unreviewed LLM usage, but they only acknowledged LLM usage on one draft, all four drafts could be nominated under this criteria.
  3. Presumed deletion of article or draft by sanctioned editor: This is inspired by CCI's WP:PDEL. An article or draft that shows WP:AISIGNS and was created by an editor subsequently sanctioned for LLM misuse, WP:UPE, and/or WP:Sockpuppetry may be nominated for G15. No documentation is required.

NicheSports (talk) 03:27, 25 November 2025 (UTC)[reply]

Re: point 3 - WP:PDEL is... not as easy to use as it first seems. Those of use at CCI tend to only use it for blocked contributors, occasionally long-inactive ones, it normally involves simply removing the text from the page or stubifying it. Full deletion must go through WP:CPN - and I only send articles through there if there's a high rate of vio, I can't access the source, and/or an established pattern of source fraud, and it has to remain tagged and blanked for at least a week. And even then people get stroppy, so I try and make sure there's a few instances of failed verification as well.
So, yeah, PDEL is used sparingly, it's not a CSD criterion, and the deletion itself typically has to be sanctioned by a copyright clerk or experienced admin, and for my time at CCI, that experienced admin has normally been Mer-C, who is awesome and seemingly has the thickest skin ever to handle all the "how dare you delete my precious content that I copied from a blogspot post in 2008 you should have just rewritten it instead why won't somebody think of the childrenreaders"-type complaints. (Which, to be clear, many people who have CCIs opened on them can actually be wonderful & valuable contributors/members of the community! But the negative stuff, like being reverted by admins or accused of vandalism or being a deletionist sticks in your mind more)
CCIs I'm doing, in case anybody thinks I'm not sympathetic towards their efforts to clean up unusable content. GreenLipstickLesbian💌🧸 04:20, 25 November 2025 (UTC)[reply]
Thanks for the feedback GLL. Any comments on the first two? I might drop you a note on your talk page tomorrow with some questions about #3. I'm surprised by the difference in deletion... difficulty?... between PDEL and G15. I G15 articles frequently and no one has ever complained, although I did learn to avoid WikiEd! NicheSports (talk) 05:33, 25 November 2025 (UTC)[reply]

Discussion

[edit]
  • Pinging editors who participated in Wikipedia talk:WikiProject AI Cleanup/Archive 3 § Idea lab: New CSD criteria for LLM content: @Ca:@Thryduulf:@LWG:@Chipmunkdavis:@Jumpytoo:@Fifteen thousand two hundred twenty four:@Newslinger:@Chaotic Enby: Please add anyone I missed or who you think should see this, such as admins who handle a lot of SD noms — Preceding unsigned comment added by NicheSports (talk • contribs) 03:32, 25 November 2025 (UTC)[reply]
  • Before adding anything to a speedy deletion criterion you need to establish that the proposed criteria meet all four of the WP:NEWCSD requirements. At first glance, most of this looks very subjective and so completely failing point 1 (note especially that AISIGNS is completely subjective and full of caveats and exceptions - it seems fundamentally incompatible with speedy deletion). I also have my doubts about frequency and non-redundancy - are there really that many pages that should be deleted according to XfD consensuses, which would meet one or more objective criteria, and which do not meet G15 or any other existing speedy deletion criteria? Thryduulf (talk) 04:00, 25 November 2025 (UTC)[reply]
  • I will second Thryduulf on this one. WP:AISIGNS, while certainly a very helpful tool, is too subjective to be relied on in a CSD. Point 2 is also worrying as something like "failing WP:NPOV" can be interpreted differently by different editors (remember, content disputes are a thing). That point could also be interpreted very widely: if an author uses LLMs in one article (from what I understand, to any extent), even their non-LLM-written articles would fall under it if they have minor policy violations Chaotic Enby (talk · contribs) 09:05, 25 November 2025 (UTC)[reply]
    @Chaotic Enby I am surprised by your feedback about WP:AISIGNS being too subjective to be relied on in a CSD. Several commonly invoked CSD, such as G11 and A7, rely on subjective criteria. More importantly, AISIGNS is already a part of the current G15 definition, as a secondary factor, and forms a critical part of my checklist when I G15 something. For example, I would not G15 a draft that contains 3 references with broken URLs (meeting the "non-existent references" G15 criteria) if it was written in broken English with many grammatical errors, which would indicate the draft was not LLM-generated and the reference issues were caused by something else. In #1 when I specify multiple material WP:V issues + WP:AISIGNS I am just making that secondary check explicit to further restrict when the criteria is invoked. Can you please reconsider? Fair point on #2. I saw Tamzin [1] suggested something similar yesterday and this was my attempt to workshop it. Note that I didn't include Tamzin's 2nd suggestion in my proposed criteria above as I thought it would be too subjective. NicheSports (talk) 14:57, 25 November 2025 (UTC)[reply]
  • I can't really comment on #3, not having had any experience of presumptive deletion for copyright infringements. For #2, as Chaotic Enby said, things like WP:NPOV or WP:SYNTH can be open to interpretation, which makes "failing" them less suitable as criteria for speedy deletion. However, I could probably get behind multiple clear WP:V failures - I would say three or more rather than two or more - together with evidence of LLM use, be that WP:AISIGNS or a previous WP:LLMDISCLOSE, as an additional criterion for G15. I would like to see some evidence that articles like this are currently taking up too much time at XfD or AfC, so as to meet WP:NEWCSD #3, Frequency. Cheers, SunloungerFrog (talk) 09:47, 25 November 2025 (UTC)[reply]
  • #2 and #3 are fine but feel like edge cases. I don't think that #1 is viable since we can't even convince people to leave the tag on, even when the person has said they used AI. Even with three WP:V failures and clear signs of AI people will still be like "well I just write like that! and I only made 3 mistakes, people can make 3 mistakes!" Gnomingstuff (talk) 14:58, 25 November 2025 (UTC)[reply]
  • I like the idea in principle of #1, but it could be more suited as a G3 expansion than G15 as even if a human wrote the fake content it still needs to go. I also feel it needs to be tightened to only cover the case where it's a seemingly well-referenced article but it turns out the content was made up and not supported by any of the sources in the article. A lot of article creations have poor sourcing or mix up the sourcing and that shouldn't be grounds to hit the article with this criteria. I disagree with #2 as it would lead to more editors to not admit they are using LLM, and #3 should be left to a case-by-case basis at ANI or whatever. Jumpytoo Talk 19:10, 25 November 2025 (UTC)[reply]

For the interested

[edit]

Noticed at [2] (you might not see what I see) that Google's AI now uses Grokipedia as a source. What could possibly go wronger. Gråbergs Gråa Sång (talk) 13:19, 26 November 2025 (UTC)[reply]

Interestingly I see it citing the Wikiwand mirror of our article ... Sam Walton (talk) 16:58, 26 November 2025 (UTC)[reply]
Well, then they can tell Elon they're not using Wikipedia. Gråbergs Gråa Sång (talk) 17:16, 26 November 2025 (UTC)[reply]
comedy is now legal on google Gnomingstuff (talk) 21:00, 26 November 2025 (UTC)[reply]

Purpose of the project

[edit]

Many participants in this project believe all use of LLMs should be banned from Wikipedia. Because of this, I wonder if this project mission statement is still accepted: The purpose of this project is not to restrict or ban the use of AI in articles, but to verify that its output is acceptable and constructive, and to fix or remove it otherwise. SuperPianoMan9167 (talk) 00:37, 28 November 2025 (UTC)[reply]

Many participants in this project believe all use of LLMs should be banned from Wikipedia. - that isn't necessarily true, and no RFC has stated that. User:Bluethricecreamman (Talk·Contribs) 01:26, 28 November 2025 (UTC)[reply]
  • I support a full, total, comprehensive, all-encompassing, draconian, butlerian-jihadist ban on LLM usage on Wikipedia. No wiggle room for vandals
  • staunch LLM abolitionist. I'm doing my part!
These are from two of the entries on the participants list. SuperPianoMan9167 (talk) 01:33, 28 November 2025 (UTC)[reply]
WikiProjects typically aren't created to advocate for policy change, but to do a specific set of tasks. Thus I think it's true that the purpose isn't to ban the use of AI in articles. ("Restrict" is a bit of a greyer area; arguably ensuring compliance with Wikipedia guidance inevitably involves some restriction.) isaacl (talk) 03:58, 28 November 2025 (UTC)[reply]
"Remove all LLM created content" is arguably doing a specific task. We need to distinguish between content creation and simple automation tasks though, LLMs are pretty damn good at things that are just one step too complicated to fully automate, and can cut down on the copy-pasting substantially. ~ Argenti Aertheri(Chat?) 06:05, 28 November 2025 (UTC)[reply]
Yes, I didn't say it wasn't. My point is that I think the stated purpose remains applicable, independent of what policy changes might be supported by participants. isaacl (talk) 17:16, 28 November 2025 (UTC)[reply]
Speaking only for myself here...a lot of times "purpose" and "what we would like to see" are two different things. I myself have said I would love it if LLMs and the like simply disappeared from the earth. They're wasteful, they encourage laziness, and their output is mediocre in the few areas in which they're actually useful. But that's my opinion, and that's not the purpose of this project. I have no interest in ridding Wikipedia of every article touched by an LLM, I just want to make sure a certain minimum set of standards are met (factual content, references that both exist and support the citation, ideally language that doesn't sound like it was written by a smart 12-year-old with a thesaurus). And on topics that are notable. If all of those are met I do not believe this project requires LLM content to be removed, and I would argue with any user who said it does. WeirdNAnnoyed (talk) 12:44, 28 November 2025 (UTC)[reply]
As one of the people in favour of banning all LLMs, I think it's good we keep the mission statement as is and encourage participation from less restrictive editors, otherwise it would be impossible to meaningfully use this project for discussion and consensus-building without it simply being canvassing/vote stacking. Athanelar (talk) 14:38, 30 November 2025 (UTC)[reply]

Can someone check this edit?

[edit]

This, as well others from same editor... ~2025-36781-52 (talk) 03:49, 28 November 2025 (UTC)[reply]

It looks like someone has reverted it, as well as the other edits by HenriqueMinuit13, because they were all poorly done AI edits. I added {{Uw-ai1}} to their talk page. N7fty (talk) 18:28, 29 November 2025 (UTC)[reply]

I believe this article is AI generated - what should I do?

[edit]

https://en.wikipedia.org/wiki/Bedtime_procrastination ~2025-36782-44 (talk) 07:54, 29 November 2025 (UTC)[reply]

Why do you believe this? CMD (talk) 08:11, 29 November 2025 (UTC)[reply]
I read over the article, and while there's a few paragraphs that are written a bit awkwardly, I don't personally believe it has AI generated content N7fty (talk) 18:43, 29 November 2025 (UTC)[reply]
It does not look like AI-generated at all. Викидим (talk) 08:52, 4 December 2025 (UTC)[reply]

Supplemental essay for NEWLLM

[edit]

Regarding our new AI guideline, I've proposed a supplemental essay to solve the criticism of there being no consensus stsndard for identifying AI text. I thought I'd plug it here since the RfC has been stagnant for a couple of days now and, while I know there's no rush, I would like to get things moving here to avoid another month-long slog like there was with the original guideline proposal. Please look it over and vote in the RfC on the talk page.

https://en.wikipedia.org/wiki/User:Athanelar/Identifying_AI-generated_text Athanelar (talk) 14:32, 30 November 2025 (UTC)[reply]

AI spelling correction websites

[edit]

Hello. I have a general question for this project: while I've never used LLMs on Wikipedia articles, are websites for AI spelling corrections useful for correcting typos on these articles? Thanks. sjones23 (talk - contributions) 05:56, 1 December 2025 (UTC)[reply]

Probably, although I've had llms confuse different spelling variants, and mess up where humans may also mess up on things like potentially outdated scientific names. CMD (talk) 07:00, 1 December 2025 (UTC)[reply]
Spelling and grammar checkers haven't historically required "AI" to function, and anything using a large language model has the possibility of predicting wrong and silently introducing errors. Use with great caution, if at all. fifteen thousand two hundred twenty four (talk) 07:17, 1 December 2025 (UTC)[reply]
In theory, if you actually review each change, maybe. In practice though? Just fix the ones you spot and move on, or tag it for copy editing. The WP:Guild of Copy Editors will get to it, likely pretty quickly for inline tags like template:copy edit inline. Personally I'd much prefer something be tagged correctly than potentially "fixed". That said, I cannot fucking spell and, with the exception of technical jargon, I've never needed more than the basic word processor built in spellchecker. The squiggly red line is actually generally correct, just targets the individual word, and is built into the visual editor, so why introduce AI? ~ Argenti Aertheri(Chat?) 14:04, 1 December 2025 (UTC)[reply]
Thanks again. While I've been an experienced copyeditor for many years (this month will mark my 19th year as a Wikipedian), I sometimes use word processors, AI spell checkers like Grammarly, and of course this project's visual editor page to correct errors and typos, though oddly enough, all of them are not really considered LLMs. sjones23 (talk - contributions) 06:25, 3 December 2025 (UTC)[reply]
I don't know that I've ever had to resort to external tools for spell checking on Wikipedia, let alone an LLM. Firefox (as with most mainstream browsers) already has a decent spellcheck feature built in. That said, I'm not usually copy editing, so I typically don't need a tool that can distinguish between the different varieties of English and whatnot. As for technical jargon, the only technical articles I contribute to are usually topics I am very familiar with. Should I ever need an external tool, I'd probably just paste the entire article into Microsoft Word. - ZLEA TǀC 06:44, 3 December 2025 (UTC)[reply]
Safari (the browser I'm currently using) also has a built in spellchecker to correct any errors. Of course, if I had to copyedit or rewrite a major section in any article, I would do it directly in the article and sometimes use either Microsoft Word (my word processor) or my sandbox as needed while checking for any potential copyvios. sjones23 (talk - contributions) 06:55, 3 December 2025 (UTC)[reply]
I do it to spell-check whenever I write in languages other than English (and Russian), and it works well (obscure terminology is always a problem, but it had fooled many human translators as well). For English I only do it to change text to British spelling (with occasional lapses - on my part, not AI). Since many spelling and grammar checkers (say, Grammarly) are in effect based on LLM-like machinery, I cannot see a problem in using the bigger and more modern engines. Викидим (talk) 17:51, 3 December 2025 (UTC)[reply]

About WP:LLM

[edit]

I have boldly tagged WP:LLM as an information page instead of an essay. Thoughts on this change would be appreciated. SuperPianoMan9167 (talk) 03:18, 4 December 2025 (UTC)[reply]

This change has been undone. SuperPianoMan9167 (talk) 00:02, 5 December 2025 (UTC)[reply]

User talk:Qcne/LLMGuideline has an RfC

[edit]
icon

Wikipedia:Village pump (policy)#RfC: Replace text of Wikipedia:Writing articles with large language_models has an RfC for possible consensus. A discussion is taking place. If you would like to participate in the discussion, you are invited to add your comments on the discussion page. Thank you.

This is an RfC to replace the current text of the guideline at Wikipedia:Writing articles with large language models with the draft guideline at User:Qcne/LLMGuideline qcne (talk) 11:33, 4 December 2025 (UTC)[reply]

Discussion at Wikipedia:Village pump (idea lab) § Scope of AI tool use

[edit]

 You are invited to join the discussion at Wikipedia:Village pump (idea lab) § Scope of AI tool use, which is within the scope of this WikiProject. Chaotic Enby (talk · contribs) 20:50, 5 December 2025 (UTC)[reply]

WikiEd project in biochemistry

[edit]

As we're all aware by now, WikiEdu projects are a vortex for LLM content. I just found the following WikiEd project covering biochemistry articles: [3], which assigned 55 different articles for editing. I would not be surprised if all of them contain AI output. I've only looked at one of them, CPLX1, so far and while it's not the worst I've seen it has some clear sourcing errors. I'm a biochemistry professor so I feel personally responsible for fixing these articles...if anyone else here has some expertise in the topic, your help would be greatly appreciated. Leave me a message and we can form a strategy.

Please allow me to apologize for other teachers who think doing things like this is a good idea. Making articles in draftspace? Fine. Mainspace edits should be off limits to WikiEd, in my opinion. WeirdNAnnoyed (talk) 12:12, 15 December 2025 (UTC)[reply]

Thanks for the heads-up. I have literally no expertise whatsoever in this topic, but I do have a lot of experience reading AI text.
I don't want to add more work to your plate -- especially during exams -- but if you or anyone else who does know the subject ends up compiling a list of articles with sourcing issues, would it be possible to post that here? These are almost definitely newer LLM edits, and the AI text indicators are changing fast. (unsurprisingly so; OpenAI has rushed out not one but two new versions in the past two months). Comparing this August edit by the editor of the CPLX1 article to the ones you flagged, the first edit pinged my AI radar almost immediately, the second not so much. So the more examples of recent AI edits we get, the more we can start to reverse-engineer the tells of them. Gnomingstuff (talk) 19:42, 15 December 2025 (UTC)[reply]
I will do that, but for right now there are no obvious "tells" that I've noticed yet. The obvious sourcing error I noted in the above article is a psychology textbook being cited for a statement about neurotransmitter vesicle trafficking...yeaaaah, that's not in any psychology textbook. A great illustration of the insidiousness of AI errors, stuff that looks just plausible enough if you're not a specialist. WeirdNAnnoyed (talk) 23:11, 15 December 2025 (UTC)[reply]
Students from this class have been repeatedly triggering the AI edit filters. I think you would have broad community support in reverting content that has even moderate signs of unreviewed LLM content. Btw, I completely agree with you that Wiki Ed needs to stay in draftspace, with a mechanism for pushing high-quality work to mainspace on an optional basis, after independent review. To move in that direction I have wanted to audit a few classes to estimate what % of Wiki Ed students are using LLMs. If that % is as high as it seems to be, we can take that analysis to Wiki Ed and suggest changes to their approach. If you are going to cleanup this entire class, 1) thank you so much 2) can you please document it sufficiently (with an emphasis on material WP:V failures that also show WP:AISIGNS) so we can use it as a data point? I have just enough domain knowledge to help you with this project. Feel free to ping me at my talk page if you want to take this on together. NicheSports (talk) 22:12, 16 December 2025 (UTC)[reply]
I could do that, but it looks like many of the edits have already been reverted by someone other than myself. But yes, I'm going to go through all the articles anyway and any changes I make will be documented in the comments (or in the talk page, if they're too long/complex for the comment box). WeirdNAnnoyed (talk) 23:44, 16 December 2025 (UTC)[reply]
This is my fault. We're getting notifications for likely-AI generated content and reverting it. I personally felt a need to separate good from bad when I could, and spent a lot of time checking references. I was deep in fact-checking biochem journal articles when things got to the crazy part of the semester, and a lot of edits from early November never got checked and reverted. I've removed what seems to be AI, and everything from students who seemed to use it heavily. Ian (Wiki Ed) (talk)/User:Guettarda 20:03, 17 December 2025 (UTC)[reply]
Thank you for doing that...if only all instructors were so responsible. In a way this has been a good thing because a lot of those articles needed work anyway, and this gives me a new project to focus on for a while. And if the students made good edits, so much the better. The article I finished this morning seemed to check out. So that leaves 50-some more to go. WeirdNAnnoyed (talk) 20:57, 17 December 2025 (UTC)[reply]

For the interested

[edit]

OpenAI has ditched the automatic model router for ChatGPT users on its free and $5 plans, so most users will now all use the same model by default. Not sure if this will affect cleanup or anything but this seems like pertinent information. SuperPianoMan9167 (talk) 15:46, 18 December 2025 (UTC)[reply]

Notified on the relevant noticeboard here. sjones23 (talk - contributions) 00:22, 19 December 2025 (UTC)[reply]
will admit I had no idea ChatGPT just sent queries to whatever model by default. makes WP:LLMDISCLOSE even harder Gnomingstuff (talk) 03:18, 19 December 2025 (UTC)[reply]

New citation markup?

[edit]

Draft:Reze (Chainsaw Man) (edit | talk | history | links | watch | logs)

This is the first time I've come across this. On rev 1328696648 to Draft:Reze (Chainsaw Man), there is something that looks like numbered footnote or citation marks. The draft doesn't include any definitions of these numbered tags. These are surrounded by 0xee 0xa8 0x81 and 0xee 0xa8 0x82: 

$ echo -ne "1" | hexdump -C
00000000  ee a8 81 31 ee a8 82                              |...1...|
00000007
$ echo -ne "2" | hexdump -C
00000000  ee a8 81 32 ee a8 82                              |...2...|
00000007

This seem like something similar to WP:OAICITE-style placeholders, but I have no idea what model this might be coming from. Has anyone seem this before, or knows where it is coming from? --Gurkubondinn (talk) 11:37, 21 December 2025 (UTC)[reply]

I've asked the submitter on Draft talk:Reze (Chainsaw Man) § Numbered tags, might be the easiest way to find out where this comes from. --Gurkubondinn (talk) 11:43, 21 December 2025 (UTC)[reply]
Seen also from prior LLMN guest NatHaddan in Special:Diff/1319055226. Based on turn0search1 type indicators in another edit of theirs, ChatGPT is my guess. fifteen thousand two hundred twenty four (talk) 12:07, 21 December 2025 (UTC)[reply]
Here's a search, swap the number to find other pages (a [0-9] range is too slow). It's been occurring since at least June, probably a similar mechanism as the oaicite issue as you've mentioned. fifteen thousand two hundred twenty four (talk) 12:17, 21 December 2025 (UTC)[reply]
Oh thanks. You are right, that's the same bytesequence:
$ echo -n "Peter Oloche David — Figshare. 0" | hexdump -C
00000000  50 65 74 65 72 20 4f 6c  6f 63 68 65 20 44 61 76  |Peter Oloche Dav|
00000010  69 64 20 e2 80 94 20 46  69 67 73 68 61 72 65 2e  |id ... Figshare.|
00000020  20 ee a8 81 30 ee a8 82                           | ...0...|
We should probably add this to WP:AISIGNS. Both of these examples are in draftspace, is there a way to request that these drafts to be preserved for the puposes of being included in documentation? --Gurkubondinn (talk) 12:21, 21 December 2025 (UTC)[reply]
{{G15-exempt}} is what other pages at WP:AISIGNS use. fifteen thousand two hundred twenty four (talk) 12:23, 21 December 2025 (UTC)[reply]
Thanks, I've tagged both Draft:Reze (Chainsaw Man) and Draft:Peter Oloche David. --Gurkubondinn (talk) 12:33, 21 December 2025 (UTC)[reply]
Author is requesting A7 for Draft:Reze (Chainsaw Man), probably best to make attributed copies of the drafts to a subpage somewhere else first. fifteen thousand two hundred twenty four (talk) 13:04, 21 December 2025 (UTC)[reply]
Copied it to User:Gurkubondinn/Draft:Reze (Chainsaw Man) for now, attributed to the user and revision ID in the edit summary. --Gurkubondinn (talk) 13:34, 21 December 2025 (UTC)[reply]
Unfortunately it was deleted. Not sure why, it was {{G15-exempt}} tagged. --Gurkubondinn (talk) 12:25, 25 December 2025 (UTC)[reply]

Fewer hallucinated references

[edit]

In the past month I am seeing fewer broken URLs or hallucinated ISBNs while patrolling the edit filters. Has anyone else noticed this? NicheSports (talk) 01:15, 23 December 2025 (UTC)[reply]

Be careful of the base rate fallacy. Have there been fewer edits in general this past month? ([4] doesn't have December numbers yet.) Apocheir (talk) 02:49, 23 December 2025 (UTC)[reply]

Phage therapy

[edit]
Moved to Wikipedia:WikiProject AI Cleanup/Noticeboard#Phage therapy
 – fifteen thousand two hundred twenty four (talk) 09:35, 23 December 2025 (UTC)[reply]