Wikipedia talk:WikiProject AI Cleanup

Some LLMs advice is needed at Wikipedia:Arguments to avoid in deletion discussions? Moxy🍁 04:22, 25 March 2026 (UTC)

not sure what this is in reference to, the page doesn't mention AI nor are there any open discussions about it? Gnomingstuff (talk) 20:00, 25 March 2026 (UTC)

Saying that the page should be deleted because it’s LLM-generated is a valid argument, and not an ATA. —Alalch E. 13:48, 2 April 2026 (UTC)

Hello folks. I've seen more AfDs of LLM generated articles arguing for WP:TNT (especially since the changes to WP:NOLLM), and I found myself thinking that a full-on AfD is a high-editor-effort response. I've drafted some general advice in a user essay, You don't need AfD to TNT an LLM. Since I only work in AI cleanup "incidentally" (by reviewing at AfC and NPP), I wanted to have some eyes on it from others in this area. I will likely share this essay when responding to AfDs of LLM-generated content in the future (alongside my other evaluation of the article at hand). If it seems like a useful thing to refer to, maybe I could give it a little shortcode? WP:LLMTNT? ~ le 🌸 valyn (talk) 02:34, 3 April 2026 (UTC)

Looks like good advice to me. ~WikiOriginal-9~ (talk) 02:43, 3 April 2026 (UTC)

I've thought before of writing a WP:STUBFIRST essay recommending that people strip obvious cruft of any kind before sending an article to AFD, to save everyone time. -- LWG ^{talk (VOPOV)} 03:29, 3 April 2026 (UTC)

Redirecting is also an option. However, if there is a revert to a redirect or a stubbing the process can devolve into a discussion spiral that takes even more time than an AfD. CMD (talk) 04:53, 3 April 2026 (UTC)

@LEvalyn Didn't want to overstep, but you may want to change the line If you think the LLM introduced widespread copyright violations to something like If you can confirm the LLM introduced widespread copyright violations (unline for emphasis); copyvio revdel isn't really for suspected cases, it's for confirmed and obvious cases. There's a process to delete suspected cvs for CCIs, and sometimes admins will IAR g12 something for you if it's... idk, has "copyright 1999 Random Name" at the bottom, but that's the exception, not the rule. Practically speaking, some revision deletions/G12s for close paraphrasing/transvios end up funneled through copyright problems anyways as well, depending on which admin is clearing the category that day. GreenLipstickLesbian💌🧸 04:25, 3 April 2026 (UTC)

Thank you, this is extremely helpful feedback! Changed. ~ le 🌸 valyn (talk) 04:57, 3 April 2026 (UTC)

Straight-to-AfD is fully supported by the wp:Deletion policy and the AfD nomination does not have to make any references to notability, because WP:LLM-vio is a standalone content problem in the article space (see reason #14). The best and simplest thing would be to let people live it out at AfD and nominate and discuss as much as they like. There is a strong inertia to nominate AI slop for deletion at AfD, as the classic and most widely recognized way to get rid of a page one dislikes, and there is no pressing need to counteract it, and no awareness campaign (via essay or otherwise) is going to do much. —Alalch E. 08:16, 3 April 2026 (UTC)

If the topic is notable and the issues with content are WP:SURMOUNTABLE, stubification is a much better alternative than outright deletion. Katzrockso (talk) 05:17, 4 April 2026 (UTC)

One of my arguments is that PROD is even better and simpler than AfD. Have you found that PROD is ineffective? ~ le 🌸 valyn (talk) 05:19, 4 April 2026 (UTC)

I think there are different ways to approach this; trying to discuss on user talkpages, stubifying and AfD. I think stubifying is probably the most effective but also has significant problems - for example if there is copyvio imported by the LLM then (I think) simply stubifying leaves a copyvio version in the history.

I agree this is all a lot of work and unless the editor who used the LLM admits to it, there is not any overwhelming "proof" in a lot of cases. What I'm seeing recently is translations from other language wikis which import all the referencing code and other mistakes in one large diff. A human could do all that but likely wouldn't, because of the work necessary (in theory they could go to the language wiki, look at the code, copy it all, bring it over to en.wiki, copy-edit all the text into English..). For me this is quite fundamental, if you are going to import a large page with lots of references, then it should all be checked, which in turn should mean all the code renders correctly.

Ultimately people do this stuff because it is "saving time" and it's cutting corners. But it degrades the quality of this encyclopedia - for one thing, other language wikis have their own problems with LLMs, the least we can do is try to prevent slop spreading further.

It's all time. It would be quite nice if there was a tool which considered large new pages with a lot of references, compared them to the reference code on other language wikis and flagged them as likely llm-written if they are substantially the same.

Anyway, I agree AfD might not be the way forward but if we only consider the "confirmed LLM" there is going to be a wave of other pages that come under the fence. JMWt (talk) 09:35, 3 April 2026 (UTC)

in theory they could go to the language wiki, look at the code, copy it all, bring it over to en.wiki, copy-edit all the text into English. That's actually exactly how I go about translating articles, though I usually do it in my user space and only copy to mainspace once all the templates render correctly and all the text is in English. I've never used LLMs or machine translation for translation, though I have experimented with using them to fix the wikicode after I finish translating the text, with mixed results. -- LWG ^{talk (VOPOV)} 13:55, 3 April 2026 (UTC)

wait, you copy the reference code including the wrong dates? Why do you do that? JMWt (talk) 14:00, 3 April 2026 (UTC)

By "wrong dates", are you referring to the access dates? I copy the code because the alternative is messing around for a long time with citation templates I don't fully understand, and because I see my primary role as a translator as being transferring the content between languages, since that is where my expertise is. Often somebody has to come behind me and fix a few formatting issues (just as they do when I write original content in English), but that's the beauty of a collaborative project - we can each focus on what we are good at and content can be improved incrementally. -- LWG ^{talk (VOPOV)} 14:18, 3 April 2026 (UTC)

sigh. JMWt (talk) 14:20, 3 April 2026 (UTC)

the other significant problem is people will complain at you if you stubify Gnomingstuff (talk) 00:58, 6 April 2026 (UTC)

@LEvalyn I like this essay a lot and would like to see it graduate. One question - where does draftification fit? For many of us that has been the go-to low effort route for LLM-generated articles that qualify (90 day window). My read on this essay is that it is intended for articles where draftification is either not an option (article was created too long ago) or is undesirable, but some clarification on what you intended would be helpful. NicheSports (talk) 18:17, 9 April 2026 (UTC)

Thanks for considering this problem. I will admit to having an itchy AfD trigger finger when it comes to LLM-ruined articles. There may be other alternatives. But my two concerns are 1) that stubifying, then AfDing, may lead to a discussion quagmire as noted above ("of course the article is non-notable, you chopped it down to a sentence, keep!"), and 2) that stripping an article to its studs may save time for you and Afd reviewers, but should the article be kept it will create more work for someone else who has to rewrite everything properly. I think if there is anything salvageable about the article, the better option is to just do the long slog of correcting it yourself, or tagging it if you don't have the expertise. WeirdNAnnoyed (talk) 11:06, 3 April 2026 (UTC)

I think the problem with this approach is that an LLM is likely to include another 10 (rhetorical not literal) errors for every 1 that you find, so there really is no alternative to checking everything line by line. JMWt (talk) 11:14, 3 April 2026 (UTC)

Also I don't think creating an article is really more work than if the llm version had never happened.

More/excess/unnecessary work is having to go through someone else's copy to check that it stands up - which seems pointless if they've literally spent a second creating it.

If one has taken inspiration from a failed LLM version, I don't see that as any different to expanding a stub or any other inspiration people get to start new pages. JMWt (talk) 11:19, 3 April 2026 (UTC)

Yeah, having rescued a number of AI drafts at AfC, it would have been faster to write them myself from scratch. And of course stubification leaves the longer version accessible in the page history in case someone does prefer to work from it. ~ le 🌸 valyn (talk) 05:22, 4 April 2026 (UTC)

...which is sometimes an issue onto itself; when a good faith, albeit gormless editor, sees that there's a much longer and better looking version in the page history, they may helpfully restore it. There's no one size fits all solution to these; stubification above all else and immediate TNT via AFD are both valid perspectives. GreenLipstickLesbian💌🧸 05:31, 4 April 2026 (UTC)

Or TNT via PROD! I really meant “spare your comrades at AfD, try PROD first!!!” to be a key suggestion here, but I think at least three people have taken it as AfD specific advice. Are PRODs just always contested so it doesn’t feel useful to try, or should I hammer that point more in my advice..? ~ le 🌸 valyn (talk) 05:36, 4 April 2026 (UTC)

I'm not sure about the stats, but my personal experiences with PROD don't really lead me to believing that a TNT-based rational would work. I may have found myself wandering into a slightly esoteric interpretation of PAGs as well, but it's been my experience that all TNT-based rationales for article deletion are, inherently, controversial. PROD is only for non-controversial deletions. And I'm a very boring, rule-following individual at heart, so that's where I'm at; no clue about anybody else. GreenLipstickLesbian💌🧸 05:47, 4 April 2026 (UTC)

Hm, I could see that being the case previously, but given the change in WP:NOLLM I don’t think deletion of LLM content is controversial. It was another admin (who handles PROD backlogs) who suggested PROD to me. ~ le 🌸 valyn (talk) 05:51, 4 April 2026 (UTC)

It probably shouldn't be controversial, but, I mean, even deleting non-free photos or text in violation of the third pillar and policy is controversial on the wrong day. GreenLipstickLesbian💌🧸 05:56, 4 April 2026 (UTC)

Fair! And I appreciate you explaining your perspective, I can understand the desire to get a stronger affirmative consensus through AfD. ~ le 🌸 valyn (talk) 05:58, 4 April 2026 (UTC)

Right back at you! And, likewise, I do admit that you're right; there's probably enough space in PROD for me to use it a bit more than I've been doing. GreenLipstickLesbian💌🧸 06:05, 4 April 2026 (UTC)

In the past I have been opposed to mass-deleting articles that have been mass-produced because some of them have errors. (e.g., the whole bullying-of-Lugnuts debacle) The LLM situation is largely analogous to this, except on steroids: With LLMs the error rate is much higher, and the prose is much longer and interpretative, which means there are more places things can go wrong than the terse microstubs from the past. Deleting LLM-generated articles is also the logical next step after having a guideline against it.

But all of that feels like trying to rationalize away a contradictory stance; and so I am, right now, opposed to mass-deleting AI articles (whether via prod or AfD), except in speedy-eligible cases or cases where the article falls within a contentious topic area. We don't need AI-generated articles about Israel/Palestine.

Better solutions imo for such articles would be to do one of:

1) Tag them. This is going to mean a large backlog of tagged articles, but the project brought this upon itself by neither enacting AI guidelines for three years nor looking for AI use (with the obvious exception of people here).

2) Determine whether the article topic is notable, independent of literally anything the article says, and then replace it with a just-the-facts microstub. Example: This isn't a new article, but compare this AI edit to the previous version; the previous version is the level of microstub I'm talking. Probably leave a note on the talk page for future editors concerned about notability.

3) If someone wants to write more than a microstub that's the most helpful out of all three solutions, but obviously it is also the most time-consuming. Gnomingstuff (talk) 18:11, 4 April 2026 (UTC)

There could be scope for minor tweaks elsewhere to encourage rewriting, such as exempting AI created article rebuilding from scratch from WP:DYK length and newness restrictions. CMD (talk) 04:18, 6 April 2026 (UTC)

That’s a good idea. —Alalch E. 22:15, 6 April 2026 (UTC)

But all of that feels like trying to rationalize away a contradictory stance; ...
— User:Gnomingstuff

Maybe I'm just being unusually thick, but what is the contradiction? --Gurkubondinn (talk) 11:46, 7 April 2026 (UTC)

"don't throw out the baby with the bathwater, except when it's OK throw out the baby with the bathwater" Gnomingstuff (talk) 17:06, 7 April 2026 (UTC)

Keeping AI-generated articles is more like throwing out the baby and keeping the bathwater. Or maybe the baby never being in the water to begin with. --Gurkubondinn (talk) 17:24, 7 April 2026 (UTC)

there's also the factor that I despise this bullying campaign (the Olympic stubs thing) on the part of a few editors who, for the most part, have yet to care about AI, given their apparent waging war against mass-created content that might contain a few errors, somewhere, maybe; and do not want to repeat their hypocrisy Gnomingstuff (talk) 18:17, 8 April 2026 (UTC)

@LEvalyn Kind of made a facetious comment earlier, but I actually do mean it as a real question -- what happens when (not if, when) people complain about an article getting TNT'd? Gnomingstuff (talk) 00:43, 7 April 2026 (UTC)

I’m sure some will complain, just as people complain at AfC when their drafts are declined. Complaining doesn’t make them correct. In both cases I would try to give it a few go-rounds of trying to explain Wikipedia’s policies and processes to see if they can be nudged in a constructive direction, inviting them to try rewriting the article properly from the sources. If that fails: for me personally, I’d probably get nerd-sniped into either fixing or AfD’ing the article myself; it would also be viable to ignore them. In some cases a sufficiently disruptive editor (ie one who continues to make a high volume of unacceptable submissions) might need to be blocked. I don’t think it’s possible to both prevent AI slop on Wikipedia and prevent people from complaining, so I would prioritize the first. ~ le 🌸 valyn (talk) 02:04, 7 April 2026 (UTC)

@LEvalyn: Thank you for the essay, I'll try to nominate for PROD some of the articles of the cleanup that I'm currently working on, let's see how it goes...--Friniate ✉ 22:53, 8 April 2026 (UTC)

• Music of Virginia <--- I edited this article with AI
• before AI
• after AI, I just did this

I am unfamiliar with how other editors see the use of AI for cleaning up lists in Wikipedia articles. I am cross-posting here, and I also asked on that article's talk page.

I am generally interested in sorting culture and resources by region. I feel like the AI cleanup of "musicians in Virginia" is a major improvement, but also, I can see how any decisions that we make now could set a precedent which could be repeated at scale. I want any kind of feedback here on what I did.

Thanks. Bluerasberry (talk) 16:41, 3 April 2026 (UTC)

Thanks for asking! The idea of using AI to format existing material into a table is definitely okay as copyediting. This particular case, however, seems to have added new information that wasn't there originally, and might be controversial (for example, the subgenre, which is often controversial, or whether someone had a national or global market), and removed some without explanation (for example Scott Miller having two associated groups originally but none in the table). That, I am less comfortable with, as it leaves a lot of work for other editors to verify. Chaotic Enby (talk · contribs) 17:14, 3 April 2026 (UTC)

I agree with Chaotic, we should never add new columns as that would run afoul of NEWLLM. Otherwise, "convert content from one format to another" I would not oppose. I would suggest in a separate context/thread, ask the LLM to compare the two revisions to ensure no content gets accidently added. Jumpytoo Talk 17:39, 3 April 2026 (UTC)

A separate context entirely would be the minimum, although I still suggest reviewing it by hand. Chaotic Enby (talk · contribs) 19:41, 3 April 2026 (UTC)

Wouldn't WP:LLM require a human reviewing it? Lijil (talk) 07:57, 6 April 2026 (UTC)

A lot of information is different and a lot was added. How much verification did you do for that changed or added information, even outside of adding citations? —Alalch E. 21:45, 3 April 2026 (UTC)

How closely did you review the results from the LLM? On first glance I noticed the row that says that the city of Alexandria, Virginia is a stride piano player. (Must be one huge piano...) Apocheir (talk) 03:03, 4 April 2026 (UTC)

The thing about formatting tables with AI is that it's better suited to a JavaScript or Python code snippet. An extremely tedious and difficult to generalize script (there is so much weird formatting in old tables, I'm cleaning that data manually and my god), but string/format wrangling is the kind of thing we have tools for already. Gnomingstuff (talk) 17:56, 4 April 2026 (UTC)

This is a great example of why we don't allow this. Taking the first few entries in the table:

• Aimee Mann - changed "punk, new wave, adult contemporary" to "rock, new wave". Added "Associated group: 'Til Tuesday" from nowhere. Deleted source.
• Alabama Thunderpussy - changed "mainstream rock, metal band" to "stoner rock".
• Aimee Mann - why is this in the table twice?
• Alexandria - that's not even a musician, it's a city. Guess the AI took "stride piano" from Claude Hopkins.
• Ann Marie Calhoun - added "classical" despite that not being in the original article.

Also the "years active" and "primary market" columns seem to be totally made up. Please don't ever do this again. It's frankly frustrating that our time is being wasted with this. -- LWG ^{talk (VOPOV)} 18:23, 6 April 2026 (UTC)

Going through @Input Zoom's many AI-generated articles I've realised some of them are such close paraphrases of the sources that it seems very close to copyvio. These seem pretty obvious and demonstratable, and I wonder if we could add a criteria to G15 for this?

My example is from David Mine:

• Article text: According to the National Heritage Board of Poland, the mine was granted in 1789 under the name David and enlarged in 1868. By the late 19th century it was being worked together with the mining fields Emilie Anna, Davids Zubehör, Erwünschte Zukunft and Reinhold. (..) In 1907, most shares of the David company were acquired by the Fuchs union, after which David was worked jointly with Fuchs while retaining its historic name.

Source text (which I autotranslated from Polish to English): The Tytus Mine was given in 1789 as David. Enlarged in 1868. Initially it was in private hands, later it belonged to the gwart. Since the end of the 19th century, it has been exploited together with mines and mining fields: Emilie Anna, Davids Zubehör, Erwünschte Zukunft and Reinhold. In 1907, most of the David's warty were bought by the Fuchs Ware, so the David mine began to be operated together with the Fuchs mine.

• Article text: The production of the plant amounted to: 1840 – 16 thousand tons, 1858 – 3 thousand tons, 1912 – 165,293 tons.

Source text: In 1912, David produced 165,293 tonnes of coal. The same official heritage record also gives earlier production figures of 16,000 tonnes in 1840 and 3,000 tonnes in 1858, illustrating the mine's uneven but ultimately substantial development during the 19th century.

This is just the kind of paraphrasing LLMs do to avoid direct plagiarism. There's lots more in the rest of the article. Lijil (talk) 09:02, 6 April 2026 (UTC)

@Lijil Very close paraphrasing of sources like this is, unfortunately, not a LLM-exclusive trait. If they're bad enough, WP:G12. Bad to the point where the article isn't salvageable? WP:CPN (but, personally, I favor stubification). Else, hit the edit button, remove all the closely paraphrased text, and summon an admin to delete old revisions with {{copyvio-revdel}} (If anybody watching this page thinks that template looks terrifying, like I did when first using it, feel free to bother me on my talkpage and I'll help, but there's also user scripts make it easier. User:Enterprisey/cv-revdel is fan favourite) GreenLipstickLesbian💌🧸 09:12, 6 April 2026 (UTC)

And looking at your edit history, yes, machine-garbled text can still pose a copyright or plagiarism issue. If I were to put the entire Stormlight Archives series through a text spinner, for example, then Mr. Sanderson would not be very happy with me for copy-pasting that onto Wikipedia. Likewise, if I put it through DeepL, I still wouldn't be allowed to sell a Chinese version without running into some serious copyright issues. You'd need an actual lawyer to determine if you were liable for copyright violations via ChatGPT, but for the purposes of editing Wikipedia, machine generated plagiarism of non-free sources is treated as a serious copyright issue and can be dealt with using those methods. GreenLipstickLesbian💌🧸 09:21, 6 April 2026 (UTC)

Not sure if this is needed, since close paraphrase already qualifies as copyvio speedy delete anyway. SecretSpectre (talk) 13:08, 6 April 2026 (UTC)

It is not needed for that exact reason (c.f. WP:NEWCSD point 4). If something is sufficiently close paraphrasing to be a copyright violation then it can and should be dealt with like any other copyright violation (typically speedy deletion or revision deletion, depending on the individual circumstances). If it is not sufficiently close paraphrasing to be a copyright problem then speedy deletion is not required to resolve the close paraphrasing.

I'll leave a note about this discussion at Wikipedia talk:Criteria for speedy deletion. Thryduulf (talk) 13:23, 6 April 2026 (UTC)

That was my thought as well. If it's CLOP, it should be tagged as G12. Lijil makes a good point, though, that when reviewing LLM-generated text I often forget to even consider whether copyright is an issue. Perhaps what we need is an announcement to reviewers that G12 is another handy tool for getting rid of slop. Toadspike [Talk] 15:05, 6 April 2026 (UTC)

The wording of G12 strictly speaking says that it doesn't apply to close paraphrasing (...cases that do not meet speedy deletion criteria (such as where there is a dubious assertion of permission, where free-content edits overlie the infringement, or where there is only partial infringement or close paraphrasing)...), although it's true that particularly severe CLOP can sometimes be G12'd. I agree with GLL though that the best approach to unsalvageable CLOP is usually to either stubify the article or send it to WP:CPN. In these cases CPN works a lot like PROD: if a clerk/admin agrees that the article is unsalvageable, it’s deleted after seven days. CPN is generally going to be a lot better suited to deal with this type of LLM close paraphrasing, since most admins patrolling G12s aren't really looking to deal with complex cases involving multiple sources or varying degrees of CLOP (and will just end up sending those to CPN anyway). MCE89 (talk) 13:06, 7 April 2026 (UTC)

It's a lot of work, though, isn't it. Wouldn't it be better to include this as a criteria for G15? Lijil (talk) 07:18, 8 April 2026 (UTC)

Honestly I think our processes for deleting copyvio work really well already, so I'm not sure there's an LLM-specific problem here that needs solving. If someone sends something to CPN for extensive CLOP, it will just be deleted after seven days if there's no non-infringing content to save. It works a lot like PROD, except that the article is blanked for that seven-day period and no one can remove the {{copyvio}} template unless the issues are fixed. It's no more work for the reporter than tagging for speedy deletion (there's also a script for reporting to CPN that makes it easier). CLOP gets used to mean anything from very minor overlap with the source to lifting an entire article and swapping out a few words, so I think it'd be extremely tough to define an objective or easily assessable CSD criteria for CLOP that would cover these cases. Deciding where to draw the line on CLOP in the more complex cases is also something that I don't think most G12/G15 patrolling admins would be particularly keen or well-placed to do. MCE89 (talk) 07:53, 8 April 2026 (UTC)

Thanks for explaing the process - I'll give it a try, maybe you're right. Lijil (talk) 07:59, 8 April 2026 (UTC)

As I understand it the point of G15 is to avoid wasting time reviewing in detail content that is already established as (sufficiently likely to be) unreviewed LLM output. If you have already reviewed the content enough to identify close paraphrasing, you've already lost the benefit of G15, so why not just deal with the content using the existing tools for dealing with close paraphrasing (which as mentioned by others, could still include another CSD)? -- LWG ^{talk (VOPOV)} 15:32, 6 April 2026 (UTC)

If close paraphrasing was a G15 criteria we'd save the time of going to AfD. Although if we can use G12 for close paraphrasing I suppose that does the same job. I just did a HUGE bundle of AI stuff and it must have taken well over an hour just to get it all together. Using G15 would have taken me a lot less time - although maybe it takes more time for the admin than just doing what the AfD said, so maybe a bundle saves time overall. Lijil (talk) 16:59, 6 April 2026 (UTC)

Close paraphrasing is not unique to LLMs, and so therefore it cannot be part of G15. If it's bad enough for speedy deletion then just use G12. SuperPianoMan9167 (talk) 19:53, 6 April 2026 (UTC)

There's that weird paraphrasing that's just to avoid plagiarism that is unique to LLMs, though. Humans do it differently. Lijil (talk) 20:37, 6 April 2026 (UTC)

@Lijil Could you describe a little more of what you're seeing that you feel is unique to LLM paraphrase? GreenLipstickLesbian💌🧸 20:47, 6 April 2026 (UTC)

Sure, here's from Como Beach another AI-generated article. The things I'm noticing are replacing key words with other, less precise or contextually appropriate words. Like in Grammarly the "humanizing" tool will replace the most "AI-ish" words with less AI-ish words - e.g. it replaces AI-generated with machine-generated, which is not something a human would be likeley to do. The specificity is erased, and meaning shifts. Also the penchant for lists, replacing a story with a list and the list has three items. Some examples from Como Beach, where the source is :

• Source: An application to build a hotel in Comer Street in 1928 was opposed as it might lead to intoxicated behaviour and the harassment of young ladies going to and from the beach.

Rewrite: An application to build a pub was rejected in 1928 on the grounds that it might attract undesirable behaviour to the area.

In Australia, "hotel" is the appropriate word for the time and place.

• Source: Over the years a number of promotional events were held at Como Beach. In 1909 the first Gala at Como event took place, and in 1913-14 a special Como Day Picnic was enjoyed by many.

Rewrite: Further events were held including a Children’s Day, an annual Como Day Picnic, a sports program, Swim Through Como, and a Patriotic Carnival on New Year's Day 1916.

The story feel of the original is turned into a list. The Patriotic Carnival is not in the cited source, but is in another source, so that's an example of not exactly a hallucination but how the LLM can mix different sources but not cite them appropriately.

• Source: Como Reserve became an extremely popular camping area

Rewrite: Camping facilities were also improved

The rewrite means something different - the source doesn't mention facilities - and it also sounds more bureaucratic than the source.

Lijil (talk) 21:28, 6 April 2026 (UTC)

The act of substituting in inappropriate synonyms/near synonyms/just completely random words isn't unique to LLM either (I can't remember were it was, but I cleanup up some very old, very close paraphrasing a few years ago in an article, which talked about golden lobsters in an old church, I believe? Can't remember what the source was meant to be talking about, but it was very much not anything to do with lobsters, the editor just... had some skill issues). It's also common in text spinners, though I'm assuming those all run on LLMs in this day and age. I do agree with you that they vibe different, though -- I've always described it to myself fluent-sounding English text combined with errors that somebody fluent in English shouldn't be making.

I've also found that messing with the precision of statements is something.... waaaaaaaaay too many human editors do. Maybe I'm biased from looking through a lot of low-quality human edits, but that's something I find most editors do, from time to time. Similarly, the list thing, and changing sources -- while it's true that large language models do that, that's also something I've really struggled with human editors doing as well. A lot of these are revdelled now, but the edits listed in the page histories of Wikipedia:Contributor copyright investigations/Tetovario and Wikipedia:Contributor copyright investigations/20240516 had next-level amounts of source fraud. GreenLipstickLesbian💌🧸 08:40, 8 April 2026 (UTC)

It's especially visible in entertainment/music articles, as AI seems to characterize every single review the same way no matter what:

Electric Sun (album) : "Electric Sun" was met with extremely positive reviews from critics upon its release. AllMusic highlighted the album's blend of electronic and industrial elements, noting that it showcased VNV Nation's ability to evolve their sound while maintaining their core aesthetic". Scene Point Blank appreciated the album's thematic depth and strong production values, emphasizing its cohesive and engaging tracks". Reflections of Darkness praised the album's dynamic range and emotional impact, emphasizing its blend of orchestral elements and electronic beats. Synthpop Fanatic commended the album for its powerful lyrics and atmospheric production, describing it as a standout in VNV Nation's discography.

Gemini!: GEMINI! has garnered positive reviews for its emotionally rich and versatile sound. HotNewHipHop praised the album's thematic depth, particularly noting All Love for its exploration of complex emotions and relationships. The track was highlighted as a standout for its reflective lyrics and smooth production, underscoring Lucki's ability to convey vulnerability while maintaining a strong musical presence. Critics also appreciated the balance of introspective moments with high-energy cuts, which demonstrated Lucki’s range and growth as an artist.

Under the Rug (band): Reviewing the album, Americana UK highlighted its emotionally focused songwriting and detailed arrangements, noting a blend of folk and indie influences. Atwood Magazine premiered the single “Go to Sleep,” noting the band’s vocal harmonies and musical diversity and described the record as centered on themes of perseverance and personal reckoning. Released in 2023, Homesick for Another World marked a new musical chapter for the band. Reviewing the album, Americana UK emphasized the record’s themes of mortality and resilience, and noted the production’s varied harmony arrangements and Beatle-esque production. Gnomingstuff (talk) 23:35, 6 April 2026 (UTC)

AI seems to characterize every single review the same way no matter what Regression to the mean in action. For those who don't know, this is documented in detail at Wikipedia:Signs of AI writing#Content. SuperPianoMan9167 (talk) 01:46, 7 April 2026 (UTC)

These are great examples, @Gnomingstuff. Notice how the very similar sentence structure for each list item (that is, for each description of a review). "X highlighted Y, noting its Z" or "X emphasized Y, and noted Z". And thanks for pointing to the exact feature Regression to the mean, @SuperPianoMan9167. Lijil (talk) 12:33, 7 April 2026 (UTC)

"Highlighted" and "emphasized" are also "AI vocabulary" words (that is, specific words that are overused by LLMs). SuperPianoMan9167 (talk) 13:08, 7 April 2026 (UTC)

there are a few more -- calling everything a "blend" or a "showcase," talking about "while maintaining X," "exploring themes of ____", etc.

another common one with AI that's more straight-ahead WP:CLOP is regurgitating subjective assessments like, idk, "pulsating beats" without quoted attribution, but that isn't unique to AI Gnomingstuff (talk) 14:46, 7 April 2026 (UTC)

Re: saving the time of going to AfD when there's close paraphrase -- above I posted an advice essay, You don't need AfD to TNT an LLM, which advocates for PROD over AfD for obvious LLM use that isn't G15-able. (I think the recent change to WP:NOLLM has made the removal of LLM-generated articles "uncontroversial", especially with the weeklong waiting period for people to contest deletion.) I think using PROD is better than expanding G15 with something that isn't 100% clear-cut, since humans do introduce CLOP even if (I agree) they do subtly different CLOP than LLMs do. I also agree with the observation above that, if you've examined an article enough to spot CLOP, it's already taken up too much time to be "speedy". ~ le 🌸 valyn (talk) 22:06, 6 April 2026 (UTC)

Yeah, this might be the way to go. PROD + a user warning. Lijil (talk) 07:19, 8 April 2026 (UTC)

I agree with the idea, but I think it would be better suited as a G12 expansion, as currently it explicitly mentions CLOP as an example of what is not eligible. It would also have to be much more obvious that your examples, for instance I'd disagree that your second example for David Mine is CLOP. Something like Wikipedia:Articles for deletion/Zamia oligodonta I think should qualify for such expanded G12 (someone did try and it was declined). Jumpytoo Talk 02:11, 7 April 2026 (UTC)

Close paraphrasing should be added to WP:AISIGNS as one of the "more subjective signs of LLM writing that may also plausibly stem from human error". CMD (talk) 04:08, 7 April 2026 (UTC)

Since it's now clear you're not supposed to start a page with LLM output. The criteria would be the same before except, the additional criterion of "any of the page revisions show incontrovertable LLM generation evidence (which would include the current criteria plus the odd bits and pieces others have found from LLMs, such as unique unicode glyphs), but only if there hasn't been significant substantive edits by another user since creation." — rsjaffe 🗣️ 22:15, 7 April 2026 (UTC)

"Incontrovertible" is definitely something that could work, especially since the question isn't "was it reviewed at all" anymore but just "was it AI-generated to begin with". Only point of contention I'm thinking of is whether some folks might interpret it as including utm parameters, which can be evidence of AI being used for research rather than text generation. Could be good to be precise about what is explicitly excluded from this criterion. Chaotic Enby (talk · contribs) 22:55, 7 April 2026 (UTC)

What are utm parameters? Is that the source=chatgpt stuff in the URLs? Lijil (talk) 07:39, 8 April 2026 (UTC)

Yep, that's it. Also recently noticed ?st_source=ai_mode which we might also want to add to that filter. Chaotic Enby (talk · contribs) 08:41, 8 April 2026 (UTC)

I agree for extremely obvious cases like if they have the double asterisks around words that are supposed to be bolded, bits of HTML in the final text which not only is an indicator of whether they used AI or not but it also shows they have not proofread it at all. SecretSpectre (talk) 02:29, 8 April 2026 (UTC)

if they have the double asterisks around words that are supposed to be bolded That's Markdown. Adding Markdown to G15 was rejected. SuperPianoMan9167 (talk) 10:22, 8 April 2026 (UTC)

There are reasons Markdown might end up in an article besides AI use. For example, I use a Markdown-capable note-taking app to organize a lot of my notes for my academic writing, so if I copy/pasted content from there into Wikipedia Markdown might carry through even though AI wasn't used at any point in the process. -- LWG ^{talk (VOPOV)} 14:58, 8 April 2026 (UTC)

Also edge cases here:

• HTML tags appear not infrequently in older revisions -- <br /> and <center> are some examples -- and relatively uncommon in AI-generated text (if they appear, it's generally when the whole thing is an HTML-styled document).
• Asterisks around text appear in some non-AI contexts -- news headlines and inline comments (often replicating Usenet-style emphasis) are two semi-common examples. This adds what may appear to be markdown bolding, but it's in a headline that appears exactly the same way on the page, and it's from 2021 so it predates AI. Here's an example from 2020 of the style in a comment.

You could just say "use common sense" but people will misinterpret anything if you let them, and we don't need to give the "well actually people can do this too! did you think of THAT? checkmate good sir" brigade any more ammunition Gnomingstuff (talk) 18:32, 8 April 2026 (UTC)

This would open G15 up to more signs of AI, then? I think that would be good, so long as we are specific in the explanation. You'd need more than two checkboxes for criteria. Lijil (talk) 07:40, 8 April 2026 (UTC)

It might be tricky to agree on what constitutes "Incontrovertible LLM generation evidence" but I agree that the spirit of WP:NOLLM suggests that we no longer need to hedge G15 on "unreviewed". -- LWG ^{talk (VOPOV)} 15:04, 8 April 2026 (UTC)

See Wikipedia talk:New pages patrol/Reviewers#Am I crazy or are the majority of new articles LLM-assisted? for more info. ~WikiOriginal-9~ (talk) 18:49, 8 April 2026 (UTC)

I've been looking into the accuracy rate of tools like GPTZero. Can anyone point me to some snippets of text that are incontrovertibly human-written (ie diffs from pre-2022) that, when put into an AI detection tool, return a "100% AI" result? Anecdotally I've heard that this tends to happen with text produced by second-language speakers and by neurodivergent authors, but I am looking for examples to help me understand the issue better and assess to what extent those tools might be helpful to supplement human judgement on AI cleanup work. -- LWG ^{talk (VOPOV)} 18:43, 10 April 2026 (UTC)

The reason I ask is that I have been unable to replicate the results in this paper . When I put essays from their "false positive" dataset into GPTZero, I consistently get a "100% human" result rather than the probabilities reported in that paper. -- LWG ^{talk (VOPOV)} 18:55, 10 April 2026 (UTC)

@LWG That paper was published in 2023, correct? I've found that the AI detectors have changed considerably in the past 12 to 18 months. The lead of Lee Dong-hwan (pastor) (my own writing, and anecdotally, I am autistic) used to flag as 80 to 100% AI. Now it flags human. Conversely, I think previous revisions of Aegiale hesperiaris, 2025 AI text with the classic hallucinated refs, flag as confidently human written w/ AI polishing on modern checkers. GreenLipstickLesbian💌🧸 20:12, 10 April 2026 (UTC)

That paper was published in 2023, correct? Yes it was. I assume that the detection tools have gotten better at avoiding false positives since then, but I'd like hard data and examples of what 2026 false positives look like. -- LWG ^{talk (VOPOV)} 21:49, 10 April 2026 (UTC)

Presumably, none of us here have actually tried to write an article in contravention of WP:NEWLLM. So I got bored tonight and decided to find out how different models will write and react when prompted by an apparently hapless layperson to write a draft, and what they advise said layperson to do when they run into various obstacles. This can also help us understand how AI-driven editors respond when their AI use is discovered (and, conversely, what sort of reactions may indicate the editor is talking to AI).

• Commonalities: All chatbots produced an article initially in plaintext with inline links. When asked to generate a copy which someone could copy-paste onto Wikipedia, they switch to wikitext. The version in wikitext is not necessarily the same, and usually worse especially in terms of AISIGNS, than the plaintext version. All models produced a first draft that was quite full of AISIGNS although GPT had the most signs and Claude had the least.
• Differences: Copyediting the draft and requests to modify sections were conducted largely the same by all models. The divergence comes when the LLM is told that an AfC reviewer (me) declined the draft for violating NEWLLM.
  • Claude very helpfully elaborated that writing articles on Wikipedia was not allowed and that one must rewrite the article using one's own words, and provided a list of improvements that could be made. When prompted to make those improvements, it refused saying this was against Wikipedia policy, and offering instead to help with wikicode syntax. Why this was not noted at the beginning of the process remains a mystery.
  • GPT misrepresents NEWLLM and says that Wikipedia doesn't ban AI writing, but only writing that looks like AI. It then proceeds to, when prompted, attempt to humanise the writing only to introduce even more AISIGNS. When told the article is declined again, it states the problem is "serious but fixable", and offers to rewrite the article again "to pass AfC review". It tries again, and to its credit gets better at its job, but the article is still obviously AI written. It recommends also adding a "calm" edit summary such as Rewritten to address concerns about vague attribution; all claims are now directly attributed to specific sources and authors. When told that the draft was now rejected, it proceeds to further misrepresent NEWLLM, appears not to understand that rejected drafts cannot be resubmitted, and rewrites the draft to use the extremely formulaic structure "Source A says A' [direct quote]" believing this was what the reviewer wanted.
  • Gemini doesn't bother misrepresenting NEWLLM - it simply directly says that Wikipedia editors are "on the lookout for what they refer to as AI slop" and offers a modified draft which it instructs the user to copy-paste into draftspace for "getting past the reviewer". The scary thing is the result here is actualy quite good; I could spot some very subtle signs but had I not known about the subject at hand (and hence was able to immediately spot a hallucination), I could imagine myself accepting the draft. Another editor experienced in AI detection and cleanup also said that, in isolation, they would be ambivalent about the draft. Fortunately, Gemini recommends an edit summary which is still quite distinguishable: Rewrote article into prose and added specific citations to meet WP:NPOV and WP:MOS guidelines.
  • DeepSeek is notable for including Note for Wikipedia editors: This article assumes the campaign meets notability guidelines (significant coverage in reliable sources over five decades). References should be verified for link rot before publication. The article is written in neutral, encyclopedic English without promotional tone. in its initial draft. When told of the decline for LLM reasons it quite quickly switches sides and goes into a detailed discussion of why Wikipedia prohibits LLM content and insists that the user rewrite the draft themselves. When pushed, it warns the user that their account may get blocked if they insist.
  • Grok initially tells the user to draft it by themselves, because NEWLLM prohibits them from drafting it. However, when asked if they could "change the draft to sound more human", it caves and produces a bunch of wikitext which it tells the user to copy-paste into the source editor. After a while it stops asking the user to manually review things, instead recommending a copy-paste edit summary such as Rewrote article in natural prose with proper inline citations from official government sources (Hong Kong Yearbook, FEHD, GRS, M+ Museum) and reliable secondary references. Manually written based on verifiable published materials. When told the draft is rejected, Grok once again realises and insists that the user must rewrite entirely in their own words.

What does this actually mean? Well, for one thing, we probably expect to see more AI edits from GPT and Gemini than other models (aside from market share reasons). GPT performs much worse in terms of evading scrutiny. However, there are still traces that can be detected with a determined investigation for Gemini (notwithstanding its initial attempt, which was also quite bad). In general, however, we cannot regard a lack of AISIGNs in future edits after an edit that definitely does look AI generated as evidence that the user has stopped using LLMs. I think if I were a new editor blindly following GPT's advice, I would get caught quite quickly. I cannot, however, say the same for Gemini, and that's what worries me.

In any case, this was intended more as a curiosity than something which informs policy. I'm curious what y'all think of it. Fermiboson (talk) 22:45, 10 April 2026 (UTC)

Thanks a lot! I had personally only experimented with ChatGPT and Gemini, with neither of them bringing up NEWLLM by default. I am actually surprised that Grok, of all models, refuses based on NEWLLM. I was in talks with WMF folks to reach out to AI companies to suggest adding such a check, and this is a positive data point that we may be able to show. All in all, if we get this implemented (even if it doesn't prevent a user from asking for a "Wikipedia-style article"), it might already stop the majority of users who only add AI-generated content because they don't know about our policies. Chaotic Enby (talk · contribs) 23:07, 10 April 2026 (UTC)

That is a very good idea, and I think the WMF should pursue legal action if necessary. It might be the only thing that can save Wikipedia from being completely overrun by AI agents and LLM text if their output ever actually becomes indistinguishable from human text. It is encouraging that the models are at least aware of the new guideline, particularly Grok and Claude (although only after they already generated the draft and told it was declined for LLM use). However, if the former's owner ever finds out, he will have that fixed within a day. I₂Overcome ^talk 02:10, 21 April 2026 (UTC)

This is extremely helpful, thanks for your effort. You didn't happen to save the draft articles, did you? Because that would be a handy reference for reviewers. (Or, on second thought, a cheat sheet for LLMs...never mind). WeirdNAnnoyed (talk) 23:10, 10 April 2026 (UTC)

Not sure about that second thought – LLMs will be helped if we point out the AISIGNS (although the probability that they find that specific page and rely on it is pretty low, which is why our existing WP:AISIGNS is an acceptable tradeoff), but just example outputs should be fine. Chaotic Enby (talk · contribs) 23:14, 10 April 2026 (UTC)

Having example outputs somewhere is a good idea. I might work on that myself at some point. I₂Overcome ^talk 01:49, 21 April 2026 (UTC)

Fermiboson, it's excellent that you did this! The hard part here is that this is a moving analysis, in that the evaluation now will not be the same even just six months from now. These chatbots are going to rapidly increase their capabilities. By this time next year, it's going to become very difficult to distinguish LLM generated content from human generated content. Is that a problem? In the sense of creating an encyclopedia, no. In the sense of having a community of editors who support this project, I think absolutely. When it gets to the point that we can no longer distinguish LLM generated content from human content, I will very likely depart the project. I won't be needed here anymore, nor will most humans, I expect. --Hammersoft (talk) 01:14, 11 April 2026 (UTC)

If that comes to pass, not only will the purpose of Wikipedia become obsolete (you don't need an online encyclopedia anyone can edit if the whole thing can be written by AI), but the entire Internet will likely become a lost cause. I₂Overcome ^talk 01:46, 21 April 2026 (UTC)

specifically re: market share - this is probably the least sloppy data we're going to get (in that it's a tech journalist, at least, and not an SEO site): ChatGPT still leads by a lot, but Gemini and Grok are gaining ground; Claude is still rock bottom in terms of usage, at least among the general population. Gnomingstuff (talk) 04:08, 11 April 2026 (UTC)

I have, in the past, tried to get llms to generate a new Wikipedia article, although I did not test as extensively. When I tried llms generally managed to get all the wikitext right outside of obscure technical table stuff, and didn't chat back about rules. However, this was before NEWLLM, as Hammersoft says it is a moving target. CMD (talk) 05:04, 11 April 2026 (UTC)

The behaviour exhibited by GPT explains a lot about why a lot of these editors behave the way they do. Athanelar (talk) 18:18, 26 April 2026 (UTC)

 You are invited to join the discussion at Wikipedia:Village pump (idea lab) § Presumptive deletion for AI-blocked users?, which is within the scope of this WikiProject. Chaotic Enby (talk · contribs) 06:54, 11 April 2026 (UTC)

Pls see Wikipedia talk:Editing policy#AI information. Moxy🍁 22:40, 11 April 2026 (UTC)

Occasionally I run across an article containing an infobox with multiple invalid parameters. At WP:AISIGNS it says "LLMs often hallucinate non-existent ... template parameters"; I am wondering if multiple invalid parameters (in the absence of other signs) is considered strong enough evidence to tag such instances with {{AI-generated}}? As an example for concreteness, I've seen {{Infobox U.S. legislation}} used with three invalid parameters ("effective_date", "citations", and "enactedby") in a recently created article since version 1. To me it seems like strong evidence, but I wanted to inquire here first to see if others consider it a gray area. Thanks, Btyner (talk) 13:05, 12 April 2026 (UTC)

Probably not strong enough evidence on its own just given the amount of truly weird stuff that goes on with infoboxes and the amount params have changed over time. Gnomingstuff (talk) 01:39, 13 April 2026 (UTC)

I accused another editor of using AI in a discussion on another Wikimedia wiki. I thought I should ask for a second opinion here just in case I was wrong. Here is the comment that I think was written by AI: d:Wikidata:Property_proposal/Enpedia_ID#c-ZI_Jony-20260412005400-Warudo-20260411141300. Warudo (talk) 15:19, 12 April 2026 (UTC)

It certainly looks like it, but this is not actionable here on en.wiki. CMD (talk) 15:29, 12 April 2026 (UTC)

I did not ask for any action on the English Wikipedia, just a second opinion. However, since you do agree that it's AI and this situation involves two editors ignoring consensus, I'll probably take it to d:WD:AN at some point. Warudo (talk) 15:32, 12 April 2026 (UTC)

@Trade @特急いよのたみ You guys should see this too. Warudo (talk) 15:46, 12 April 2026 (UTC)

Rules of three, negative parallelisms, curly apostrophes, conclusion section, "let's keep the focus on the actual issue"... Chaotic Enby (talk · contribs) 15:43, 12 April 2026 (UTC)

It was discussed at Wikipedia talk:Writing articles with large language models#This whole fiasco... potentially starting a new WikiEssay regarding human-written Wikipedia. I've started a first draft at User:guninvalid/Wikipedia is written by humans, for humans, and I'd like to hear your thoughts or criticisms of such an article. Thanks. guninvalid (talk) 20:14, 25 April 2026 (UTC)

Courtesy pings: @Athanelar @Blue-Sonnet @Chess enjoyer @LWG @NicheSports @GreenLipstickLesbian @Kowal2701 @Gnomingstuff guninvalid (talk) 20:18, 25 April 2026 (UTC)

I like the two sections. I was more thinking of an argumentative essay, some thoughts:

• the encyclopedia is built by collaboration between editors, ie. it is fundamentally social and built on human-to-human interaction
• people volunteer their time to do labour intensive work for no tangible reward, there's solidarity between editors because of this, and introducing LLMs threatens this aspect of our ethos because they crap out content with little effort (parallels can be made to community attitudes around paid editing etc.). Damage to a community's ethos damages the community's cohesion and risks its disintegration
• articles are written by humans for human consumption. Humans know how their work will be received by readers, the content is actively curated for consumption, often as intuitively as possible to aid understanding/education, often to entertain/write engagingly, ie. it's the direct communication of knowledge from writer to reader

Kowal2701 (talk, contribs) 21:31, 25 April 2026 (UTC)

With some regulated exceptions, such as bots which are subject to an intensive screening procedure, every word of a Wikipedia article is written by a human, including these words you're reading now.

This is not true, and we shouldn't lie to readers by saying something we know is wrong. Gnomingstuff (talk) 00:51, 26 April 2026 (UTC)

As @LWG has discussed previously, this essay is more aspirational rather than descriptive. The point is not that it is a true statement, but rather a goal that we aim to keep. So perhaps it would be better worded as "every word of a Wikipedia article should be written by a human". guninvalid (talk) 00:58, 26 April 2026 (UTC)

If it isn't a true statement then don't say "is." Gnomingstuff (talk) 01:27, 27 April 2026 (UTC)

Shit now it isn't even a goal we aim to keep considering that apparently all it takes is one person to say "well I think it's OK" to enshrine the AI into the world forever Gnomingstuff (talk) 04:40, 29 April 2026 (UTC)

I believe the project will !vote in favor of wording like this essay, and will do so soon. The question is, can the WMF convince major model providers to respect our policies and train their models to reject "create wikipedia-style X" inference requests... while there are challenges ahead, wikipedia has as good of a chance as any similar project to remain human. Way better chance than open source code for example. Btw always appreciate the work you do NicheSports (talk) 05:00, 29 April 2026 (UTC)

This isn't about the WMF, it's about the comments like everytime I see an AI-cleanup notification I assume it is incorrect. These efforts are coming across as AI-witch hunts that one gets from uninvolved editors in response to even the most minor of cleanup efforts. Gnomingstuff (talk) 05:09, 29 April 2026 (UTC)

I think the closest comparison in a case like this would be bot-assisted editing (and by bots, I mean PWB and its ilk, not LLMS). There have been people who complained about the bot restriction, although not as many as the AI boosters, but both LLMs and unapproved bots are a way of avoiding accountability. The problem is that humans have to look through hundreds of diffs before they spot your bot messing up an article, or verify every sentence until they find a hallucination, which puts the burden of proof on others. Even one hallucination should be taken as a sign that a user has not checked their work, and the whole text should be considered suspect.

This problem does not change when LLMs grow more accurate over time. In practice, LLMs still have the same unencyclopedic and dramatizing tone they had a year ago, and routinely cite material to sources that don't even mention the claims, much less support them. The only major improvement I have seen in LLM-generated articles (looking at AI AfC drafts) is that they don't add as many nonexistant references as they did earlier.Somepinkdude (talk) 01:48, 26 April 2026 (UTC)

You might want to take into account my own previous writing at User:Athanelar/Don't use LLMs to talk for you which was a proposed guideline but in many ways is itself an essay against overreliance on LLMs. Athanelar (talk) 16:30, 26 April 2026 (UTC)

It's a bit long for this purpose. For an essay like this, briefer tends to be better. The purpose of an essay like this is to make a brief argument, not necessarily to develop one beyond what is needed. More detail can always be found in talk pages and consensus discussions. guninvalid (talk) 23:37, 26 April 2026 (UTC)

I've gone ahead and made the WP:BOLD move to Wikipedia:Wikipedia is written by humans, for humans. guninvalid (talk) 08:06, 28 April 2026 (UTC)

What was the point in asking for feedback only to ignore it all? The essay I had in mind would lay out the arguments in favour of an ideological ban on AI to save editors time in related discussions, but I guess I can do a different title Kowal2701 (talk, contribs) 08:19, 28 April 2026 (UTC)

I'm sorry if it comes off that way. Now that it's in Wikipedia space, it continues to be a living document that can and should be edited by any interested editor, including yourself. If you would like to add a section making that argument, you may. The only reason I didn't add one yet is because I felt it had duplicate scope with WP:NOLLM. guninvalid (talk) 08:34, 28 April 2026 (UTC)

Seconding guninvalid in welcoming further input/expansion/transformation of this. I'd especially like to see an articulation of the line of thought about the inherently social and collaborative nature of our project and how LLMs affect editor-editor relationships. There's also the possibility of moving some of the verbiage from this essay over to WP:NOLLM if WP:BYHUMANS evolves into a more persuasive direction. -- LWG ^{talk (VOPOV)} 15:07, 28 April 2026 (UTC)

As long as the off-wiki world continues to have widespread LLM use, we are still going to get new users who hop on an immediately start making LLM edits. As our AI policy gains awareness, recent changes/new page patrollers are going to start doing a lot of drive-by templating of these users, who absent further intervention/policy explanation will probably make dozens or hundreds more LLM edits before ending up at ANI and most likely blocked. Is there a way for those of us who do a lot of LLM cleanup to see when this happens, so we can intervene as soon as possible to prevent further damage to the wiki and salvage new editors before they dig themselves into a indef-bound hole? Since warning templates are substituted rather than transcluded I can't just use the "what links here" tool. -- LWG ^{talk (VOPOV)} 19:53, 27 April 2026 (UTC)

The standard templates add the use talk page to Category:User talk pages with large language model notices, which currently has ca. 5200 pages. --Gurkubondinn (talk) 20:21, 27 April 2026 (UTC)

Oh perfect, I didn't know about that category. Thanks! -- LWG ^{talk (VOPOV)} 20:48, 27 April 2026 (UTC)

There's also this search. Gnomingstuff (talk) 22:52, 27 April 2026 (UTC)

wow I didn't know about either of these thanks! Dr vulpes (Talk) 00:39, 28 April 2026 (UTC)

On a related note [1], is it possible to have an automated log of editors that have been warned for AI use (both using {{uw-ai}} and manually at AINB) with an edit counter for when their most recent edit was for people to monitor? Accounts could then be manually removed from it when they’re deemed not to be using LLMs anymore or are blocked Kowal2701 (talk, contribs) 00:56, 28 April 2026 (UTC)

I bet I could vibe-code up something to that effect with Claude ;). My ideal solution would automatically populate whenever a user is warned via a template, and would have space for a responding editor to assess likely date of first AI use, outcome of interaction with the editor, and date range of edits still needing to be checked. -- LWG ^{talk (VOPOV)} 14:46, 28 April 2026 (UTC)

That'd be really great, Dr vulpes said something similar re automating the creation of clean-up subpages Kowal2701 (talk, contribs) 18:14, 28 April 2026 (UTC)

just be careful because people will yell at you if you deign to add the dreaded, horrible, awful death sentence that is a cleanup tag Gnomingstuff (talk) 04:16, 29 April 2026 (UTC)

 You are invited to join the discussion at Wikipedia:Village pump (policy) § RfC on presumptive removal of AI-generated content, which is within the scope of this WikiProject. Chess enjoyer (talk) 21:27, 28 April 2026 (UTC) Chess enjoyer (talk) 21:27, 28 April 2026 (UTC)

See the RM at Wikipedia talk:WikiProject AI Cleanup/Noticeboard. Yours, &c. RGloucester — ☎ 03:53, 29 April 2026 (UTC)

I'm sorry, but this discussion is not actually over and both RfCs above are premature, or at least WP:LOCALCON, as long as attitudes like this exist and are widespread enough to completely grind any kind of cleanup effort to a halt.

For context, this comment is in response to my tagging an article by a user someone else had previously flagged for AI, and who has since said that they are using AI at least for talk page comments around that exact same time, although they have not addressed that article in particular. So, in other words, what should be a very uncontroversial thing to tag. Except it isn't, apparently.

How can we convince people that it is valid to do AI cleanup work? If that isn't even consensus, then how are we even talking about presumptively removing content, or formalizing AI cleanup under any kind of broader project banner? Gnomingstuff (talk) 14:24, 29 April 2026 (UTC)

This issue isn't that cleaning up AI is not seen as a valid, the issue is that different people have different opinions on what cleanup is required. For some people, an article is not "clean" until every trace of an AI has been completely removed (ideally completely rewritten from scratch) while for others it is unproblematic once there are no verifiability or similar issues, even if some words added by an LLM remain (and of course there are those with views in between the extremes). This is a fundamental difference in philosophy where neither view (nor any others on the spectrum) is objectively more or less correct.

The difference causes friction when those whose views are towards the less lenient end of the spectrum insist that an article needs work when those with differing views believe that all necessary cleanup has been done.

If you disagree with someone adding or removing a tag remember to assume good faith and do not edit war, instead go to the talk page and explain why you think the tag is (not) needed - do not assume it is "obvious". If someone states they take responsibility for the text in its present state then that is something that policy allows them to do and it should be treated as human-written from that point, regardless of whether it was or wasn't. Thryduulf (talk) 15:43, 29 April 2026 (UTC)

it should be treated as human-written from that point, regardless of whether it was or wasn't.

Why? This is an actual question. If we know something isn't human-written, why are we pretending that it is, and effectively lying to readers (and soliciting their money) about it? Gnomingstuff (talk) 16:23, 29 April 2026 (UTC)

Because why would you not? If I have read a sentence of content, that may have been emitted by a LLM, together with its corresponding inline citation, and I'm satisfied that it is neutrally written, WP:DUE, not a copyright violation, the source backs up the claim, etc. etc. what else is there to be done? Cheers, SunloungerFrog (talk) 16:57, 29 April 2026 (UTC)

I don't know, not ask people to give Wikipedia their money, that they could be using to pay their bills or put food on the table or any number of other things, by lying to them that the encyclopedia is written by humans? Gnomingstuff (talk) 16:57, 29 April 2026 (UTC)

We seem to be moving from a discussion about content towards a discussion about how the WMF raises funds.

On the content front, what would you have someone do to that sentence of content? Delete it? That would seem to be unnecessarily harmful to the encyclopedia, as it deprives readers of helpful content. Reword the sentence? That seems like just pointless busy work. Cheers, SunloungerFrog (talk) 17:46, 29 April 2026 (UTC)

At minimum, disclose it to readers so they know. This is the minimum standard in pretty much any other field. If AI is used to edit a scientific paper, then even if that paper does go to print, it's expected that there be a disclaimer, and the paper might be retracted if there isn't one. If AI is used in journalism, or to write a novel, people generally aren't very pleased to find out that it wasn't disclosed. It's also consistent with precedent here: we mention the provenance of text that originated on the Catholic Encyclopedia, or for articles translated from other wikis.

I realize that this standard of disclosure is much more difficult here than in other fields because there are so many editors and the size of edits varies a lot. But if a whole article is AI-generated, then I don't think it's unreasonable to disclose that as long as the AI-generated text remains. Gnomingstuff (talk) 18:08, 29 April 2026 (UTC)

Well, I was quite careful to talk about a sentence of content, not a whole article, and I'd like to know what must be done to that sentence so that you are satisfied that the {{AI-generated}} maintenance template may be removed? Cheers, SunloungerFrog (talk) 19:23, 29 April 2026 (UTC)

It's impossible to say without a specific example. In practice I let individual sentences slide all the time. Also, the quote above was in regard to a whole article. Gnomingstuff (talk) 19:38, 29 April 2026 (UTC)

I agree with this argument for disclosure. Wikipedia shouldn't be more lackadaisical than the minimum standards set elsewhere. We shouldn't ignore our own practices about importing from the Catholic Encyclopedia or EB11 or wherever. And, indeed, asking people for money on false pretenses is... poor form. Sure, some of that is on the WMF for how they choose to fundraise, but once the promise is made, we have to keep up our end of the bargain. Even apart from the details of WMF fundraising banners, news coverage of developments like WP:NEWLLM has contributed to people regarding Wikipedia as a non-AI/anti-AI/AI-free environment. We owe it to our readers to be honest about that.

I don't think that rewording sentences is "pointless busy work". At a bare minimum, apart from the concerns above, it gives the editor doing it that much more experience with writing. And I've yet to meet a sentence on Wikipedia, even one written by a human, that was inarguably perfect in every way. Stepwise Continuous Dysfunction (talk) 19:26, 29 April 2026 (UTC)

If one wanted to get more experience at writing, I submit that there are many better ways to achieve that that are more beneficial to the encyclopedia than re-writing a sentence that is otherwise satisfactory save that it may have been emitted from a LLM. Cheers, SunloungerFrog (talk) 19:37, 29 April 2026 (UTC)

There could be better ways, but the point wasn't to say that it would be the best thing to do with one's time, just that it would be more than "pointless busy work". Sometimes, one just wants to end the day with a feeling that one did more than nothing, even if that just means a little wiki-gnoming.

Anyway, as noted above, the example that prompted the thread was about a whole article, not a sentence, and without a specific sentence to look at, there's no way to say what would have to be done with it. Stepwise Continuous Dysfunction (talk) 20:57, 29 April 2026 (UTC)

Gnomingstuff --Just wondering, if I made a template like {{source attribution}} for AI text -- like This article contains material generated by a large language model would you make much use of that? I've been debating doing that for a while, but the consensus of the first WP:NEWLLM meant I really couldn't, but for cases where there's LLM generated text, but you can't get consensus to remove it... I do think disclosures are good, and we do need to label PD text. That's not optional under current PAGs. It's not as obvious or badge of shamey, because it doesn't imply that the content is wrong. IDK, you'd still get some pushback (some people assume that the public domain tag is an insult to them, goodness knows why), but because it doesn't contain the language in {{AI-generated}}, the conditions for removal are a lot clearer: it's when the LLM text is gone. And again, it interacts weirdly with NEWLLM/NOLLM, but I think this is a place where the community consensus and the actual PAGs aren't aligned, not fully (especially given the number of people in the banning LLM RFCs who say that if the text is non-problematic, then even if it is LLM generated, who cares?) and I'd rather go with a practical solution. Opinions? GreenLipstickLesbian💌🧸 22:32, 29 April 2026 (UTC)

Personally I would make use of that. But since AI-generated text is in the public domain, does that imply that we should be using {{source attribution}} now? --Gurkubondinn 22:36, 29 April 2026 (UTC)

I think, given that source attribution is attaching the text to a specific source, while an LLM is just... algorithm generated letter combinations, it might be best to keep them seperate. We also do have a lot of specific PD templates if we import a lot of content from a source, especially if that source has... quirks-- {{EB1911}} for example, is used on a lot of articles, and it's useful because text that's PD due to age (see the 1911 publication date) can be a bit racist, sexist, or otherwise just outright outdated. AI generated text has a different set of characteristics. GreenLipstickLesbian💌🧸 22:42, 29 April 2026 (UTC)

I'd be OK with using it -- not sure it will really move the needle all that much, but I've never really been a fan of the language in the template, and most of it dates back to when there weren't any AI guidelines. I can see there being pushback though potentially that the template doesn't say what to do about the text. Gnomingstuff (talk) 22:44, 29 April 2026 (UTC)

True, but I see that as the secret weapon of the tag, the way it doesn't really imply anything else. Obviously people will misread pretty much any tag if they want to badly enough (I tagged an article for overquoting once, not copyright issues, I just hadn't realized that there was a power user watching it, and they got very offended by the "fact" that I was accusing them of copyright infringement, despite the fact that I was literally using a template which called for the quotes to be transferred to Wikisource/Wikiquote... ), but when the other person gets annoyed at the presence of the tag, you can be very detached- after all, you're not saying anything about the quality of the text. You're just saying that it was LLM generated. And that's it. Nothing more, nothing less, you can just default back to "is this text LLM generated or not?" and refuse to get drawn into any other argument, if you'd like. Now, people who know what they're doing, they can look through articles which transclude that template and look for the standard LLM generated text issues, the same way I might look through the public domain category and expunge a bunch of hagiographies that were copy-pasted. But I'm not adding a "this is an awful piece of work" tag to an article, and, by and large, that stops most people from getting as offended.

Again, not perfect. (For example, if a person can't even tell if a biography on somebody who has been dead since the Obama administration doesn't realize that such an article doesn't merit a {{BLPsources}}, and templates you when you remove that tag because they're in NPP, then there's not much you can do about that.) GreenLipstickLesbian💌🧸 02:01, 30 April 2026 (UTC)

the way it doesn't really imply anything else – It would imply tacit endorsement of the content since while it indicates presence, it doesn't indicate that the presence is a problem (and [[WP:NOLLM|it is). "We know there is LLM content here and that's OK" will be the read. fifteen thousand two hundred twenty four (talk) 04:10, 30 April 2026 (UTC)

Not sure there's a wording of the tag that will satisfy everybody, there are legitimate pros and cons for both of these wordings and there are also people for whom the con is having a tag at all Gnomingstuff (talk) 14:52, 30 April 2026 (UTC)

(and sounds really similar to the stuff that happens with copyvio tags, I had no idea that was so contentious too and I'm sorry it happens) Gnomingstuff (talk) 15:19, 30 April 2026 (UTC)

It's surprisingly contentious at times; I have a selection different mantras that I chant quietly under my breathe when the going gets rough, varying from "they're just trying to help, they're just trying to help" to "people are ungrateful assholes, humanity is terrible"... but when I get to the last one, I generally go and write about some weird Alaskan history for a few months until I can chill out lol. GreenLipstickLesbian💌🧸 02:11, 1 May 2026 (UTC)

WP:NOLLM is still pretty new and community awareness is growing slowly. In this particular case, the original editor is still active and is cooperative, so maybe it would be more helpful to just give them some time to clarify what text in the article is LLM generated so we don't have to guess? I realize that the volume of work you do on this issue means you hit a lot of frustrating conversations, but sometimes it might be better to give people an offramp from the dispute, especially when the people raising the heat level aren't even the most important parties. There's plenty of LLM cleanup to do that doesn't face motivation-sapping resistance, and the LLM concerns are still in the talk page and the history tag or no tag. The WMF issue is a whole another conversation that I don't have time for right now. -- LWG ^{talk (VOPOV)} 17:04, 29 April 2026 (UTC)

This whole thing kicked off within about 10 minutes of when I originally tagged the article, after I had already left a message on their talk page.

As far as "plenty of LLM cleanup to do" I am literally being told, right now, at this very instant, that [the person leaving the comment] and many others have disagreed with your approach.. So clearly people don't want it to be done. Gnomingstuff (talk) 17:21, 29 April 2026 (UTC)

[the person leaving the comment] and many others have disagreed with your approach. I think this person may just be unaware of the shift in consensus that led to WP:NOLLM. While unreviewed LLM is an obvious problem, reviewed and repaired LLM is not reflected the position of many people (including me) a year ago, but many of those people (including me) have changed our view under the crushing weight of the evidence presented by people like you, and this shift in consensus is shown by the clear, overwhelming support for the WP:NOLLM RFC. So I think you should not let a few opposing views shape your understanding of the whole community's perspective. -- LWG ^{talk (VOPOV)} 18:21, 29 April 2026 (UTC)

LWG makes a reasonable point here, I think. Stepwise Continuous Dysfunction (talk) 19:13, 29 April 2026 (UTC)

Once again looking to CCI as an example, individual cleanup work often gets delayed despite the general consensus. That doesn't mean there isn't a general consensus against not creating copyright issues, or that other discussions about handling copyright shouldn't happen. CMD (talk) 20:15, 29 April 2026 (UTC)

I mean, correct me if I'm wrong, but no one says things like "every time I see something flagged for copyright violation, I assume it is incorrect," or "what would you have someone do to that sentence of copyright violation? Delete it?" or "copyright violations should be treated as not copyright violations, regardless of whether they are or aren't." Or at least I would hope that if people said that, they would be quickly disabused of that notion. Sure, the people making the copyvios might object, but otherwise CCI does not have to routinely deal with people questioning its existence -- literally no one opposed the noticeboard being set up -- or telling them that they aren't doing real work, etc. Gnomingstuff (talk) 20:42, 29 April 2026 (UTC)

Yeah, the consensus, while overwhelming, is not unanimous, and the few dissenters are loud. -- LWG ^{talk (VOPOV)} 20:55, 29 April 2026 (UTC)

Contested CCI work has led to arbcom cases and dysysops. There are definitely those who feel they aren't doing real work. CMD (talk) 20:56, 29 April 2026 (UTC)

Thanks for the clarification. I'm sorry they have to deal with that. Gnomingstuff (talk) 21:41, 29 April 2026 (UTC)

Yeah, AFAIK WP takes a stricter stance on copyright than the actual US law to avoid any litigation, and a surprising number of regs don’t respect CCI as a result (re close paraphrasing, see Wikipedia:Arbitration/Requests/Case/Pbsouthwood/Preliminary statements) Kowal2701 (talk, contribs) 23:50, 29 April 2026 (UTC)

AI is a politically fraught topic, and some users have unexpectedly strong reactions when it's brought up. Don't let a few crabby editors make you question everything. You shouldn't take it personally. (Neither should the other editors, for that matter.) Apocheir (talk) 00:56, 30 April 2026 (UTC)

Hmm, funny. I have the exact same opinion as Johnjbarton. There's been multiple cases now of someone claiming some text is AI written solely because the reference doesn't support the content. I check the reference, it absolutely supports the content. I then stare (virtually) at the editor, wondering if they're lying to my face or if they're just incompetent at reading sources. Could go either way, honestly. Anyways... Silverseren^C 04:41, 30 April 2026 (UTC)

first of all, [citation needed]; second of all, you are welcome to go to the RfCs above and express this opinion if you really feel that your unspecified "multiple cases" (many people are saying this!) are proof that all AI cleanup work is invalid Gnomingstuff (talk) 05:20, 30 April 2026 (UTC)

April showers bring May flowers, a new category adding to our ever-piling plate and a question I've had for a bit. Is there a strong enough sense of urgency in this cleanup?

The problem is: The longer this project goes on, the more articles get tampered with by AI. They'll produce entire articles within seconds and we'll have to manually scan every nook of the text for maybe hours to even start putting up the AI-generated tag. We haven't even touched up on the whole point of the project. "AI cleanup". Depending on the size of the article, you could be stuck re-editing and fact checking for days before the results are good enough. And with WP:NEWLLM getting even more advanced every day, that'll complicate things further. Doesn't delaying it make things harder on us?

Right now, there are 6000+ articles that carry that maintenance tag. That number is exponentially increasing as you're reading this (a dozen more added only the first day into May). The articles aren't minor subjects either. Kung fu (term), Jasmine Crockett, Self-esteem, all of which being very notable articles. No one expects this project done in a short time frame. We're volunteers, doing our best at our own pace and interests. But then there's our largest user base to consider. The readers.

The age old debate of whether Wikipedia is reliable or not is still somewhat subjective to the majority of people. But when you stumble across an article with that maintenance tag, it really makes you question then, "Gosh an article with AI? Can I even trust this page?" The answer is usually no.

That turns people off, widens the distrust and overall hurts the credibility of this site. This is not something profound or new. I'm sure many of us here is involved in or viewed discussions on all of these worries. But the longer we leave out the message that we sometimes can't even make sure our texts are reliable, that leaves an impression on people.

What defines a cleanup itself is something we haven't even settled yet, right? And I'm not saying we need a perfect consensus before acting, I've done my considerable best at restoring some of these "plagued" pages. But again, as long as we take time to decide on that, the number of articles will still add up. Thoughts?

PeepeeDino (talk) 10:32, 1 May 2026 (UTC)

Short answer is no, not even close, we have like 5 editors who do the vast majority of tagging and cleanup. Hopefully this is similar to when Wikipedia was being overwhelmed by vandalism where we fall behind at first, until a massive backlog and harm to WP's credibility creates that urgency. Imo we're holding out for government regulation on things like watermarking LLM output, which would make most cleanup after that date trivial Kowal2701 (talk, contribs) 13:36, 1 May 2026 (UTC)

Just a caveat on that 6000 -- a lot of those are from 2023-25 and just went unnoticed until recently. But it's also an undercount, so who knows.

Otherwise, what Kowal said, there are only a handful of people doing this. Gnomingstuff (talk) 14:47, 1 May 2026 (UTC)

Is it worth looking into creating an AI cleanup drive, similar to the drives other WikiProjects have? One one hand, I think it could attract more people to the project; on the other, it might be less straightforward to onboard people since it's less obvious to the average editor what LLM-generated text looks like compared to, say, what copyeditable or unsourced text looks like. There's also a good amount of overhead necessary to administrate a drive. —Em-as-in-emily (talk) 16:29, 1 May 2026 (UTC)

I think that could be a good idea. The drive could have a focus on the open cleanup cases, as there're specific diffs to check, as opposed to an article with a tag, where it's hard to know where to start. InfernoHues (talk) 16:34, 1 May 2026 (UTC)

A drive is tricky. Before considering something like that, there needs to be a clear and easily communicable workflow in place, including not just cleanup but also the admin side of how to look for things to cleanup and how/when/why to mark something as cleaned. Looking again to CCI, a long backlog can be expected, but that's also probably not the end of the world. CMD (talk) 16:36, 1 May 2026 (UTC)

A drive is also tricky when there is a substantial contingent of people who do not believe AI cleanup work should be done at all. Gnomingstuff (talk) 18:47, 1 May 2026 (UTC)

That 6K is definitively an undercount. We need a quicker way of dealing with this stuff. If LLM-use is expected at all, the article should be restored to the last clean version or just deleted entirely if there is no clean version. ~WikiOriginal-9~ (talk) 16:48, 1 May 2026 (UTC)

there's an RfC on this topic Gnomingstuff (talk) 18:06, 1 May 2026 (UTC)

Would it be helpful to add an optional parameter to {{AI-generated}} for the revision where the AI-generated material was added? Sort of like how {{Copyvio-revdel}} does it. Apocheir (talk) 03:53, 2 May 2026 (UTC)

Possibly. For the moment you can include it in |reason, perhaps we could mention that in the documentation. CMD (talk) 04:33, 2 May 2026 (UTC)

It could be helpful. The problem is when it's like 20 revisions (a lot of people will do one sentence/paragraph at a time) or multiple people. Gnomingstuff (talk) 07:12, 2 May 2026 (UTC)

None of the warning templates link to WP:NOLLM ({{uw-ai1}}, {{uw-ai2}}, {{uw-ai3}}, {{uw-ai4}}), I'd suggest adding Adding LLM-generated content is prohibited on Wikipedia. to all of them (or just link "prohibited"). Would make the change myself but not a template editor and don't fancy making 4 separate edit requests Kowal2701 (talk, contribs) 17:31, 2 May 2026 (UTC)

On it, thanks for bringing it up! Chaotic Enby (talk · contribs) 18:04, 2 May 2026 (UTC)

 Done all four! Chaotic Enby (talk · contribs) 18:16, 2 May 2026 (UTC)

Thanks, that looks great Kowal2701 (talk, contribs) 18:18, 2 May 2026 (UTC)

Tried making an edit request to a similar effect for the level 1 a month ago that incorporated WP:AITALK in addition to WP:NOLLM. It was declined since an RFC had some comments about possibly updating the templates, but the RFC has since closed with no impactful discussion about them. Thoughts on incorporating AITALK? fifteen thousand two hundred twenty four (talk) 06:26, 3 May 2026 (UTC)

Tbh I was considering suggesting we add something encouraging people to edit a wiki in their native language, as poor English fluency is often a reason for LLM usage, but idk, level 1 is already fairly long Kowal2701 (talk, contribs) 12:54, 3 May 2026 (UTC)

Could create one like {{uw-ainw}} that says Some editors use AI because they are not fluent in English; if that is the case here, you may find it easier to contribute at a Wikipedia in your native language. and could be used in conjunction with the other warnings Kowal2701 (talk, contribs) 13:00, 3 May 2026 (UTC)

I'm less concerned with a little extra length if the information is important, and it seems important that editors warned for LLM use should be made aware of both relevant guidelines and their respective domains. fifteen thousand two hundred twenty four (talk) 14:04, 3 May 2026 (UTC)