This page is within the scope of WikiProject AI Cleanup, a collaborative effort to clean up artificial intelligence-generated content on Wikipedia. If you would like to participate, please visit the project page, where you can join the discussion and see a list of open tasks.AI CleanupWikipedia:WikiProject AI CleanupTemplate:WikiProject AI CleanupAI Cleanup
To help centralize discussions and keep related topics together, all non-archive subpages of this talk page redirect here.
If there's any LLM-generated content involved (which there well may be since LLM-generated content tends to be very promotional), this would be in scope for this project. Otherwise, this might be an issue to take to WP:NPOVN or inform WikiProject Artificial Intelligence (which focuses on coverage of AI in articles).
I can help with this, although it will take me a few days on my own. I was waiting for some type of admin approval for a mass-ish revert at the ANI thread, which I think we got today [1].
My approach will be to completely revert anything that shows a sign of LLM-introduced issues - hallucination, puffery, etc. From what I have seen from the user's lead rewrites this will mean 90%+ are reverted. Some, such as the Haboob article (which is where I discovered these edits) have been rewritten already. If a manual reversion is necessary I will start my edit summaries off with "LLM CLEANUP" so anyone who wants to collaborate on the cleanup can easily see that a particular article has already been handled. NicheSports (talk) 18:03, 9 September 2025 (UTC)[reply]
No, go right ahead. I was mainly hitting the ones that were the most recent revisions that I could easily rollback. There were still some that had been edited after the LLM versions, so it would take more work to untangle the revisions. Trivialist (talk) 10:52, 10 September 2025 (UTC)[reply]
Update. Of the ~480 leads this user rewrote with LLMs, ~320 have been reverted, including ~55 that I handled manually. Almost all of the remaining ~160 leads will require manual reversion. I will keep working on these but wouldn't mind some help. Feel free to message me here or on my talk page and we can discuss how to split them up. NicheSports (talk) 23:16, 15 September 2025 (UTC)[reply]
Thanks! A few other users (mostly Trivialist) went through their edit history and reverted everything that could be done automatically. I have been following behind, working from Jaravedr's most recent edits backwards, to cover anything that requires manual reversion. I just handled Government_of_India_Act_1935, so everything more recent than this edit is handled.
I believe it's necessary to review all of their edits from 2025, as far as I can tell, nearly every one of them, including the non-lead expansions, used AI. ARandomName123 (talk)Ping me!03:17, 19 September 2025 (UTC)[reply]
I went backwards including all edits up to and excluding Lend-Lease Sherman tanks. I reverted most due to citation issues and because a lot of it was just saying the same thing but in a more wordy manner with some unsourced examples sprinkled in. ARandomName123 (talk)Ping me!04:47, 19 September 2025 (UTC)[reply]
Their 50 most recent edits are now handled, with help from a few others. I'll keep working on the rest of the leads/expansions tomorrow. I could use some help with their edits before they kicked off their mass rewriting efforts on August 28 though. Those earlier edits might be harder to resolve unfortunately NicheSports (talk) 22:41, 20 September 2025 (UTC)[reply]
How should we be proceeding regarding people with hundreds/thousands of edits made with AI?
After about a month of trawling for possibly-AI edits, something I've observed is that unless someone's a student who edits one article and then leaves, the people who are making AI edits are often making them to the tune of dozens, hundreds, or even thousands of edits. Alarmingly this includes articles that have made it through AfC.
I haven't been making threads for every single such person, or posting on ANI, etc., because there are just so fucking many of them, and not all of them have been active recently (though a lot have). There also continues to be no AI policy, and no help seems to be arriving on that front for the foreseeable future, so it's unclear whether anything is even enforceable. Gnomingstuff (talk) 23:02, 8 September 2025 (UTC)[reply]
If the only issue is that the edits may have been done with AI and there are no policy/guideline/MOS issues, there isn't anything that can be done. It is difficult, but possible to use LLM and be constructive. But usually if it's a sloppy use of LLM, there will be some other problem with the edits that can be used as justification to revert/delete/take action. If it's a case of large scale editing with LLM with problems, WP:MEATBOT could be an applicable guideline depending on the situation. JumpytooTalk01:04, 9 September 2025 (UTC)[reply]
Constructive edits are less likely to be noticed, so the ones that end up under discussion are likely the ones with issues. That said, aside from developing a few edit filters, there's little way to bulk address issues other than individual action. CMD (talk) 01:52, 9 September 2025 (UTC)[reply]
Yeah exactly -- if I'm finding these at all, it means that there's at least enough slop and/or unchanged formatting remaining that it doesn't take much scrutiny be 90% sure an edit was AI. But from there I don't actually know how much review was actually done. And there are so many of these editors compared to the handful that get written up at ANI that it feels... disproportionate, like "ok, here's our daily fall guy."Gnomingstuff (talk) 02:11, 9 September 2025 (UTC)[reply]
create something better than the massRollback script that allows people to undo the changes of those posting AI slop. A script that can take into account edits that came after and provides a convenient interface that allows people to decide what to restore/keep
Problems are: it is difficult to force the WMF to do something, and especially to do it quickly. Creating a firm consensus for an AI-policy that says "if reasonable people expect AI, then trash it" is gonna be difficult because some of us are reflexive contrarians, some of us don't understand the issue and everyone needs to have their say.
I assume there have been previous attempts to get #1 introduced as policy. What failed? Can we use the recent ANI thread as a trigger for a renewed effort?
I'm newer here but I assume the correct approach would be Village Pump? Perhaps we could get some admins to take a look at a draft of an RfC before we post it there? Sorry if any of these ideas are stupid NicheSports (talk) 18:07, 9 September 2025 (UTC)[reply]
If someone is actively editing I post a comment on their talk page asking if they're using AI, what tools/versions/prompts, and what review they're doing. Usually they're somewhat receptive if you ask nicely without a template. Problem is that not everyone is actively editing since we are clogged up with 3 years' worth of slop by now, and so I usually don't template anyone who hasn't been around in the past month. The other problem is that even people who do answer the questions don't always give a full accounting of the generation/review process, and I'm not sure whether that's a problem with my approach.
Of your suggestions:
2) is the easiest but depends upon 1)
1) should not be hard in a reasonable project but is proving to be excruciatingly difficult, partly because I don't think people realize how bad the problem has gotten
What I've usually been asking is just a three-parter asking people to disclose:
- What if any AI tools were used, and what version
- What features and/or prompts were used
- What review was done on the output
Usually if you ask nicely and don't slap a template down, people will be willing to at least give some answer (and in my experience it's usually that AI was indeed used) Gnomingstuff (talk) 15:12, 9 September 2025 (UTC)[reply]
Flagging a thread on the {{upscaled images}} talk page for wider input, as it's a template that's used by members of this WikiProject.
Should the template documentation tell users to explain your reasons on the page's talk page whenever they add it, or is the template sufficiently self-explanatory?
Another editor objected to me taking that line out of the documentation a couple of days ago (it's only in there because I accidentally copied it over from {{AI-generated}} when creating the template last January), but they aren't really saying why, and have asked for other views. Belbury (talk) 20:25, 15 September 2025 (UTC)[reply]
yet another large swath of AI edits
There are a bunch of edits by Bookleo that are all but certainly AI-generated; given that one of their edits left in the chatbot response.
These are largely edits to award-winning books -- making this case high priority because literary authors.... uhhhhh, do not tend to like AI, and so the shitstorm potential if this escapes containment is greater than norma Gnomingstuff (talk) 21:04, 16 September 2025 (UTC)[reply]
It's worse than that - in the last few weeks they have been mass-rewriting sections of articles, primarily leads, within minutes of each other. The following 6 rewrites were made within 65 minutes: [2], [3], [4], [5], [6], [7]. These are bot-like edits that cannot be adequately human-reviewed and therefore violate multiple core policies. Because they are also using the {{lead too short}} tag to find targets they have actually hit a few articles that we had just cleaned up from the previous mass-lead-rewriter. Examples: [8], [9]
We need to get this user to stop editing immediately, the cleanup here is going to take forever because from their edit history it looks like they have used AI tools inappropriately on hundreds of articles in the last year. I suggest taking this to ANI and asking for an immediate temporary block to prevent further disruption while we discuss with the user. NicheSports (talk) 21:42, 16 September 2025 (UTC)[reply]
Yeah I figured it would be good to have subject matter experts take a look since I haven't read most of the books, and lo and behold someone oh wait that was you found a hallucination in like 20 minutes. Gnomingstuff (talk) 03:08, 17 September 2025 (UTC)[reply]
Is there something particular about the lead too short tag that attracts this? Perhaps there needs to be some background tracking of very rapid removals of that tag. CMD (talk) 03:12, 17 September 2025 (UTC)[reply]
Based on the tags in their edit history I assume it has something to do with newcomer tasks -- the lead too short is one of the possibilities for "expand" tasks -- although I may be wrong. (And also they're not a newcomer so I don't know why they're even getting shown those.) Gnomingstuff (talk) 04:35, 17 September 2025 (UTC)[reply]
I handled all of this user's edits back through August 28, which is when they started, after a few months off, came back to rapidly editing articles. I could use some help addressing their earlier edit history NicheSports (talk) 17:34, 22 September 2025 (UTC)[reply]
Thank you for looking into this. The AI edits goes back quite far, unfortunately - I suspected that Bookleo was making AI-generated edits to the My Brilliant Friend article over a year ago in this change: [10] which had characteristically ChatGPT-style writing and broken citations. Jordan Eldertalk21:08, 2 October 2025 (UTC)[reply]
Yet _another_ enormous swath of AI edits, this one even larger
I am in the process of tagging the edits by Thefallguy2025 as AI generated and I am not even close to being done. Similar as the above: a lot of rewrites from the newcomer edits copyedit/expansion interface, a few smoking guns (ChatGPT parameters left in on some of the ones earlier in the year), extremely AI-esque edit summaries, and enough suspect text elsewhere that I'm pretty sure it's all AI. To the tune of over 1,000 edits. Gnomingstuff (talk) 21:24, 17 September 2025 (UTC)[reply]
Thanks for finding these and please keep it up! We need more LLM policies and being able to demonstrate the magnitude of the problem to the wider community will help move in that direction NicheSports (talk) 03:13, 18 September 2025 (UTC)[reply]
imo there should be an AI version of WP:CCI. Makes it easier to identify who needs to be cleaned up, their edits, and the progress. Hopefully it doesn't become as backlogged as CCI though. ARandomName123 (talk)Ping me!04:58, 19 September 2025 (UTC)[reply]
That would be helpful but I'm worried about the inevitable "you have no proof why are you putting me on a list" backlash, and I also don't want this to turn into scapegoating or insulting editors. Gnomingstuff (talk) 06:35, 19 September 2025 (UTC)[reply]
According to Wikipedia:Contributor copyright investigations/Instructions, a newly filed CCI has to be evaluated by a clerk or administrator, who would subsequently open a case only if one is warranted. We could implement a similar triage system for LLM cleanup cases, although I think a consensus on the LLM noticeboard (this page) would be sufficient to open an LLM cleanup case (without needing an administrator or designated clerk). — Newslingertalk15:51, 19 September 2025 (UTC)[reply]
This makes sense to me. It would help with the cleanup process and make it clearer to the wider community how this type of LLM misuse should be handled, likely leading to cases being caught earlier NicheSports (talk) 13:29, 20 September 2025 (UTC)[reply]
Had a chance to look into this and it is pretty bad too. These four substantive article expansions with obvious LLM-style editorialization were made within 45 minutes: [11], [12], [13], [14]. This article expansion [15] is particularly absurd, and likely involves extensive hallucinations as it is entirely unsourced. This is just a small sample NicheSports (talk) 15:21, 19 September 2025 (UTC)[reply]
yeah, I checked all of their hundreds of edits to make sure at least some of the original text was still in the current version, and I think I only found a handful where either the text had been substantially edited, or where someone had noticed and reverted. Gnomingstuff (talk) 17:54, 19 September 2025 (UTC)[reply]
Ugh.. lots and lots of 'Plot' sections too, which are a nightmare because we rely on the source material for the plots so, short of going to watch the film, independent editors have no way of verifying the content.. I'm assuming that ChatGPT hasn't actually watched the films so must be pulling the summary from somewhere else, or hallucinating it. JeffUK02:09, 24 September 2025 (UTC)[reply]
long time editor, i think i did a quick review of a few of their recent edits, think they started using AI somewhat early, but would like others to weigh in to confirm Bluethricecreamman (talk) 18:39, 20 September 2025 (UTC)[reply]
Mentioned this on the page, but this does not seem to be AI, or at least not fully AI. The "editorializing" thing is specifically about an inappropriate tendency of "Wikipedia-style" AI articles -- talk page comments are supposed to "editorialize," that's literally what they are. Text that is "confusing" is an anti-tell for AI. Uncommonly used acronyms are an anti-tell. References to Wikipedia minutiae are an anti-tell -- when LLMs mention Wikipedia policies they do it in a superficial way, they don't talk about arcane historical Wikipedia redirects from 2012. Even the quotes you mention are wrong: you criticize them using "stands as," but that AI tell is referring to a specific, narrow kind of AI Wikipedia article puffery like "stands as a testament," "stands as an enduring reminder," etc., not just the literal words -- and more importantly, the comment doesn't even include them. Sloppy stuff all around. Gnomingstuff (talk) 06:53, 21 September 2025 (UTC)[reply]
The threshold of evidence needed to file an LLM-related complaint against an established editor that results in any action is higher than the threshold needed for a new editor. (This is an observation, not a normative statement.) It's certainly plausible that the comment Special:Diff/1310258505 was written with LLM assistance, when taking into account how its style of writing differs significantly from the style used in previous comments, e.g. Special:Diff/1197568358. But, it's unlikely that an ANI complaint would result in any action against any established editor on the basis of LLM use unless the complaint contains at at least one high-confidence sign of AI writing, such as a WP:G15 element or Markdown use. — Newslingertalk16:50, 21 September 2025 (UTC)[reply]
Few months ago, I tried to create a page (Wikipedia:WikiProject AI Cleanup/AI or not) in which editors can test their AI detection skills á la Tony1's copyediting exercises. At first I was copy-pasting novel examples that I generated myself, but I think compiling real-life AI-generated and human-generated articles would create a more accurate testing environment. Any help collecting examples (whether be it writing from pre-November 2022 or editors who disclosed their LLM usage) will be welcome. Catalk to me!00:24, 23 September 2025 (UTC)[reply]
AI with substantial review: here (this user has disclosed their use of LLMs, including a few they apparently trained themselves, and their review process)
Guide dog, specifically the Discrimination->Australia bit, might seem like its written by ChatGPT, especially the "What to do if you experience discrimination", because Wikipedia is not meant to be your guide. I think the first section of the Australia section is alright though. N51 DELTATALK08:13, 23 September 2025 (UTC)[reply]
Concur. Section was added by Anaman12321 in Special:Diff/1301116787, Anaman has also created Western Mediterranean which shows signs of being LLM-generated including a ?utm_source=chatgpt.com reference. Their userpage which states they are "thrive on transforming seemingly impossible ideas into tangible concepts, often blending scientific principles with imaginative thinking" is also clearly LLM-derived.
Fortunately there don't appear to be many contributions that need review, here's a list, anyone please feel free to edit my comment to strike one to claim it for review:
I began reviewing and rewriting that entry on Guide dogs, only to see that it was already covered in another section further up, so I've just removed it entirely. Nil🥝02:56, 24 September 2025 (UTC)[reply]
Given that this wiki project is becoming a bit more like a AI noticeboard, somewhat akin to the CCI or COIN noticeboards, any thoughts about adding the standard boilerplate "notify people if you bring them up here, and try to talk through the problem with them first" notice and banner? Not a hard and fast rule, of course, and I'm sure its something experienced editors already do, but anyways. Thoughts? GreenLipstickLesbian💌🦋08:57, 23 September 2025 (UTC)[reply]
ARandomName123 and Newslinger both suggested setting up a CCI for LLM misuse which makes sense and is maybe where your suggested language could go? I don't think this project should only be a noticeboard, I'd like to see a dedicated page for that and keep this open for broader topics like research to quantify the impact of LLM use (#/% of AfC declines for AI content over time, etc.) and discussing potential policies NicheSports (talk) 15:18, 23 September 2025 (UTC)[reply]
I fully agree with this proposal – it would still be very helpful to be able to discuss broader topics, and they can easily get drowned in the amount of individual reports. Chaotic Enby (talk · contribs) 03:56, 3 October 2025 (UTC)[reply]
Guidance on handling article with mostly minor edits subsequent to LLM-rewrite
I started working on Bookleo's earlier edits which unsurprisingly have been subsequently edited by other users. My plan is to revert to the pre-AI version unless there have been multiple material modifications to the LLM-generated text. But what about the case of Lincoln in the Bardo? Here is a diff [24] of all subsequent changes made to the article post AI-rewrite. With the exception of the Adaptations section these are all minor copy-edits of the AI-generated content itself. My instinct here is to revert the changes to the pre-AI version and then manually update the Adaptations section to incorporate the added content. What do people think? NicheSports (talk) 20:02, 24 September 2025 (UTC)[reply]
I stumbled across an article created by Tbound2 that was nominated for deletion (Adelaide Metro bus route 100), and have discovered a large number of articles created by this user, in implausibly short succession (while obviously this does not inherently signify generative AI editing, all of these were created within 2 hours of each other), and which I believe to be generated by AI large language models due to their gushing prose (and which already have AI tags):
India Rasheed, created at 10:54 on 10 August ("Rasheed is noted for her clean hands, decision-making, and effective left-foot kicking. She is considered capable of impacting games both at ground level and in aerial contests. Coaches have also noted her endurance and adaptability to different positions.")
Keeley Kustermann, created at 11:19 on 10 August ("She was known for her “clean by hand” play, strong decision-making, and versatility.")
Georgia McKee, created at 11:22 on 10 August ("McKee is recognised for her agility, goal sense, and forward craft. She applies strong defensive pressure inside 50 and is regarded as a creative and opportunistic small forward.")
Brooke Smith (footballer), created at 11:26 on 10 August ("Smith is described as a reliable utility with strong marking ability for her size, a penetrating kick, and the versatility to adapt to multiple positions.")
Lily Tarlinton, created at 11:32 on 10 August ("Tarlinton is known for her athleticism, marking ability, and adaptability to both forward and ruck roles. Her height and mobility offer Adelaide flexibility in matchups, enabling them to rotate key position players across the ground as needed.")
Keeley Skepper, created at 11:37 on 10 August ("A midfielder noted for her contested ball-winning ability, composure under pressure, and precise kicking...")
Poppy Scholz, created at 11:47 on 10 August ("A tall utility renowned for her intercept marking, versatility, and athleticism...")
The Original Pancake Kitchen, created at 12:14 on 10 August ("aims to deliver the brand’s signature all-day breakfast, diner-style charm, and thick milkshakes in a modern, family-friendly setting.")
There are a number of additional articles created by this user that likely fall under the same category of speedily-generated articles entirely written by a generative AI LLM. I don't really know what to do here. I don't think they're eligible for speedy deletion under G15, given these articles obviously could have plausibly been created a different way, they're not transparently nonsense and they've all been reviewed. But I believe this editing to be highly problematic and I'm not sure what to do about it. What is supposed to be done when an editor is repeatedly generating articles using large language models? LivelyRatification (talk) 00:43, 30 September 2025 (UTC)[reply]
Cases like this are where I've generally been tagging articles. It looks like I've tagged some of this user's contributions but not all, I assume that here, as in similar long-term LLM users, I stuck to the most obvious cases even though it's highly unlikely they stopped using AI. (Because then I would have to deal with "BUT YOU HAVE NO PROOOOOOOF POINT OUT WHAT IS WRONG WITH THIS TEXT" type comments which I do not want to do.) Gnomingstuff (talk) 02:01, 30 September 2025 (UTC)[reply]
@Gnomingstuff: I suppose my question is, what good is tagging articles if the user in question is continually using generative AI to create articles? I could, and indeed might if I have the time, individually fix every article to remove problematic LLM insertions (as was done on Lauren Young (footballer)), but is it not a problem of disruptive editing if a user is repeatedly using generative AI to create articles? LivelyRatification (talk) 02:11, 30 September 2025 (UTC)[reply]
The thing is we have no policy against LLM use currently, so within our existing system, they're not actually doing anything that isn't allowed. If someone isn't responding to attempts to discuss the situation then ANI threads have been opened, but there feels like there's a mismatch in what we enforce and what our policy actually is(n't). Gnomingstuff (talk) 02:18, 30 September 2025 (UTC)[reply]
Hm. That's frustrating, but thank you for your help. I guess I'll just keep an eye out and try and fix what's there. I've also left a message on this user's talk page in regards to this thread, which has not yet been responded to. LivelyRatification (talk) 02:25, 30 September 2025 (UTC)[reply]
Gnomingstuff is the OG of LLM text identification and tagging (muchas gracias) but I slightly disagree with their perspective on this situation. We don't yet have policies about LLM use in articles but unreviewed use of LLMs is highly likely to lead to violations of WP:V, WP:RS, WP:OR, and WP:NPOV, while high-frequency misuse of LLMs is likely to violate WP:MEATBOT. I would recommend going through some of the user's edits and finding a few examples of the LLM hallucinating a tangible claim or reference - if they have been adding unreviwed LLM-generated content to articles, you will find a bunch. Then we should follow Tbound2's edits and if they continue to make (likely) unreviewed LLM-generated edits without responding to attempts to discuss with them, just notify one of the admins who follows this page, like Newslinger. Thank you for bringing this up! NicheSports (talk) 04:52, 30 September 2025 (UTC)[reply]
Looking at the articles now, Rasheed has 5/11 sources that 404 and when validating the first source there is some information that is hallucinated (the article only mentions her quote saying she is improving her "opposite" foot, but the LLM took that and said "her left foot" in article), but I was unable to tag G15 as there was content written by other editors. I was able to tag G15 for the third article as it both had 404 sources and irrelevant sources, and no other editors added prose to the article.
Thank you so much! If I have the time I'll also try and take a look at some of these articles, as I said I didn't have time to investigate them deeply beyond their prose. I do hope Tbound2 is able to respond to the criticisms left on their talk page, but I'll keep an eye for any future problematic editing. LivelyRatification (talk) 04:58, 30 September 2025 (UTC)[reply]
I've looked through a number of other articles created by Tbound2 and none that I saw seemed to immediately qualify for G15 speedy deletion, as the references seemed to work. I think Mitchell Sariovski is not notable enough to warrant an article so I'm going to AfD that one. I'll also leave a message on Tbound2's talk page again expressing my concerns. --LivelyRatification (talk) 07:32, 30 September 2025 (UTC)[reply]
Whilst I understand everyone's concern, please draftify or userfy, rather than delete these if you feel they aren't suited to publication yet. Whilst this editor's efforts are misguided to some degree, ever since the removal of WP:NAFL and stricter enforcement of WP:GNG, the coverage of female AFLW players has been very poor here (see how few bluelinks are on List of AFL Women's debuts in 2024 and List of AFL Women's debuts in 2025), and this is at least an effort to address the systemic bias that exists here. The-Pope (talk) 06:58, 4 October 2025 (UTC)[reply]
An article which meets G15, the criteria any of the above have been deleted under, is an article which is essentially raw output from an LLM, it will be better to make a new article than to require other editors carefully review and fix the neutrality, formatting, and hallucination issues typical of model output. If an editor wants to fix up raw LLM output, then they can visit any number of chatbots and have it delivered hot and fresh to themselves, without using up the project's time on cleanup as has happened here. fifteen thousand two hundred twenty four (talk) 08:58, 4 October 2025 (UTC)[reply]
it's the AI pumpkin, Charlie Brown
Wk3v78k23tnsa has made many edits recently primarily to Vince Guaraldi/Charlie Brown-related articles. I noticed some AI tells on one of them while searching, and more jumped out when skimming the other articles. And then I found this. This diff adds a lot of supposed quotes, such as this: Trotter does not just arrange, he amplifies. That bossa nova piece could have been background filler, but instead it feels deeply emotional—like it is telling its own story. Which sounds exactly like ChatGPT coming out of someone's mouth. Annoyingly, the interview it is cited to is a 40-minute video, but it does have a transcript; while I'm pretty sure the transcript is auto-generated, I CTRL-F'd multiple words out of that quotation and not one of them shows up in the transcript, and while machine-learning transcripts can mess up they usually don't mess up that much. So I did the same with some other quotes from the article and still have yet to find a hit. Fucking sheesh.
I don't know how far back this issue goes. Before the Charlie Brown stuff they did a lot of plot summary revisions, which I'm not sure are AI -- for instance, this edit is tonally glib, but it sounds more like human ad copy than LLM slop (this is a gut feeling), and there's at least one grammatical error uncharacteristic of AI. (Also, they seem to have radically changed the way they write edit summaries between now and then.) Gnomingstuff (talk) 00:07, 3 October 2025 (UTC)[reply]
Update: There's also this edit with the chatbot response Here is the revised version with all formatting removed while maintaining the academic tone and word limit:, so I guess their rewrites may now be in play too. Gnomingstuff (talk) 00:42, 3 October 2025 (UTC)[reply]
This one is tough because the user has a high level of English fluency (see this exchange on their talk page from 2016). Then you have this, which feels more like a troll than an LLM. But based on the smoking gun you found that included the chatbot response, I think we have to assume that the majority of their plot summary rewrites involved LLMs. I mean on March 2 2024 they rewrote 108 plot summaries, often 1-2 minutes apart, without grammatical errors. There are many examples of this. On Feburary 17 2024 they rewrote these 7 plot summaries in 9 minutes: [26], [27], [28], [29], [30], [31], [32]. And of course they never made any edits like this prior to November 2022. NicheSports (talk) 02:35, 3 October 2025 (UTC)[reply]
Just so everyone knows, the supposed Trotter quote listed above is back in the article. Either it was there before the last reversion, or someone put it back in manually. I don't have time to watch the video to correct it myself, but want to flag it. WeirdNAnnoyed (talk) 11:05, 6 October 2025 (UTC)[reply]
today in large swaths of maybe-AI edits: world politics
This is another big one.
For the past year-ish, Ritwik Deuba has been adding hundreds of instances of text with consistent AI indicators (though nothing unambiguous that I've found), primarily to world politics articles. One other person has pointed this out, but they didn't respond; I left another edit on their talk page.
I'm less sure about this one than some of the others. It looks like they are probably editing the output somewhat -- although not completely enough to be detectable. It also seems like they may be using a newer LLM model and/or different prompts than most do; the older edits seem more problematic than the newer ones. In particular, most of the "reflecting the significance" AI opinion-ese is attributed as other people's opinion rather than just dangling as unattributed editorializing like it usually does, but if it's AI it still may not be an accurate summation of what those people said.
The big challenge here is that almost literally all of these edits involve very contentious geopolitical topics, and some of them have been dragged into edit wars (that don't seem to be about AI). Because of this I haven't touched them with a 1000-foot pole, besides adding the AI generated tag if the indicators were clear enough. The articles are also high-traffic enough that "just reverting" would be difficult to impossible, and might re-ignite the edit warring. So frankly I don't know what to do. Gnomingstuff (talk) 17:41, 3 October 2025 (UTC)[reply]
hmmm. do you have any diffs we can look at? I checked the edits that the other person pointed out and didn't find any hallucinations - all claims are backed by the cited source. LLMs hallucinate at high rates so I think that edit would have been significantly human reviewed. This is just one diff of course so could be missing something NicheSports (talk) 18:17, 3 October 2025 (UTC)[reply]
ok finding problems. I noticed that these three edits [33], [34], [35] were made within 15 minutes (although with significant overlap in content), without any subsequent copy-editing or typo-fixing. This one also has clear copyright violations - to the point it should be reported, which I have never done NicheSports (talk) 18:29, 3 October 2025 (UTC)[reply]
Will take another look tomorrow. I mentioned this on a few of the talk pages but some of the edits are the same paragraph added to multiple articles (which isn't necessarily a problem by itself).
For stuff like this I'm less concerned about blatant hallucinations so much as any NPOV issues/interpretations that might have been introduced by the AI. Gnomingstuff (talk) 05:58, 4 October 2025 (UTC)[reply]
Article on Product-family engineering has heavy signs of AI usage.
This article on Product-family engineering does not seem to be meeting Wikipedia's standards, and reads like AI writing significantly. Almost no citations, and "Here's a list of some of them", "The Nokia case mentioned below" , random bullet points, long weirdly convoluted series of topics, the whole "example" section about Nokia, it all sounds very odd. 81.214.164.185 (talk) 07:26, 7 October 2025 (UTC)[reply]
(edit conflict) Most of the article dates from far before chatgpt existed. It doesn't have many of the hallmarks of AI to me, just slightly awkward marketing language. Sarsenet•he/they•(talk) 07:36, 7 October 2025 (UTC)[reply]
Not AI: the 2021 version reads much the same, and that predates mainstream LLMs.
The verbiage and structure (including the style of bullet point text here, really terse, no fluff) also isn't really characteristic of AI. Gnomingstuff (talk) 12:29, 7 October 2025 (UTC)[reply]
