Wikipedia talk:WikiProject AI Cleanup

With the recent amendment to WP:NEWLLM to prohibit the use of LLMs to generate or rewrite article content, I think its time to update G15 to encapsulate AI-generated articles generally, not simply unreviewed ones.

Some clear cut signs I'm thinking of: presence of turn0search0, attributableIndex, oaicite, or having a majority of citations include utm_source=chatgpt.com.

It might also be worth emphasizing that WP:AfD or getting consensus for WP:LLMPROD are better for less clear-cut cases. Ca ^{talk to me!} 00:44, 29 May 2026 (UTC)

Yep, all the unambiguous cases (like turn0search0 or oaicite) should definitely go there at the very least. I was thinking of also revisiting the Markdown criterion, but it doesn't seem to be a thing with newer models anymore, so I'm not sure how relevant it would be. Chaotic Enby (in solidarity · talk · contribs) 00:49, 29 May 2026 (UTC)

Markdown is something that I still routinely see, mostly as **markdown bolding**. But there's one specific Markdown element that I want to see added as a G15 criteria, when an edit is surrounded in a markdown code block:

```wiki

```

There's usually perfectly valid Wikitext between ```wiki and ```,^[1] and I'm pretty sure that this happens when someone asks a chatbot to generate a page in wikitext markup and the bot delivers it in a formatted code block for the user.

The leading ```wiki isn't always present (it's more obvious and so users are more likely to see and remove it), so the trailing ``` would need to be sufficient on it's own for G15. --gurkubondinn 10:21, 29 May 2026 (UTC)

Huh, didn't expect it to still be so common, but that's definitely a good catch and should be added! Chaotic Enby (in solidarity · talk · contribs) 10:27, 29 May 2026 (UTC)

I don't really know how common it is, because the edit filter 1369 (hist · log) doesn't match for triple backticks (and I think it would be better to have a separate filter for these code block markers). This is one of the reasons that I've been considering requesting permissions for edit filters, but the granting criteria seems hard to meet. Finding these is almost always a surefire way of spotting an AI-wielding user though, yesterday I ended up draftifing c. ~25 AI-generated articles after I spotted an edit with a markdown codeblock. --gurkubondinn 10:37, 29 May 2026 (UTC)

By coincidence I spoke to @Daniel Quinlan about the triple backticks just a couple of days ago and he added it to 1369, so hopefully that should catch these going forward. There are a very few existing cases where it's used in legitimate wikitext (usually to mark up an ASCII code sample) but overwhelmingly it seems to appear in definitely or at least plausibly AI edits.

I think most of the examples I found had only the start or end backticks, not both. Interestingly, some were relatively small edits including the ``` - so presumably some people are using it to generate individual snippets (eg single paragraphs or references) and not just entire pages. Andrew Gray (talk) 12:50, 29 May 2026 (UTC)

The utm_source could have come from someone using ChatGPT to find sources, and not to write content, so I don't think that should be in G15. Adding the rest of the things you mentioned makes a lot of sense. InfernoHues (talk) 01:41, 29 May 2026 (UTC)

Forgot to mention -- G15 really needs a disclaimer about the need to check Wayback Machine to ensure citations are genuinely nonexistent. Ca ^{talk to me!} 05:37, 29 May 2026 (UTC)

That depends on when the article was created/when the citation was added. If someone adds a link today to somewhere that existed five years ago, that is still a nonsensical citation (probably hallucinated), and they did not read the source themselves. Because if you are reading a now-dead source, then you are reading it on Wayback Machine (or on any other internet archive), so you would copy that link. You wouldn't strip out the https://web.archive.org/... part first. --gurkubondinn 10:10, 29 May 2026 (UTC)

Not necessarily, e.g. you could be copying the reference from somewhere else (on or off Wikipedia), you could have viewed it a few days ago when it was working, etc. It's not something reliable enough for speedy deletion. Thryduulf (talk) 10:26, 29 May 2026 (UTC)

That's why I said "five years" in my reply, because I meant in a timeframe that is unreasonable between reading a source and then using it in an edit. Fair point on copying from elsewhere, though. But I think that this would be reliable enough for speedy deletion if the criteria would require multiple (maybe 3+) instances of this? But just on its own, I agree with you that it doesn't work as a speedy deletion criteria (thought it is an AISIGN in and of itself, and it most likely also means that the user has not read the source that they are citing). --gurkubondinn 10:41, 29 May 2026 (UTC)

I never seen AI cite the Wayback Machine before. Ca ^{talk to me!} 11:16, 29 May 2026 (UTC)

It seems to have "learned" about it somewhat recently. But afaik, The Internet Archive are not super happy about being scraped by these companies (and neither should they, their scrapers are ridiculously aggressive and badly implemented). Maybe some of them are using the API to search for archived snapshots now? --gurkubondinn 11:24, 29 May 2026 (UTC)

Never mind, I just sent ChatGPT the prompt search up in Wayback Machine for 2008 version of wikipedia's page on Cats and it dutifully gave this Wayback Machine link. So, yeah, I agree it seems to know how to cite WM.

However, I do suspect its hallucinated (instead of making use of RAG/actually searching in WM) as the WM does not appear in the "sources" list and the timestamp on the url doesn't seem to match the one on the actual snapshot. Ca ^{talk to me!} 11:40, 29 May 2026 (UTC)

Yep, the Wayback Machine will redirect you to the "nearest" snapshot if you give it a non-existing timestamp.

$ https --headers "https://web.archive.org/web/20080705195056/https://en.wikipedia.org/wiki/Cat" | grep -E "^(HTTP|location)"
HTTP/1.1 302 FOUND
location: https://web.archive.org/web/20080625213031/http://en.wikipedia.org/wiki/Cat

$ AI_TIMESTAMP="20080705195056"
$ https --headers "https://web.archive.org/web/${AI_TIMESTAMP}/https://en.wikipedia.org/wiki/Cat" \
    | grep "^location" 
    | awk '{ print $2 }' 
    | awk -F'/' '{ print $5 }'
20080625213031

$ WAYBACK_TIMESTAMP=$(https --headers "https://web.archive.org/web/${AI_TIMESTAMP}/https://en.wikipedia.org/wiki/Cat" 2>/dev/null \
    | grep "^location" \
    | awk '{ print $2 }' \
    | awk -F'/' '{ print $5 }')
    
$ if [[ "${AI_TIMESTAMP}" == "${WAYBACK_TIMESTAMP}" ]]; then echo "Same timestamp"; else echo "Different timestamps"; fi
Different timestamps

I've noticed this with AI-generated references to Wayback Machine (haven't seen AI-generated references to any other internet archives yet) for a while now. --gurkubondinn 12:03, 29 May 2026 (UTC)

that is still a nonsensical citation

Just to clarify my own words here, I didn't mean this as "meets the 'nonsensical references' criteria for G15". I just meant it in the sense "that still doesn't make sense and therefore it is nonsensical". Didn't make the connection to how it was basically the name of the G15 criteria, probably could have phrased this better. --gurkubondinn 21:13, 29 May 2026 (UTC)

Anything listed under WP:OAICITE, WP:AISIGNS § turn0search0 and WP:AISIGNS § attribution and attributableIndex are valid G15 criteria already because they are all "nonsensical references" that no human would ever write down (especially the byte sequence), but it would not hurt to explicitly list them as examples. You sometimes get people incorrectly removing {{db-g15}} over this, even for the very common OAICITE. I've had this happen a handful of times for all of them, but I've requested CSD for hundreds of pages under all of these criterias.

For the UTM parameters, those only "count" when every single reference has utm_source=chatgpt.com (or the second most common example that I see, copilot.com). But these should still be tagged with {{AI-retrieved source}}, though people are very prone to just removing them. If it is just a handful of them it doesn't even really reliably mean that the user is even using AI as a search engine anymore, because these links have spread over the internet by now (in no large part due to how commonly they appear on Wikipedia). You can even find them in Google search results now.

As a sidenote, the m:Cite Unseen userscript is very useful, and it marks references with the UTM tags with a little orange robot symbol. That's also why they shouldn't be removed from citations. --gurkubondinn 10:04, 29 May 2026 (UTC)

Oh my god, how have I not seen Cite Unseen before. That's awesome, thanks for sharing. EatingCarBatteries _{(contribs | talk)} 01:46, 30 May 2026 (UTC)

I'd support adding the more "smoking gun" AI signs. I don't fully agree with Gurkubondinn that these are already covered by "nonsensical references", as in many cases the links work and back up the text they're cited for, which is far from the original intent of that phrase. Regarding old sources or bizarre access dates, while they can be an indicator, I'm not sure that they're enough for G15. At a minimum, we'd have to require checking whether the article is a translations from other Wikipedias. Toadspike [Talk] 20:49, 29 May 2026 (UTC)

Oh, I meant that stuff like OAICITE, grok_card, lenticular brackets + dagger sign, turn0search0, and the 0xee 0xa8 0x81 byte-sequence is already covered by "nonsensical references". I agree with the dates etc (I think we should do something about that but I don't know what). Sorry that wasn't clear, --gurkubondinn 21:03, 29 May 2026 (UTC)

While I agree G15 should be expanded, as it is very narrow, adding utm_source to the criteria isn't a good idea. If someone types in "find me reliable, secondary sources for X topic" and ChatGPT spits out valid references, but they have the tracking tag, that is totally fine. It doesn't mean that the chatbot wrote the article. EatingCarBatteries _{(contribs | talk)} 01:43, 30 May 2026 (UTC)

Just a note that MW:Who Wrote That? is a browser extension I recently discovered that aids in detecting who added a particular bit of text to an article and may assist with AI cleanup efforts. It was recommended for use in cleaning up copyright infringements but can probably also help break down which editor may have added text when doing AI cleanup. It may be worth adding to the cleanup guide as well. ASUKITE 20:33, 11 June 2026 (UTC)

Pretty amazing to see it brought up here! I'm working on an AI cleanup-tailored diffbrowser and this would be a very neat tool to integrate there. Chaotic Enby (in solidarity · talk · contribs) 21:32, 11 June 2026 (UTC)

This is amazing. It roughly doubled the speed at which I can handle LLM with minor editing on top that blocks reversion. My workflow is one window open to Who Wrote That and another editing the article text. No more going crosseyed looking at the diffs. M kuhner (talk) 14:35, 26 June 2026 (UTC)

I wish there was a way to have similar capabilities without needed to download a browser extension. -- LWG ^{talk (VOPOV)} 20:05, 23 July 2026 (UTC)

These edits made in 2020 have all the hallmarks of LLM, they include citations that obviously cannot verify the claim (2008 citation for 2017 claim), hallucinated references with DOIs not matching the reference, LLM style comments 'Immediately contact/seek a veterinarian if any signs are present.', there are no spelling/punctuation issues but many sentences that are awkward and unnatural etc..

The edits themselves are problematic regardless of if an LLM was used to create them or not but I cannot see these edits as being written by a human. Traumnovelle (talk) 21:44, 13 June 2026 (UTC)

Those are pretty clearly not LLM-written. Beyond the references (which I haven't checked), the only real LLM-like aspect is the bullet list with headers, but even that doesn't seem very solid. For example, semicolons as separators aren't something this LLM formatting often comes with. A definitive indicator is the inconsistent formatting in:

* Type of drug;  brands of Tepoxalin available on the market<ref name=":7" />
	
* Wide use; Not only in veterinary medicine, also in humans. <ref name=":7" />

Beyond that, I disagree with the claim that comments like Immediately contact/seek a veterinarian if any signs are present. are necessarily signs of LLM writing. It strikes me as more similar to professional writing not familiar with how Wikipedia addresses the reader, rather than comments intended for the LLM user. More generally, many LLM "quirks" are amplifying existing patterns that were present in human writing (especially outside of Wikipedia), so it isn't surprising that some less careful writers copying the styles they were used to may encounter some superficially similar traits. Chaotic Enby (in solidarity · talk · contribs) 22:02, 13 June 2026 (UTC)

Any text in Wikipedia that was present before ChatGPT came out (November 30, 2022) can be safely assumed to not be LLM-generated. GPT-3 did exist back then but it was still relatively obscure. SuperPianoMan9167 (talk) 22:05, 13 June 2026 (UTC)

In any case, I remember playing around with early text generation models around GPT-2 or so. I can't remember now what the website was called, I think it was textsynth? At the time it didn't 'respond to prompts' but rather tried to continue from where you left off in whatever you typed, so "Once upon a time there was a..." would prompt something resembling a fairytale. "1, 2, 3, 4, 5" would prompt an endless list of counting integers.

Either way, those early models were much more rudimentary and had far more limited training data. I doubt you'd've had any success trying to use them to give you anything that would work for Wikipedia; and people weren't really using it for that back then, either. It was a novelty and indeed far more obscure than anything after the release of ChatGPT. Athanelar (talk) 09:55, 14 June 2026 (UTC)

Talk to Transformer, I think Gnomingstuff (talk) 20:23, 14 June 2026 (UTC)

When I went to Talk to Transformer (which I found out about from Two Minute Papers' video), I would give it a prompt and use the last part of its output as a new prompt. For laughs, I would start with something silly for a prompt and see if GPT-2 can come up with other silly things, which it did a lot. It often generated text that looked like it couldn't decide if it was an online news article or a wiki article. – MrPersonHumanGuy (talk) 13:47, 8 July 2026 (UTC)

That's how LLM interfaces were before someone came up with the idea of wrapping them in "chatbot" interfaces instead. They still work the same way. The system prompt comes first, then the previous parts of the "conversation", and then the current "message" that you send. Then the algorithm predicts what should come after that.

The chatbot wrapper interface makes it seem as if you are "talking to someone", and my theory is that it is this is what gets people really engaged/hooked on using them. Like it triggers something in peoples brains or "hijacks" their thought cycles, because we're so used to using chat interfaces to talk to other people. ‑‑gurkubondinn 13:59, 8 July 2026 (UTC)

• Four years is the answer. And as others have noted, all of the "quirks" that LLMs have in their writing style was obtained because they were copying the most common generic writing style in the data given to them. Meaning that said writing style is also the most common, generic manner of writing out there, meaning it's not unexpected to run into it in the wild even without LLMs involved. It also means there's a decent chance even nowadays to have a false positive when claiming someone's writing is LLM-generated. And that's without considering the recursive effect that all the LLM use is having on molding people's actual writing style now because of it. Silverseren^C 22:23, 13 June 2026 (UTC)

  Regarding all of the "quirks" that LLMs have in their writing style was obtained because they were copying the most common generic writing style, that is a common claim, although I believe a hyperbolic one. Training methods lead to specific writing styles being reinforced and patterns popping up way more often in LLMs than in human writing, rather than them averaging out all of their training data or even broadly reflecting it. Chaotic Enby (in solidarity · talk · contribs) 22:27, 13 June 2026 (UTC)

  Yeah, I think a lot of the "I hope this helps!" and other sycophantic behavior emerges because those kind of responses score higher with RLHF reviewers during fine-tuning. SuperPianoMan9167 (talk) 22:35, 13 June 2026 (UTC)

  Yep, RLHF was the main pipeline I had in mind here. Chaotic Enby (in solidarity · talk · contribs) 22:51, 13 June 2026 (UTC)

  Two of the studies cited on AISIGNS (the Juzek/Ward) corroborate this — this stuff doesn’t show up as much even in output from base models. I will beat this drum until people start listening (i.e. forever) alas Gnomingstuff (talk) 20:21, 14 June 2026 (UTC)

  That's without even mentioning that these models are not even really trained on "human writing" in general, but whatever digitised, machine-scrapable human writing is out there, which is not necessarily a representative dataset. Athanelar (talk) 09:57, 14 June 2026 (UTC)

The main reason I suspect an LLM is the hallucinations. A 2008 journal issue is cited for a claim that the drug was withdrawn in 2017, with the 2017 claim being incorrect as the withdrawal is mentioned in my 2016 source.

Claim about antihistamines, which isn't in the source (probably hallucinated from , which doesn't support such a claim but is the only study mentioning both tepoxalin and antihistamines)

'The Committee for Medicinal Products for Veterinary Use (CVMP) approves Tepoxalin to be used as a drug for animals to reduce inflammation and pain control', cited to what is literally just an image of the synthesis of the chemical. Said citation is repeated several times for other claims that it obviously cannot verify.

is cited, which is a completely different drug (FTY720)

'As a result, the usage of carprofen was replaced with Tepoxalin in 1998', cited to something published in 1995.

Pretty much everything in the edit is a hallucination. I cannot see a reason why a human would make such mistakes unless they were intentionally trying to pretend they are using an LLM. Traumnovelle (talk) 22:51, 13 June 2026 (UTC)

Before LLMs were a thing anyone knew about? Silverseren^C 23:08, 13 June 2026 (UTC)

Or maybe the human author was just atrocious at source-text verification. SuperPianoMan9167 (talk) 23:19, 13 June 2026 (UTC)

In practice, it started on Wikipedia around March 2023, ratcheted up around fall 2023, really surged around late 2024, and has remained high. A good barometer here is spammers as they are early adopters of this stuff; AI spam only really emerged in early to mid 2023 Gnomingstuff (talk) 20:19, 14 June 2026 (UTC)

Well, technically GPT-2 was released in February of 2019, and was the first model that could generate plausible-sounding nonsense resembling Wikipedia articles. Practically, LLMs were obscure before GPT-4 release, when the whole hype began. Considering the vastness of Wikipedia, I'm sure there are edits which were generated or assisted by GPT-2/GPT-3, but there's a very small (<1000) number of them. sapphaline (talk) 13:56, 8 July 2026 (UTC)

Personally, I have wondered if Wikipedia may have been used as a "live testbed" for these companies. Inserting generated text and seeing if it gets removed as some sort of measure of how "good" the output is. ‑‑gurkubondinn 14:25, 8 July 2026 (UTC)

•  Looks AI-generated to me. (Template:Looks AI-generated)
•  Sounds like an AI chatbot beeping into a megaphone to me (Template:MegaphoneAI)
•  1.75x amplified ultimate beep of ultimate destiny (Template:MegaphoneAI with parameter ultimate)

Inspired by {{duck}} and {{megaphoneduck}}. – MrPersonHumanGuy (talk) 23:06, 26 June 2026 (UTC)

Love them! Would they fit better with the current megaphone icon (in line with their duck counterparts) or with something like or (to keep the design unity with the robot icon)? Personally, I like both, and the current ones are very fine to me, just spitting out ideas! Chaotic Enby (in solidarity · talk · contribs) 23:46, 26 June 2026 (UTC)

 Done. I went ahead and replaced the current megaphone icon with the first megaphone icon you brought up because it's the most similar. – MrPersonHumanGuy (talk) 10:33, 27 June 2026 (UTC)

FYI, those don't display well in dark mode. I don't know if there's any way to make the color dynamic. ChompyTheGogoat (talk) 05:32, 14 July 2026 (UTC)

I love these. Maybe the analogy with SPIs/LTAs will make the point that this has become a serious problem for us. —In solidarity with Wiki Workers United · ClaudineChionh (she/her · talk · email) 04:58, 27 June 2026 (UTC)

How are these meant to be used? They come across as snarky to me, especially the last one. SenshiSun (talk) 19:03, 7 July 2026 (UTC)

They're inspired by the {{duck}} series of templates which are often used in WP:SOCK investigations. I'd say these templates are meant for editors comparing notes and suspicions with each other, rather than directly communicating with people we suspect are using AI. —ClaudineChionh (she/her · talk · email) 02:08, 8 July 2026 (UTC)

I've made a table indexing user pages which follow a particular pattern. Several of these contain indications like lists, title case, em dashes, emoji, and Markdown.

More information Userpage, Contents ...

Userpage	Contents			Section titles			Month of revision
	L	E	M	Introduction	Hobbies and interests	Communication
404hugo	N	N	N	Welcome to My Wikipedia User Page!	—N/a	—N/a	February 2025
AbdulRehmanMemon	Y	N	N	About Me	Editing Interests How I Contribute	Collaboration	November 2025
Ahsanghaffar101	Y	N	N	About Me	Expertise and Interests Contributions to Wikipedia	Let's Connect	November 2023
AnjitAgarwal	Y	Y	N	👋 About Me – Anjit Agarwal 🎓 Education	💻 Skills & Expertise 🧭 Goals & Vision 🔍 Hobbies & Interests 🌟 Personal Traits	📬 Let’s Connect	June 2025
AlexisCdR	N	N	Y	1. About Me	—N/a	—N/a	November 2024
Alwayscaffinated	Y	Y	N	Welcome to My User Page! 🌍📚☕ A Little About Me	My Passion for Open Information My Contributions: A Blend of Interests Collaborations and Community Engagement Languages and Interests	Let's Connect and Collaborate	March 2024
Archromeo	Y	Y	N	👋 About Me	🧠 Interests 🔧 How I Contribute 🎯 Goals 🏆 Milestones 🧰 Tools I'm Learning	🤝 Let's Collaborate	June 2025
Arumobileworld	Y	Y	Y	👋 Welcome to My Wikipedia User Page! ✨ About Me	🏆 My Contributions	💬 Let's Connect!	March 2025
Ashu2909	Y	N	N	About Me	Areas of Interest How I Contribute Philosophy Future Contribution Goals	Disclaimer	May 2025
BantikumarWiki	N	N	N	Welcome to User talk:BantikumarWiki! About Me	Editing Philosophy Contributions Awards and Recognitions	Contact Information Let's Collaborate!	August 2023
Bapanwon	Y	Y	N	About Me Personal Background Professional Life	Wikipedia Contributions Fun Facts Wikipedia Philosophy	Contact External Links	January 2024
Beingratnakar	Y	N	N	About Me	Why I'm on Wikipedia How I Contribute	—N/a	July 2025
Bhaskar sunsari	Y	Y	Y	👋 Welcome to the User Page of Bhaskar Sunsari 🧑‍💻 About Me	🌏 My Interests 🎯 What I’m Working On	📬 Let’s Connect!	April 2025
Celtsystem	Y	N	N	About Me	Interests Contribution Philosophy Conflict of Interest (COI) Statement How I Contribute	—N/a	August 2025
Cosmicom01	Y	N	N	About Me	Editing Approach AI Use Tools I Use	Let’s Collaborate Thanks	June 2025
Danieloliver7	Y	Y	N	—N/a	🧠 Interests How I Contribute External Involvement	—N/a	May 2025
Dr Pius Adie	Y	Y	N	👤 About Me 🎓 Academic & Professional Background	📚 Areas of Interest 🌍 My Vision	🤝 Let’s Collaborate	June 2025
Dreamblue69	Y	Y	N	About Me	My Interests	Let's Collaborate	October 2025
EdEscaMar	Y	N	N	About Me	My Involvement in Wikipedia Areas of Interest Related Projects	Let’s Collaborate!	June 2025
Erikdarrin	Y	N	N	About me	Contributions Why I edit	Let's connect	May 2025
Garreth van Niekerk	Y	N	Y	About me	Conflict of interest How I contribute	Contact	November 2025
Guggger	Y	N	N	About me	Editing interests How I contribute	External link	December 2025
HKG 2026	Y	N	N	About Me	My Interests How I Contribute Editing Principles	—N/a	June 2026
Iamadityaalive	Y	N	N	About Me – Aditya Kumar	What I Do My Mission	Why Follow Me? Let's Connect	December 2024
Indiepostrockmegazine	Y	N	N	About Me	My Interests How I Contribute	Let's Connect	July 2025
James.aminian	Y	N	N	About Me	Interests Contributions	Contact	May 2025
Jonathan3340	Y	Y	N	👤 About Me	🛠️ Current Projects & Focus Areas	📬 Let's Connect	June 2026
JustineHC20	Y	N	N	—N/a	Areas of Editing Interest How I Contribute Current Projects User Philosophy	Contact	November 2025
Khokhar1977	Y	N	N	About Me	Interests Goals	Contact	July 2025
Lalit bc	Y	Y	N	👋 About Me 📚 Academic & Professional Background	🛰️ Interests 🌍 Projects & Initiatives 🛠️ Tools & Platforms 🖋️ Contributions on Wikipedia	💬 Let’s Connect!	April 2025
Luca-pattern-98	Y	N	N	Welcome to My User Page	My Focus Current Projects About Me	Let's Collaborate	July 2025
MaahirSehgal	Y	N	N	About Me	Areas of Work Editorial Philosophy Notable Contributions Useful Links	Talk to Me	April 2025
Mainno Mpanza	Y	N	Y	*About Me:* *My Story:*	*Performances and Achievements:* *What I Do:*	*Let's Connect:*	November 2025
MamertusSheez	N	N	Y	About Me	Contributions Interests	Contact	February 2024
Miss Hope so	Y	N	N	About Me	My Interests Skills I’m Developing	Let's Connect	May 2025
Mr Hedgehog UA	Y	N	N	About Me	My Interests How I Contribute to Wikipedia Useful Links	Contact	February 2025
MrMonkEdits	Y	N	Y	About Me	Editing Style Interests	Let's Collaborate!	December 2024
Nicholasgyamfi	Y	N	N	About Me	Contributions Areas of Interest	Let's Collaborate	April 2025
PaigeLangton	Y	N	N	—N/a	Paid Editing Disclosure How I Contribute What I Do Not Do	—N/a	November 2025
PeaceLoveKumbaya	Y	Y	N	Welcome to My Awesome User Page! About Me	My Contributions My Editing Philosophy	Contact Me	May 2024
Prasenjeetl69	Y	N	?	—N/a	Skills Selected projects Publications & writing Open-source contributions Code of conduct / Wikipedia note How I contribute to Wikipedia	—N/a	September 2025
Rana DG	Y	Y	N	👋 Hello from RanaDG!	What I Do (In and Out of Wikipedia) Skills & Passions A Bit More About Me	Elsewhere on the Web Let’s Collaborate	July 2025
RaviTejaAlchuri	N	N	N	About Me	Wikipedia Contributions	Contact	August 2023
RezainShafa	N	N	N	About Me Personal Life	Interests and Background Contributions Collaboration	Contact	May 2023
S.H.&.SONS	Y	N	N	About Me	My Interests on Wikipedia Languages	Let's Collaborate	June 2025
*sherazzee*	Y	N	N	About me	Editing focus Drafts and contributions Disclosures Outside Wikipedia	Contact	September 2025
SproxtheWriter	Y	Y	N	About Me	My Editing Garage	Let's Connect	November 2025
Sukhleen mallan	Y	N	Y	About Me	Areas of Interest Wikipedia Best Practices Contributions Barnstars & Recognition	Collaboration & Contact	February 2025
Tanak001	Y	N	N	About Me	Interests and Hobbies Contributions to Wikipedia	Let's Connect	February 2025
UncleReyRey	N	N	N	About Me	How I Contribute Why I Edit	—N/a	May 2025
UroojWrites	Y	Y	N	🌟 Welcome to My Page ✍️ About Me	🎯 My Purpose on Wikipedia	💬 Let’s Connect	June 2025
Virilikestea	N	N	N	About Me	Contributions	Let's Connect	August 2023
Wasim Khalil Ali	Y	N	N	About	Editing interests How I contribute Disclosures	Contact	January 2026
ZeonDev	Y	Y	N	👤 ZeonDev 🛠 About Me	📝 Contributions 🎓 Skills & Expertise 🌍 Goals	📬 Reach Out	October 2024
苍野恰诺	Y	N	N	Welcome to Aono Chano's User Page! About Me	My Interests Current Projects	Let's Connect!	September 2024

Close

Just figured I'd compile my observations. I use PermanentLinks to avoid accidentally sending mention notifications to anyone on the table. If your name appears on this table, it doesn't necessarily mean that I'm certain you used AI to make your user page, although if you appear to have used Markdown on it, then I may be more confident that you have. – MrPersonHumanGuy (talk) 19:59, 2 July 2026 (UTC)

These always set off my spidey sense when I stumble across them. Have you found these by happenstance or are you searching for specific text? —ClaudineChionh (she/her · talk · email) 23:47, 2 July 2026 (UTC)

I came across a couple of them a while ago, which made me curious to see how prevalent their apparent format was. Out of this curiosity, I would occasionally try to find more userpages like those by typing in search terms like these:

• "About Me" "Let's Connect"
• "About Me" "Let's Collaborate"

Today, I decided to do this yet again so I could compile this table that I was originally going to post at WT:AISIGNS, but after a while, I realized that the table would have lots of entries, so I saved what I had been working on up to that point at a new subpage on my userspace instead and continued expanding it there. Whilst working on this table, I went over several user pages I had seen in the past and found many user pages that I hadn't seen before. – MrPersonHumanGuy (talk) 01:23, 3 July 2026 (UTC)

I don't have many userpage-related AI edit summaries, but they sometimes look like these (if something is here it's because the user has a pattern of AI edits with AI edit summaries):

• Create concise user page – bio, workflow, contact info, licence note (policy-compliant) (June 24, 2025)
• Created a new user page with a neutral and informative self-introduction. (July 5, 2024)
• Created an informative Wikipedia user page content highlighting the establishment, product offerings, and operations of [redacted], focusing on its role as a trusted e-commerce store in Pakistan. (November 18, 2024)
• Created user page for [redacted] outlining editing interests, contributions to Philadelphia hip hop articles, sandbox drafts, and collaboration goals. (July 29, 2025)
• Creating user page: focusing on general maintenance and WP:BLP compliance (February 16, 2026)
• Revamped user page: added personality, humor, focus areas, sandbox link; WikiProject banners moved to talk page. (October 11, 2025)
• Updating my User Page to reflect professional expertise in SaaS solutions within the 'Professional Roles' framework. Anchoring technical background to support ongoing Data Validation and Forensic Auditing projects (1963 Gazette/UNESCO/KNBS) per WP:USER. (February 27, 2026)
• Updating user page: adding interest in new article creation and general maintenance (February 16, 2026)

Gnomingstuff (talk) 23:48, 2 July 2026 (UTC)

Another thing I find funny is that some of these pages have a section about having a conflict of interest, but they just say something along the lines of "I will disclose any conflict of interest I have in cases where I would have to do so" and don't disclose which topics they have a conflict of interest on. – MrPersonHumanGuy (talk) 10:07, 8 July 2026 (UTC)

Another potential candidate: Syedhashimpak ChompyTheGogoat (talk) 05:38, 9 July 2026 (UTC)

This one also looks dodgy: Townsaiso, with the telltale Markdown fenced code block in the first revision. But these last two don't have the abundance of emojis that some of the earlier ones do, so not sure if it's the same behavioural pattern. —ClaudineChionh (she/her · talk · email) 05:46, 9 July 2026 (UTC)

Wow - on a 15 year old account, with seemingly valid (if limited) historical contributions. That's disappointing.

A number of the ones in the table don't use emojis either, but the phrasing is a dead giveaway - and both have already been called out for it in articles. ChompyTheGogoat (talk) 06:51, 9 July 2026 (UTC)

this stuff is just really easy to find in general; there are false positives in this search but not many Gnomingstuff (talk) 02:39, 11 July 2026 (UTC)

Here's a good one:

Here is the updated draft bio with wallsil2k7 as your primary professional moniker, replacing the previous name.
— User:Wallsil2k7

Originally found this user when they tried to submit a chatbot-generated autobiography: AbuseLog/44648204. ‑‑gurkubondinn 11:29, 15 July 2026 (UTC)

The bot even tried to warn them. ChompyTheGogoat (talk) 12:43, 15 July 2026 (UTC)

• Special:AbuseFilter/1407 logs the removal of the "ai-generated" family of tags including {{AI-generated}}, {{AI-generated source?}}, {{AI-generated span}} and {{AI-generated inline}}. View the log here.
• Special:AbuseFilter/1408 logs the removal of {{prod llm}} tags. View the log here.

The request that created these can be seen here. fifteen thousand two hundred twenty four (talk) 06:42, 4 July 2026 (UTC)

I keep seeing people trying to use the results of an AI detector as the only justification for a G15 tag (recent example), which is insufficient. I created the shortcut WP:AIDETECTION and added some info at the signs of AI writing page stating this, but I think it would be even better if G15 explicitly disallowed the use of AI detector results as the sole justification for speedy deletion, because these tools are notoriously unreliable. (That is, any G15 tag that has "100% AI detected" as the only given reason can be summarily declined as invalid.) Do you think adding this to G15 is a good idea, and if so, how should this be worded? (This is separate from the proposal to amend G15 above.) SuperPianoMan9167 (talk) 23:02, 5 July 2026 (UTC)

That is, any G15 tag that has "100% AI detected" as the only given reason can be summarily declined as invalid. This describes the status quo. G15 only applies in two cases: when there is "communication intended for the user" and "non-existent or nonsensical references".

Additionally, the end of G15 already says: "In addition to the clear-cut signs listed above, there are other, more subjective signs of LLM writing that may also plausibly stem from human error or unfamiliarity with Wikipedia's policies and guidelines. While these indicators can be used in conjunction, they should not serve as the sole basis for applying this criterion." voorts (talk/contributions) 23:24, 5 July 2026 (UTC)

The problem is that people still seem to think that AI detector results are a valid justification for G15 even if the policy doesn't allow for it. I think G15 would benefit from explicitly stating "do not use AI detector results" as it would likely cut down on these sorts of invalid G15 tags, even if this problem is caused by people not reading the directions. SuperPianoMan9167 (talk) 23:27, 5 July 2026 (UTC)

The problem is that people still seem to think that AI detector results are a valid justification for G15 even if the policy doesn't allow for it. And they quickly learn that they're incorrect when another editor declines their CSD. What's the problem? voorts (talk/contributions) 23:29, 5 July 2026 (UTC)

It happens frequently enough that clarifying it on the policy page would most likely be worth it. Unfortunately, gathering evidence for this observation is quite difficult because you can't directly search for edit summaries or diffs making a particular change across multiple pages. I tried finding more examples to no avail. SuperPianoMan9167 (talk) 00:06, 6 July 2026 (UTC)

People misuse the CSDs all the time. I don't expect this change will change anything, particularly since G15 is limited and already says not to rely on LLM detectors. voorts (talk/contributions) 15:04, 7 July 2026 (UTC)

Support, for now, pending any kind of partnership/integration pipeline. (The problem is that when people actually are using AI, they use the existence of the detector as an attempt to dismiss the whole thing.) Gnomingstuff (talk) 23:34, 5 July 2026 (UTC)

Support. Clarifying this means users will need to explain why content is AI generated instead of pointing to a checker with opaque logic. SenshiSun (talk) 18:54, 7 July 2026 (UTC)

Are there really that many people doing this? I don't do G15s, so I don't know the stat there, but I have now gone through each of the 6,000 articles with the AI-generated tag (as of May), and only a handful of those mentioned detectors at all. Gnomingstuff (talk) 06:39, 8 July 2026 (UTC)

Yes please - they currently just don't cut it ~ Squawk7700 (talk) 00:02, 14 July 2026 (UTC)

Support. It won't harm anything to make this explicit. Thryduulf (talk) 01:22, 6 July 2026 (UTC)

How should it be worded? Maybe something like Do not use the results of artificial intelligence content detection tools as the sole basis for applying this criterion. with one footnote giving some popular examples (GPTZero, Turnitin, etc.) and another footnote explaining these tools' non-trivial error rates. (I didn't have any specific wording in mind when I started this discussion; hence why I started it here instead of at WT:CSD.) SuperPianoMan9167 (talk) 01:29, 6 July 2026 (UTC)

Support and the above text sounds OK. Graeme Bartlett (talk) 23:54, 6 July 2026 (UTC)

Support making it explicit. Above text sounds fine to me. Netstars22 (talk to me!) 15:01, 7 July 2026 (UTC)

Comment I keep hearing about this "notoriously unreliability" of AI detection software, and I agree that it shouldn't be sufficient as the sole basis of a G15, and I don't rely on tools like that in my own personal judgements about text, but I have yet to see a single example of verifiably non-AI text that pings 100% AI on GPTZero or a similarly-reputable tool. I would really love to be shown such an example. The articles I've read about AI false positives have all been cases where the study counted a 51% AI confidence level on human text as a false positive. GPTZero at least claims that their most recent models have a false-positive rate of classifying human text as AI of 0.24%. If there's evidence refuting that claim I'm all ears to hear it. -- LWG ^{talk (VOPOV)} 20:32, 23 July 2026 (UTC)

The good AI detectors have a fairly high accuracy rate, the problem is that they’re also not unlimited so people use the bad AI detectors Gnomingstuff (talk) 01:03, 24 July 2026 (UTC)

 You are invited to join the discussion at WP:VPI § Retention of editors who've been caught using LLMs. Kowal2701 (talk, contribs) 18:54, 7 July 2026 (UTC)

The current AI user templates focus on recent edits, which does not make sense if someone made AI edits years ago, but they've just been identified. SenshiSun (talk) 19:06, 7 July 2026 (UTC)

Is a warning necessary if the edits were years ago? I would expect that in that situation, either they've stopped making AI edits, in which case nothing needs to be said, or they're still making AI edits, in which case you could warn them about one of their recent ones. Justin Kunimune (talk) 20:48, 7 July 2026 (UTC)

There might be a case if a user's current edits are too small to tell, or if there's an investigation. SenshiSun (talk) 00:17, 9 July 2026 (UTC)

If there is an investigation about their recent edits then wait until it's finished, there are three possible outcomes:

• They're not currently using AI in which case there is nothing to be said
• They still are using LLMs, in which case it doesn't matter for the purposes of warning them whether their old ones also were or not any warnings about their current edits will explain the rationale, etc. and don't say or imply that it's only recent edits that have issues.
• The outcome is inconclusive, in which case templated messages are not going to be appropriate due to a combination of assuming good faith, and needing to explain what issues there are with their writing that makes them look possibly AI-generated.

If you can't tell whether the edits are AI-generated or not then it doesn't matter whether they are or not. If they're otherwise good edits then they benefit the encyclopaedia, if they're otherwise bad edits then they can and should be removed for whatever non-AI-related reason they're bad. Thryduulf (talk) 08:55, 9 July 2026 (UTC)

@SenshiSun You're right that the vast majority of warning templates assume whatever issue is one that's quite recent; {{AINB-notice}} doesn't, if it's an issue that you've directing to AINB, but I do get that this only works in the context of cleanups. If you do need to let somebody know that you've had to undo/extensively clean up after an older edit of theirs (which is something I do, from time to time, though not typically in an AI context ), then this is a circumstance where I've found it's often better to leave a short, personalized note. It's okay if it's not perfectly worded. GreenLipstickLesbian💌🧸 09:16, 9 July 2026 (UTC)

Good catch. Thank you. SenshiSun (talk) 13:29, 9 July 2026 (UTC)

One of the ways that I do AI cleanup is regularly check for new introductions of utm_source=chatgpt. I often place {{uw-ai1}} on the talk pages of editors who added the offending content, and context-dependently may revert their edit in whole, in part, or just check the source directly and remove the UTM param if the source does actually match the text (and the text is not obviously/likely LLM written).

Sometimes I get objections from people that they only used ChatGPT to retrieve the source and not to write the content; in some cases, the text is still fishy, but sometimes it seems a reasonable defense. Still, I regularly have to explain that LLM-retrieved citations often do not align with the text they are being attributed to, and thus require editors (like myself) to have to double check that they do.

I think a series of UW templates along the lines of {{uw-aicite1}} etc. would be useful for not having to repeat the same thing over again, and would be slightly more fitting than simply {{uw-ai1}}. However, I also recognize that the UTM parameters are something of a honeypot, and making offenders aware of what flagged their additions may just make them try to conceal it more (not that what I am doing now doesn't already have this caveat).

Would appreciate hearing others' thoughts on this! ~ oklopfer (💬) 12:40, 8 July 2026 (UTC)

In past discussions there, a mixture of detection utility (honeypot) and actual utility of using llms to find sources has been raised regarding utm parameters. I think your process is a good example of the former, creating templates may not be helpful in that respect, especially as the real problem is often in the source use. CMD (talk) 12:51, 8 July 2026 (UTC)

Side note: the "possible AI-generated citations" auto-tag does not seem to be the same as Special:AbuseFilter/893. Is there a separate edit filter that does similar to my search pattern/the auto-tag? ~ oklopfer (💬) 12:55, 8 July 2026 (UTC)

Special:AbuseFilter/1346 also exists and tags "possible AI-generated citations", however the only utm_source that it matches to is copilot. You may want to ask at WP:EFN for an update. Special:AbuseFilter/893 only tracks websites that have content generated by AI, and only has one website tracked (http://publifye.com/). ARandomName123 (talk)^{Ping me!} 23:57, 8 July 2026 (UTC)

No need for EFR, 1346 also matches utm_source=(chatgpt|askpandi|deepseek|copilot\.microsoft|m365copilot|gemini\.google|groq|grok). fifteen thousand two hundred twenty four (talk) 08:26, 9 July 2026 (UTC)

In my experience, only copilot, chatgpt, and perlexity add their own utm_sources. I can't find any instances of the other sites being linked by a utm_source. FlammablePizza (talk) 14:06, 9 July 2026 (UTC)

Grok sometimes adds referrer=grok.com to links. ‑‑gurkubondinn 14:10, 9 July 2026 (UTC)

Thanks, that's that one I was looking for! ~ oklopfer (💬) 15:57, 9 July 2026 (UTC)

Would a bot that adds {{AI-retrieved source}} to such sources be helpful? I'm aware of the edit filter tags, but automatically adding this template would make it more obvious what needs to be cleaned up. It also wouldn't affect the honeypot too badly due to not warning the offender (and the AIs will keep adding it due to their programming). I'm currently busy with an unrelated BRFA, but I have been playing around with a prototype. Here is an example diff: . It can also manage a table of un-checked ai-retrieved sources: . Around a third of all obviously AI-retrieved sources appear to already be tagged. FlammablePizza (talk) 14:49, 9 July 2026 (UTC)

For anyone curious, here are the identifiable untagged AI-retrieved sources (2,347 of them). [7] [8] [9] [10] FlammablePizza (talk) 14:54, 9 July 2026 (UTC)

Maybe, but the problem is that when you add that tag, people tend to just remove the tag. What theyre supposed to do is to set |checked={{CURRENTMONTHNAME}} {{CURRENTYEAR}}. They'll often also remove the tracking parameters from the reference URL as well, which prevents tools like m:Cite Unseen from highlighting the reference as being AI-retrieved. ‑‑gurkubondinn 15:27, 9 July 2026 (UTC)

Could be worth doing an edit filter for that as well? InfernoHues (talk) 00:33, 11 July 2026 (UTC)

Certainly wouldn't object to that. ‑‑gurkubondinn 00:41, 11 July 2026 (UTC)

Usually whenever I work on a movie which has been tagged, I go for the entire article. The problem are the movie plots which frustratingly don't have sources. They dont need to, I know, but it's hard to discern between AI texts and the ones written by the movie fanatics who summarizes it after their fourth of fifth viewing of the movie. The latter of course might have original research but beats being written by a bot.

WWT helps a ton but to parse all the AI text, connect with the existing one, rewording it all without even knowing the contents of the movie, is a slow process to be honest. What I've been doing is to just empty the section but there goes the plot which might've been written years beforehand. Would that be fine then?

PeepeeDino (talk) 08:19, 10 July 2026 (UTC)

I'll be honest, I would just not touch the Plot section if I were you (unless there's an earlier, pre-LLM version of the Plot section you can revert to). There are some editors who are very enthusiastic about cleaning up and editing Plot sections. Much like editing categories or MEDRS pages, such things are not meant for us mere mortals. I would save yourself the headache. Cheers, Suriname0 (talk) 14:30, 10 July 2026 (UTC)

Well, I suppose from an entire article being flagged to just that section is still something nonetheless.

PeepeeDino (talk) 22:51, 10 July 2026 (UTC)

Hi PeepeeDino, what is WWT? I'm not familiar with what you are referring to. Softlavender (talk) 00:18, 11 July 2026 (UTC)

Oh, Who Wrote That? Great extension an editor introduced to me. Saves some time in going through the revision history.

PeepeeDino (talk) 01:46, 11 July 2026 (UTC)

 You are invited to join the discussion at Wikipedia talk:Speedy deletion § RfC: Updating G15 ("LLM-generated pages without human review"). Ca ^{talk to me!} 16:23, 10 July 2026 (UTC)

Check out WikiTomte-LLM on Toolforge, an easy (and possibly fun!?) tool for finding undetected LLM text in articles, based on Gnomingstuff’s excellent guide to finding AI-generated text. It runs searches for random combinations of LLM vocabulary words and gives you lists of potentially-suspicious articles that aren’t yet tagged for AI cleanup. It’s new and still a little slow and janky, but let me know what you think!

I’m also interested in how to help chip away at the problem of 7000+ articles tagged for AI cleanup, so I tried running the list of articles through Projo, a tool that can rank a set of articles based on page view counts: https://projo.toolforge.org/jobs/08eaa?min_pageviews=20000&names_only=1&max_quality=1&sort=pageviews&dir=desc. This is a way of finding articles that would be especially helpful to clean up for our readers. Dreamyshade (talk) 02:15, 11 July 2026 (UTC)

Really cool stuff! I just tried it out, and it found some LLM text, in addition to promotional writing more generally. InfernoHues (talk) 02:28, 11 July 2026 (UTC)

Just wanted to chime in here, thanks for making this Gnomingstuff (talk) 14:40, 11 July 2026 (UTC)

Sounds amazing, thanks a lot! We definitely need to make a cool "tools" tab for all of these (update our "resources" page maybe?) Chaotic Enby (in solidarity · talk · contribs) 15:27, 11 July 2026 (UTC)

There is an ongoing discussion at WT:AFC#Option to reject AI generated submissions considering the option to be able to reject AI-generated AfC submissions. You may be interested to participate in the discussion. Thank you. Fortek67 (talk) 13:02, 12 July 2026 (UTC)

I wrote some suggestions at Wikipedia talk:Presumptive removal of AI-generated content#Clarifications for reversing removal of content and would like to invite input from anyone interested. Dreamyshade (talk) 17:20, 12 July 2026 (UTC)

 You are invited to join the discussion at WT:AISIGNS § Signs of AI generated edit summaries. Athanelar (talk) 05:26, 18 July 2026 (UTC)

I read a post and couldn't see any way of interacting. ~2026-40439-19 (talk) 19:30, 18 July 2026 (UTC)

Are you referring to WP:AI or not quiz? Ca ^{talk to me!} 09:40, 22 July 2026 (UTC)

 You are invited to join the discussion at WT:LLMRESP § Section on copyediting. Kowal2701 (talk, contribs) 18:16, 19 July 2026 (UTC)

Possibly the question here is "where do I discuss this?"

If appropriate, the main question is this:

Has anyone else started noticing random temporary accounts suddenly commenting on your old user talk sections? This appears as a personalized message of understanding, as if I needed emotional comfort when having a perfectly normal disagreeing with other users? You can find reverted additions on my user talk if you need examples.

Assuming the answer is "yes, we've noticed" where is this discussed? Thanks CapnZapp (talk) 12:03, 21 July 2026 (UTC)

Interesting. Here’s a sample For others, here’s a sample

I don’t know. This is a bad person you are dealing with. Very dishonest him and his partner just do some deep checking. If they would throw their own family under the bus matching what they would do to anybody else.

Dw31415 (talk) 12:09, 21 July 2026 (UTC)

This is not AI-generated, though. sapphaline (talk) 12:40, 21 July 2026 (UTC)

Unless… (And I’m stretching here)… someone is training their own model and testing it out on talk pages. Dw31415 (talk) 15:06, 21 July 2026 (UTC)

Or instructing it to make grammatical errors to appear like a user and "warm up" the account (this is stretching that requires bad faith readings, but I haven't even looked at any of these accounts, just coming up with a hypothetical explanation for why someone would do this with an LLM). Not all LLM output is required to be grammatically correct or free of typos and spelling errors. These are algorithms of mimicry, so you can feed them instructions to mimic grammar errors if that is what you want. ‑‑gurkubondinn 12:05, 23 July 2026 (UTC)

Checked your talk page history now and the first example that I found was by an TA, so it's not about "warming up" an account. ‑‑gurkubondinn 12:19, 23 July 2026 (UTC)

I haven't noticed this, it feels like a situation that may be specific to you Gnomingstuff (talk) 00:17, 22 July 2026 (UTC)

Thanks. Just random coincidence then. (Obviously I wouldn't have reacted if it was just the one message. It's only when several unrelated accounts suddenly start offering unsolicited advice using florid language I'm starting to wonder of this is a wave of LLMs) Good to know CapnZapp (talk) 11:09, 23 July 2026 (UTC)

I've been going through the users in Category:User talk pages with large language model notices and checking their contribs to see if cleanup is needed. Once I've confirmed that any damaging contribs from the user have been dealt with, it would be nice to indicate that in some way so future AI patrollers don't waste time. In other cleanup categories I remove the tag after completing cleanup, but here that would involve editing the warning on another user's talk page. What do you all recommend to minimize duplication of labor? -- LWG ^{talk (VOPOV)} 16:32, 24 July 2026 (UTC)

This discussion (to me) looks to have some LLM responses in it: Talk:Battle of Lumë#RFC on Infobox for Battle of Lumë. I don't want to collapse/ask myself because I've !voted in the RfC. InfernoHues (talk) 20:53, 25 July 2026 (UTC)

I skimmed it but didn’t see signs beyond the length. Any specific passages? Dw31415 (talk) 23:48, 25 July 2026 (UTC)

Some of FranéRogoz and Magapetro's comments seemed weird to me. Especially how they seem to kind of but not really respond to what people said. InfernoHues (talk) 02:43, 27 July 2026 (UTC)

I thought I'd get to work on some namespage pages after all my blathering at the guideline. I tried one at Talk:Abas (mythology)#Topos Text is AI? and was thankfully rescued by another editor. It dawned on me that it'd be much easier to do on pages I care about.

I created User:Dw31415/AI Cleanup to contain the projects I care about. It was a good exercise to continue to learn the db schema so I didn't spend much time looking to see if this was already done.

Probably the best place to put this is on a subpage of each project I'm interested in (like Wikipedia:WikiProject Medicine/Deprecated archival service).

Any thoughts, recommendations, requests?

Thanks to all you editors on the this project! Full disclosure, I created the sql using an agent. Dw31415 (talk) 12:27, 26 July 2026 (UTC)

@Dw31415: Nice. I have no idea how these work, but would it be possible to create a template so that interested WikiProjects can just simply substitute that and create a page like this? Kovcszaln6 (talk) 13:18, 26 July 2026 (UTC)

The projects go right at the top so it’s pretty easy to modify (see below). I like the idea of the template but won’t have time to do that for another week or so. I’d be happy if someone else wants to do that. It’d be helpful to have one even without parameters because it’s not obvious where the sql template starts and the results table ends.

WITH selected_projects AS (

/* Edit this list to choose projects. Titles are PageAssessments titles. */

SELECT 'Aviation' AS project_title

UNION ALL

SELECT 'Film' Dw31415 (talk) 14:15, 26 July 2026 (UTC)

maybe Template:AI Tag By Project is a good name? Dw31415 (talk) 22:11, 26 July 2026 (UTC)

Hi all, I've got a couple user scripts for anyone interested in identifying AI media.

• AIMarker adds an AI watermark to images that have been identified as having been AI generated, based on the templates and categories on the filepage. This doesn't reveal anything new, but more readily surfaces the fact that an image has been AI generated or modified.
• Over on Commons, AIDetect lets you search a file's metadata for signals of AI generation (e.g. C2PA statements, standard diffusion metadata, etc.). This can be helpful to find confirmation that an image has been AI-generated/modified (but of course it's easy for metadata to be stripped from an image by screenshotting, downloading from social media, etc., so it's only as good as the metadata that's present).

Thanks, ~SuperHamster Talk Contribs 13:15, 27 July 2026 (UTC)

Self-portrait (Yayoi Kusama) - this is a new one for me. I was like "why does this article describe an acrylic painting as a digital photograph?" Then I looked at the infobox image on Yayoi Kusama and it all became clear. -- LWG ^{talk (VOPOV)} 18:32, 27 July 2026 (UTC)

Is there a truer self-portrait than the face that stares back at you as you look at your Wikipedia page? CMD (talk) 00:38, 28 July 2026 (UTC)

The Simple Summaries did this sometimes, couple examples: A cuckold is a term used to describe a husband whose wife is unfaithful. It comes from a 19th-century painting by Cornelius Krieghoff., or This ancient painting shows Hercules carrying his son while facing the centaur Nessus. Hercules is a famous Greek hero, the son of the god Zeus. He's known for his strength and bravery, fighting monsters and completing tough tasks. In Roman times, he was called Hercules and was very popular. This picture tells a story from his adventures Gnomingstuff (talk) 20:15, 28 July 2026 (UTC)

The first version was somehow even worse. I've tagged the article for G15 because other than the issue you pointed out, there is communication intended for the user present in the infobox. Athanelar (talk) 12:15, 30 July 2026 (UTC)

 You are invited to join the discussion at WT:AISIGNS § New AIS: shortcut format. Athanelar (talk) 12:11, 30 July 2026 (UTC)

I'm aware there is a banner saying to report AI content on the AI noticeboard. I couldn't find instructions for how to list a single article. I have no intention of engaging in any cleanup or discussion of this article but Al Sa'dun is clearly LLM made and has some issues. If there is another process I should go through to report this, let me know and I'll do it, but I just wanted to inform the right people and stop thinking about it. I would recommend to the project to create instructions for the AI noticeboard for someone less familiar with the systems in place, but y'all probably have more than enough on your plate as is. – Ike Lek (talk) 20:09, 30 July 2026 (UTC)

Hi -- for something like this you can add the AI-generated tag to the top of the article and list the reason in the parameter, for instance, {{AI-generated|date=July 2026|reason=whatever your reason is}}. You will probably want to add this reason text to the talk page as well. Gnomingstuff (talk) 22:28, 30 July 2026 (UTC)

It's worth looking at the user who created it to see if there's much more to cleanup, in this case it's AytWaygher232 (talk · contribs) who's also created which appears to have a hallucinated category (WP:AIREDCAT). I've warned them but haven't done any cleanup Kowal2701 (talk, contribs) 17:03, 31 July 2026 (UTC)

Yeah, I think this might be someone who has read Wikipedia:Signs of AI writing and is trying to avoid the major signs. The red cat was a good catch since no category exists for any of the alternative spellings of the place, making a typo unlikely. Even then, it would be easy to argue that they were planning on creating the category. I can tell with a pretty high level of confidence that LLM generated text was used heavily in the article, but it is difficult to put together a case to prove it. I find this kind of work frustrating to a point that I choose not to do it most of the time. I respect what this project does, but I believe I can do a lot more good for Wikipedia as a whole if I don't burnout doing something I hate. Ike Lek (talk) 17:33, 31 July 2026 (UTC)

Probably more likely to be just an AI translation -- translated material can sometimes deviate more than non-translated text from the standard set of AI signs since there's at least some original text to influence the final product.

(but yeah, for every article I tag there are two or three I don't because, as you say, it is difficult to explain) Gnomingstuff (talk) 05:03, 1 August 2026 (UTC)