Background (G15)

• Preliminary discussion at WikiProject AI Cleanup

Survey (G15)

• B as proposer. In particular, the name change would better align the criterion with the guideline WP:Writing articles with large language models, which now disallows AI-generated content altogether, reviewed or not. Ca ^{talk to me!} 18:22, 2 July 2026 (UTC) Edited 14:08, 10 July 2026 (UTC)
• Oppose due to the text about any previous revision. If the page has now been fixed, ie reviewed and problems removed then there is no reason to delete. Also in regards to Q2, both current and proposed seem to be titled "LLM-generated pages without human review" so some clarification is required. Graeme Bartlett (talk) 00:21, 3 July 2026 (UTC)
  • If the genuine problems are removed, then the article is no longer substantially similar like that text explicitly requires. The unambiguous signs that turn a plainly LLM-generated article into something speedyable aren't the problem: anyone could easily remove "Here is your Wikipedia article on", while making the article truly acceptable is at least as much work as writing it from scratch. The usecase for this clause is for when someone removes only the unambiguous tells without fixing anything else. —Cryptic 00:44, 3 July 2026 (UTC)

  @Graeme Bartlett I've restructured the RfC to better clarify the distinction between the two proposals. Ca ^{talk to me!} 14:15, 10 July 2026 (UTC)

• Status quo The issue is with objectivity and there is no way to easily distinguish AI from human writing as many signs of human writing are similar to AI writing after all AI was trained on the latter. We can use G3 and G11 if the page is written as spam or vandalism regardless of AI writing, and should use XfD in less clear cases. Aasim (話す（はなす）) 16:02, 10 July 2026 (UTC)

  I feel like the presence of "oaicite" is pretty objective. G15 doesn't apply (with or without this change) when the only signs are writing style quirks that could plausibly be human-written. Justin Kunimune (talk) 16:47, 10 July 2026 (UTC)

  Oaicite may be objective, but the other signs such as ```wikitext should be taken on a case by case basis, as an online manual explaining wiki text could have a code block like this:

  ```wikitext
  = Heading 1 =
  == Heading 2 ==
  === Heading 3 ===
  [[Article link]]
  [https://example.com External link]
  ''Italics''
  '''Bold'''
  '''''Italics and bold'''''
  <parsertag foo="bar"/>
  {{Template name here|a|param=foo}}
  {{#if:a|foo|bar}}
  ```
  

  or something similar. But I do agree there likely isn't much use in wrapping an output in such tags unless if one is coding an LLM which automatically parses Markdown (like ChatGPT), also one can try to remove AI signs manually before publishing which can make it much harder to detect. Aasim (話す（はなす）) 17:53, 10 July 2026 (UTC)

  as an online manual explaining wiki text could have a code block like this

  Yes, they could, but do they? If not, then we shouldn't make policy around hypothetical infinite monkeys. Gnomingstuff (talk) 02:30, 11 July 2026 (UTC)

  I mean, it also wouldn't. This hypothetical manual (written by an infinite series of monkeys), might have been written in Markdown, but the webpage that the hypothetical user (one of an infinite series of monkeys) would be reading would have been converted into HTML. The actual markdown syntax would only be present in the Markdown source that was used to generate HTML on a webpage. ‑‑gurkubondinn 10:05, 11 July 2026 (UTC)

  Perhaps if someone created a manual on GitHub pages but I digress Aasim (話す（はなす）) 18:02, 11 July 2026 (UTC)

  Thats what I am talking about, markdown to HTML renderers. The markdown files are converted to HTML, this hypothetical reader would be looking at an HTML page in their browser. The markdown files are in the git repo, not on the "github sites" page. ‑‑gurkubondinn 19:14, 11 July 2026 (UTC)

• B – This third subcriterion seems to fit in cleanly with the other two as an objective way to distinguish AI from human writing, and being able to apply all three to previous versions also makes sense since if it was AI generated before then it's still AI generated now (barring substantial revision). I also agree that the proposed name change better reflects WP:NOLLM. — Preceding unsigned comment added by Justinkunimune (talk • contribs) 16:44, 10 July 2026 (UTC)
• B – The new "Technical indicators" are unambiguous signs of generated content and are perfectly in line with existing criteria. Removal of "reasonable human review" is also appropriate as it no longer aligns with our guidelines. At the time G15 was originally adopted there was no strong prohibition against LLM use, now with WP:NOLLM human review doesn't matter as generated content is prohibited regardless.
  The "substantially similar" addition is also reasonable. The literal presence of signs like turn0search0 isn't the problem, and removing them without also addressing the rest of the content which they indicate is generated should not be enough to block G15. fifteen thousand two hundred twenty four (talk) 17:00, 10 July 2026 (UTC)
• B per fifteen thousand two hundred twenty four. I actually wrote a fairly lengthy argument in favor of A at first but then I realized I still wasn't quite grasping the significance of the new guidelines; use of AI is now a conduct issue not a content issue so this should be seen as analogous to G5 in which use of AI renders the edit void ab initio in the same way as sockpuppet edits are. * Pppery * _{in solidarity} 17:05, 10 July 2026 (UTC)

  Moving to A only since Cryptic below and Aasim above made a convincing case. I had viewed this as abrogating this subjective component entirely and leaving it only as a page which has (or has ever had) specific tells. But I see now that was naive wishful thinking, and I've Cryptic points out below (re Markdown) that you can never underestimate the ingenuity of fools so that what I see as an objective AI tell may not be one. * Pppery * _{in solidarity} 18:15, 10 July 2026 (UTC)

• B, consistent with our broader guidelines and only using unambiguous technical indicators. Chaotic Enby (in solidarity · talk · contribs) 17:25, 10 July 2026 (UTC)
• B, per fifteen thousand two hundred twenty four (note: I wrote the "additional explanation" on the markdown code blocks). ‑‑gurkubondinn 17:30, 10 July 2026 (UTC)
• A only. The change to the title is going in precisely the wrong direction. The title is already a mismatch for the criterion itself and this makes it more so, and a substantial number of CSD taggers (and a handful of deleting admins!) only pay attention to speedy deletion criteria's titles. This is the same root cause of many of the problems we had with U5 - everyone saw "Non-contributor's misuse of Wikipedia as a web host" and treated it like that, even though the only thing it had in common with the criterion itself was that it targeted new users.
  The intent of the criterion and wording of its body is clear: it's meant for material that's subjectively LLM-generated that has at least one of a short list of specific tells. Remove either of these and the page isn't speedyable. That the current title isn't clear enough was enough to torpedo adding markdown and emoji to the tells - when someone used to using either puts that in an article, it doesn't look at all like when an LLM does, but people were afraid that just the tell would be enough, because "without human review" is such a poor match for the idea of a "specific tell". This proposal removes even that, so that people who skim the actual criterion, misremember it, or don't read it at all are left with just "subjectively looks LLM-generated". The pointer below to Wikipedia talk:WikiProject AI Cleanup#Proposed update to G15 to disallow using AI detector results as sole justification is a symptom of the same problem. —Cryptic 17:57, 10 July 2026 (UTC)

  What exactly makes the presence of ":contentReference[oaicite:16]{index=16}" "subjectively LLM-generated"? Are you suggesting that it is realistically possible that someone lovingly, handcraftedly typed that out without AI? Gnomingstuff (talk) 02:34, 11 July 2026 (UTC)

  Well, you just did.

  The entire point of the criterion is to remove subjectively LLM-written material. Something that blatant makes it both objective and subjective. But we don't want to delete a plainly human-written page that happens to meet the letter of one of these tells, or else we'd be speedying pages that have the wrong url pasted in or that have an author mistranscribed. —Cryptic 02:56, 11 July 2026 (UTC)

  (I actually copy-pasted it) Gnomingstuff (talk) 14:15, 11 July 2026 (UTC)

• B on account of the significantly increased barrier to the infection. -The Gnome (talk) 18:00, 10 July 2026 (UTC)
• Option D: Technical indicators + "LLM generated pages - This applies to any page (including previous revisions that are substantially similar) that exhibits any of the following signs which indicate that the page could only plausibly have been generated by large language models (LLM) and which unambiguously violates all of Wikipedia's core policies including neutral-point-of-view, verifiability, and no-original-research". I do not want to encourage indiscriminate application as that can also be disruptive especially if the LLM content is high quality, heavily copyedited, and thoroughly checked against Wikipedia's core policies. We are building an encyclopedia first, and if an LLM's output is heavily factual and encyclopedic and perhaps a great starting point, we should not punish new and/or experienced editors for using it as a foundation. On the other hand most LLM output is WP:TNT material. Aasim (話す（はなす）) 18:05, 10 July 2026 (UTC)

  especially if the LLM content is high quality, heavily copyedited, and thoroughly checked against Wikipedia's core policies

  Except that our guidelines prohibit this. Why do so many people not know what our guidelines are? Gnomingstuff (talk) 02:36, 11 July 2026 (UTC)

  WP:IAR and WP:POINTy exist as well. If one tags high quality material satisfying core policies for deletion, it is disruptive to delete it, regardless of whatever policies or guidelines say. We are building an encyclopedia, not policing productive AI use. Aasim (話す（はなす）) 04:10, 11 July 2026 (UTC)

  As another example, one can use Copilot, Codex, Claude, etc. to write a template, module, or user script to accomplish a needed task. It seems to work quite well, Dave Plummer for example on Dave's Garage used Copilot to create a Notepad clone. Aasim (話す（はなす）) 04:46, 11 July 2026 (UTC)

  You should have weighed in, then, when the community was making the guidelines. You didn't. Too bad. I do not know why people insist on coming out of the woodwork trying to undermine a policy that strong consensus was developed for without their involvement. Gnomingstuff (talk) 19:01, 11 July 2026 (UTC)

  You should have weighed in, then, when the community was making the guidelines. WP:NOTBURO and WP:5P#5. And as I said, almost all LLM content is AI slop that can be TNT'd. I concur with Pppery that using LLMs to write poorly generated articles is a conduct issue which is why WP:LLM advises against it.

  The other problem I have yet to discuss is how can one be certain that a piece of text is AI-generated or human generated? Especially because one can write a program which removes all of the obvious telltale signs that a piece of text is AI generated. This is a question of enforcement. CSD also needs to be objective, meaning that it includes only the pages that are problematic, and no more. The last time I successfully suggested an expansion to the CSD was with a WP:G5 expansion to include enforcement of general sanctions. Aasim (話す（はなす）) 21:53, 11 July 2026 (UTC)

  Open contempt for community consensus does not improve Wikipedia. * Pppery * _{in solidarity} 23:01, 11 July 2026 (UTC)

  Dismissing reasoned comments that are simply not unambiguously toeing the party line as "open contempt" actively harms Wikipedia, even more so when that criticism is founded in policies and guidelines. Wikipedia is not a one-party state where dissent is not tolerated, even if some would prefer to denigrate (or worse) anyone who prefers to act cautiously and with consideration for minimising collateral damage when it comes to matters like AI or archive.today. Thryduulf (talk) 23:12, 11 July 2026 (UTC)

  We are building an encyclopaedia. We have decided that chatbots and generated text are not welcome in the encyclopaedia that we are building. Any claims or comments about the "quality" of some hypothetical text that may possibly exist if someone just uses some unspecific tool in the right way is immaterial and irrelevant. The guidelines prohibit this.

  They also prohibit promotional material, doesn't matter how well it's written.

  If you want an LLM generated encyclopaedia, that already exists. That's not what this project is. ‑‑gurkubondinn 19:18, 11 July 2026 (UTC)

  Exactly. I'm going to state this bluntly. If you can't write a Wikipedia article without an LLM, you have no business writing a Wikipedia article at all. Competency is required. That includes the ability to write. I'd rather cleanup a good-faith but weakly written contribution than have to hunt down LLM hallucinations adding false info (sometimes the complete opposite of the facts!) and non-existent "sources". oknazevad (talk) 21:41, 14 July 2026 (UTC)

  Amen. During the early LLM ban debates, there was a lot of griping about "what if a person can't write in encyclopedic English but wants to contribute anyway?" I'm not sure why it's such a sticking point to suggest that proficiency in writing English should be a prerequisite to taking up contributing to an English-language encyclopedia as your hobby of choice. It's not a hobby that suits everyone, and that's fine. One wouldn't go to a cycling club and insist they let them use a motorcycle because they aren't proficient at riding conventional bicycles, nor would one expect such a club to make every effort to accommodate someone unwilling or unable to cycle. Athanelar (talk) 07:20, 15 July 2026 (UTC)

• B, as per most other people. The additions are good, and are very common to see in AfC/NPP. EatingCarBatteries _{(contribs | talk)} 21:27, 10 July 2026 (UTC)
• B, as per above, but also do want to say that I like the clarification on dead URLs a lot. Gnomingstuff (talk) 02:32, 11 July 2026 (UTC)
• B. I support the overall proposal because more and broader tools to remove AI content is always good, we shouldn't be spending longer on the deletion process than was spent to excrete the content in the first place. I support B specifically due to the reason given by 15224 et al; our guidelines on AI use no longer make any allowance for "human review," G15 shouldn't be any different. G11 isn't a narrower standard than WP:NOPROMO, it just requires unambiguity. Same goes here; G15 should follow NOLLM. As for thr criticism given by Cryptic, I think it is generally a bad idea to design rules for Wikipedia based on hypothetically imagining the worst possible way a person might interpret and implement them. WP:AGF also applies here; we have to assume that people are going to engage with our systems in good faith; if they don't, that's a conduct issue to deal with in itself. Athanelar (talk) 04:39, 11 July 2026 (UTC)
• B. Good rewrite. We need to recognize, as Cryptic pointed out near the top, that technical indicators are very reliable evidence of LLM use, and that their removal alone does not resolve the risks inherent to LLM-generated content. Toadspike [Talk] 20:37, 11 July 2026 (UTC)
• B, as per above. FaviFake (talk) 23:27, 11 July 2026 (UTC)
• B, as per above. Particular appreciation to those that identified the issue in the current status quo and suggested the proposed changes. — ♠ Ixtal ^{( T / C )} ⁂ Non nobis solum ♠ 14:25, 12 July 2026 (UTC)
• B. Indicators such as oaicite are now commonly used as a basis for G15, even if they aren't quite "communication intended for the user", and it would help if those tells were given their own sub-criterion. However, it would be useful to clarify that general markdown is not included, since the distinction with "markdown code blocks" could confuse new users. Somepinkdude (talk | contribs), in solidarity 15:46, 12 July 2026 (UTC)
• C, I'm personally concerned about the clause (including previous revisions that are substantially similar) since the rest of the criteria does not define how this particular clause will be applied and I potential for misinterpretation especially in the context of rewriting LLM articles to be human-written. More broadly, minus that specific claude, I find myself between A and B. The one thing I keep coming back to when looking this criteria is "unambigousness" and whether it can be red-teamed against abuse. To me, even if we don't enshrine "reasonable human review" in our global LLM policy anymore, it is what makes these pages identifiable, unacceptable and unambiguous in the first place. If we were generically looking for "AI-gen" pages, almost everything with a em-dash or a curly quote or vague language would be up for deletion. We clearly don't, (and I would oppose us expanding the criteria in that direction) which is what makes this criteria accurate to current practices even with the "reasonable human review" clause. As a result, I find myself leaning towards A (minus the first sentence) in keeping the CSD criteria header the same going forward. -- Sohom (talk) 16:35, 12 July 2026 (UTC)

  I at least would interpret that clause as applying the same body of rules we already apply in G5 cases when assessing whether non-socks have made substantive edits. But I agree the wording is kind of a bit unclear. * Pppery * _{in solidarity} 17:27, 12 July 2026 (UTC)

  I would be open to a WP:G5 like interpretation requiring substantial edits in between! Sohom (talk) 18:52, 12 July 2026 (UTC)

• B, per various comments above. AI is becoming more and more prevalent and with the WP:AFC queue becoming over 5000 for the first time ever we need to make it easier to delete slop that gets posted and wastes time. CoconutOctopus talk 14:10, 14 July 2026 (UTC)
• B, per above. In line with WP:NEWLLM, and AFC and NPP need the additional tools to combat the flood of AI articles. 🌸⁠wasianpower⁠🌸 (talk • contribs) 15:14, 14 July 2026 (UTC)
• B: I think the need to be able to act decisively outweighs any benefit from keeping the human-review reference in the criterion. Simply removing the tells is not enough, as pointed out by 15224 above, we need to ensure the underlying issues are fixed. The current wording also has a two-paragraph EFN plastered on it explaining what reasonable human review is in the first place, and then it trips over itself to remind the reader that LLM-generated content is not allowed regardless of review, so removing that is exactly the sort of simplification the policy can benefit from. ASUKITE 18:43, 14 July 2026 (UTC)
• B. I'd say that is the best approach in regards to anything written by an LLM, and is consistent with Wikipedia's guidelines. User:EdiotRando
• B I like the proposed version more as it is more specific and thorough. jolielover♥talk 14:22, 16 July 2026 (UTC)
• B. G15 came in rather early when you look at Wikipedia’s LLM-related guidelines, and the proposed revision brings it more in line with the current state of play. —pythoncoder (talk | contribs) 12:50, 17 July 2026 (UTC)
• B LLMs have no place on Wikipedia and updating to this will help enforce that. in solidarity, Chloe (talk) 02:00, 20 July 2026 (UTC)

  "LLMs have no place on Wikipedia" and similarly blunt arguments are very poor rationales for updating a speedy deletion criterion (or for making any LLM-related policy change, in fact) as, without any explanation of why LLM output is generally incompatible with Wikipedia, they are effectively equivalent to "I personally dislike this technology, so I think it should be restricted or banned." The exact same arguments can be used to support a policy like The use of LLMs is banned on the English Wikipedia. No exceptions. SuperPianoMan9167 (talk) 01:48, 2 August 2026 (UTC)

• B. The counterarguments above are not persuasive, IMO. We could invent endless hypotheticals that have no bearing on reality ("What if there's a bomb that goes off when this blatantly LLM-slop Wikipedia article is deleted?! checkmate atheists"). Stepwise Continuous Dysfunction (talk) 22:06, 22 July 2026 (UTC)
• Option E: Use option A, except remove the clause "including previous revisions that are substantially similar" because it appears to be either redundant or actively harmful. OAICITE and similar reference markup bugs (except UTM parameters, because search engines may occasionally index URLs with these parameters, and an editor may write article content themselves based on sources obtained via an LLM, which is actually recommended as an acceptable alternative to copy-pasting LLM output) are even more obviously inappropriate for inclusion than communication intended for the user, and are among the most objective signs of AI writing possible. I feel wary of removing "reasonable human review" because of the very real possibility that it will encourage overzealous speedy deletion tagging, as despite the instruction to avoid relying on more subjective signs, many G15 tags do so anyway and treat the criterion as a catch-all "delete because AI". "Including previous revisions that are substantially similar" seems to be unnecessary because the speedy deletion policy already states that A page is eligible for speedy deletion only if all of its history is also eligible. and Use common sense when applying a speedy deletion request to a page: review the page history to make sure that all earlier revisions of the page meet the speedy deletion criterion, because a single editor can replace an article with material that appears to cause the page to meet one or more of the criteria. (emphasis added) On the other hand, if the intent is to allow for the speedy deletion of any page with at least one revision consisting solely of unambiguously LLM-generated content (i.e. if any previous revision was eligible for speedy deletion, the page is eligible for speedy deletion regardless of its current revision) then I oppose the clause even more, because that's, plain and simply, not how speedy deletion is supposed to work. If a spammy article is rewritten to remove all the promotional content, and it becomes a well-written article, should it be deleted via G11? No. If an LLM-generated article is overwritten with a human-written stub, should the stub be deleted via G15? Also no. Of course, someone cannot save a page from speedy deletion by simply removing the LLM signs, but we should still allow for the possibility that a well-written human page can overwrite a G15-eligible page. SuperPianoMan9167 (talk) 01:32, 2 August 2026 (UTC)

  I think the whole idea of G15 is similar to G11 spam. Yes spam can be rewritten so it is not promotional, but no we should not reward the creator of spam by keeping a non-notable spam topic on Wikipedia. Aasim (話す（はなす）) 03:42, 2 August 2026 (UTC)

  Whether content about a topic was written by an LLM is completely independent of how notable that topic is or isn't. An LLM-written article about my sister's cat could not be rewritten to be a notable article. An LLM-written article about Highveld Alluvial Vegetation (one of our most wanted articles) could be rewritten by a human and become a good article. We should not penalise the human who improves the encyclopaedia just to spite the LLM-user. Thryduulf (talk) 10:18, 2 August 2026 (UTC)

  If someone did generate that article with an LLM, and then someone else rewrote it, and someone then nominated it for G15 after it was rewritten, then the G15 would be declined. But if the sequence of events is different and it is nominated for G15 while it is still just LLM-generated, then it should be deleted (assuming that it meets the specific G15 criteria). Rewriting it presumably takes longer than it takes to process a G15 request, but if someone starts rewriting it in the mean time then draftification seems like a reasonable option. ‑‑gurkubondinn 10:24, 2 August 2026 (UTC)

• B, which enables reviewers on the AI noticeboard to remove LLM-generated content more efficiently. We see these problems on a regular basis and it would be helpful to have them explicitly stated in the policy. There have been too many cases of editors who misuse LLMs trying to conceal their LLM misuse by removing the most obvious AI signs without spending any effort to revise the remainder of the affected articles, and removing "and would have been removed by any reasonable human review" brings the amount of effort needed to remove LLM-generated content a bit closer to the very low amount of effort needed to generate it. Frankly, our policies concerning LLM misuse are still not strong enough, but these changes are a step in the right direction. — Newslinger talk 00:30, 12 August 2026 (UTC)

Discussion (G15)

• This would have benefitted from a pointer to the rfcbefore discussion from here. In particular, most of the changes are entirely uncontroversial and could have been severed from the RFC. —Cryptic 00:35, 3 July 2026 (UTC)

  Good idea--I've directly edited in the uncontroversial changes into G15 in this revision. Ca ^{talk to me!} 14:13, 10 July 2026 (UTC)

• Pinging Wikiproject AI cleanup discussion participants: @Chaotic Enby @Gurkubondinn @EatingCarBatteries @Thryduulf @Toadspike Ca ^{talk to me!} 16:25, 10 July 2026 (UTC)
• Comment, largely in favour of this but I'm not a fan of calling it "unambiguous" since that implies other articles determined to be LLM-generated but not eligible for G15 are ambiguous. "indisputable"/"incontrovertible" or "obvious" might be better? Kowal2701 (talk, contribs) 16:49, 10 July 2026 (UTC)
• I'm leery of the specific wording "This may include" in the new "Technical indicators" subcriterion. Such wiggle room has been abused in the past, and I don't doubt it would be here too. ("This is plainly llm generated! Using the rule of three is a technical indicator of ai slop! How dare you remove my db-g15 tag!") I'd prefer limiting it to the model-specific citation bugs and the wiki(text) markdown subset, rather than using those as examples. —Cryptic 18:40, 10 July 2026 (UTC)

  Agreed. * Pppery * _{in solidarity} 18:43, 10 July 2026 (UTC)

  Whether or not this RFC passes, While these indicators can be used in conjunction, they should not serve as the sole basis for applying this criterion. should be changed to Those subjective signs cannot serve as the basis for applying this criterion. — rsjaffe 🗣️ 19:53, 10 July 2026 (UTC)

  This is plainly llm generated! Using the rule of three is a technical indicator of ai slop! How dare you remove my db-g15 tag I have often seen this hypothetical overeager AI cop brought up in LLM policy discussions as a cautionary tale, but I am somehow yet to see anything resembling this argument happen in actual AI cleanup cases. Athanelar (talk) 04:45, 11 July 2026 (UTC)

  I mean... I saw one of the admins active in G15 tagging delete a one line stub with exactly 0 promotional language as G11. So, like.... worries that whatever criteria we decide here is going to be stretched to its breaking point isn't that unfounded? GreenLipstickLesbian💌🧸 05:59, 11 July 2026 (UTC)

  If an admin improperly applies speedy deletion criteria, then we should challenge the deletion and, if necessary, issue reprimands and revocation of tools against said admin. I'm all in favour of making sure we use clear wording without unnecessary ambiguity, but we're not a dictatorship, and we don't need to worry that some admin might maliciously apply the criteria. If the criteria are being applied in a way obviously contrary to consensus, people can and will challenge that. Athanelar (talk) 06:08, 11 July 2026 (UTC)

  And I do agree with you, @Athanelar, that challenging incorrect deletions/removing sysop from admins who repeatedly abuse/misuse speedy deletion criteria is, ultimately, the way forward. However, I do disagree that people can and will challenge that, especially in a timely manner. And I also know that not everybody agrees with us; U5 was repealed rather recently because, although its intentions were good and it could be a useful tool, it was openly misused for years to delete... pretty much any arbitrary newbie's userpage? Especially if said userpage revealed the newbie to be from a non-Anglophone, non-Western country. So I also don't agree that we don't need to worry that some admin might maliciously apply the criteria (or, not maliciously, just... blatantly incorrectly). Dealing with abuse by good-faith actors is hard at the best of times, and so it's not unreasonable for somebody to keep the criteria as narrow and as hard to stretch as possible. GreenLipstickLesbian💌🧸 19:22, 11 July 2026 (UTC)

  I can't speak for G15s -- I suspect it's going to be hard to get any data on this given that, you know, things get deleted -- but after going through the AI tags, I only saw maybe a few dozen that I felt the reasoning was weak on, and they were almost always by the same person. (nobody here and not someone active in AI discussions) Gnomingstuff (talk) 14:19, 11 July 2026 (UTC)

• For fun I got ChatGPT to generate an encyclopedia article about Wikipedia and here is what it came up with:

# Wikipedia

**Wikipedia** is a free, multilingual online encyclopedia written and maintained by a global community of volunteer contributors. It is operated by the Wikimedia Foundation, a nonprofit organization, and is one of the world’s most widely used reference websites. The name “Wikipedia” combines *wiki*, a term for collaboratively editable websites, with *encyclopedia*.

## History

Wikipedia was launched on January 15, 2001, by Jimmy Wales and Larry Sanger. It initially developed as a complementary project to Nupedia, an earlier online encyclopedia that relied on a formal expert-review process. Wikipedia’s open editing model allowed articles to be created and revised much more rapidly, and the project soon surpassed Nupedia in size and activity.

Over time, Wikipedia expanded into hundreds of language editions. Its growth also contributed to the development of other collaborative knowledge projects within the broader Wikimedia movement.

## Content and editing

Most Wikipedia articles can be edited by members of the public. Contributors add information, correct errors, reorganize material, and discuss disagreements about article content. Editors are generally expected to follow principles concerning verifiability, reliable sourcing, a neutral point of view, and restrictions on original research.

Wikipedia does not employ a traditional centralized editorial board to approve every article before publication. Instead, content is governed through community-developed policies, discussion, consensus-building, and various forms of review and moderation. Some pages may be temporarily or permanently protected from unrestricted editing when they are frequent targets of vandalism or disruptive activity.

## Organization and governance

Wikipedia is hosted by the Wikimedia Foundation, which provides technical infrastructure and institutional support. The encyclopedia’s day-to-day content decisions, however, are largely made by volunteer communities within individual language editions. Editors may take on additional responsibilities, such as administrative or technical roles, according to community procedures.

Wikipedia is financed primarily through donations and does not generally display conventional commercial advertising.

## Reliability and criticism

The reliability of Wikipedia varies among articles and subjects. Because articles can be edited continually, errors may be corrected quickly, but inaccurate, incomplete, or biased information can also be introduced. The quality of an article often depends on factors such as the availability of reliable sources, the level of editorial attention it receives, and the maturity of the article’s development.

Wikipedia has been criticized for systemic bias, uneven coverage, editor disputes, vandalism, and the difficulty of representing controversial subjects neutrally. Supporters emphasize its broad accessibility, extensive scope, transparent revision histories, and capacity for rapid collaborative improvement.

For academic and professional research, Wikipedia is often used as a starting point rather than as a sole or final authority. Its citations and reference sections can help readers locate primary documents and other published sources.

## Cultural impact

Wikipedia has significantly influenced how people seek and share general knowledge online. Its collaborative model has become a prominent example of large-scale volunteer knowledge production. Articles from Wikipedia are also used in search engines, educational contexts, digital assistants, and numerous other information systems.

The project’s open-content model permits much of its text to be reused and adapted under applicable free licenses, provided that licensing requirements are followed.

## See also

* Online encyclopedia
* Wiki
* Wikimedia Foundation
* Open content
* Collaborative knowledge
* Free knowledge movement

• Yeah all alone it is unlikely to satisfy Wikipedia's core policies. My view above still hasn't changed that if one uses AI as a start and they rigorously add citations and content to comply with Wikipedia's core guidelines it may be acceptable. But again, just standing by itself, WP:TNT. Aasim (話す（はなす）) 01:07, 12 July 2026 (UTC)
• If we are broadening this category, which it looks like will happen, and which is probably a good thing, I think we need to put some more thought into AI detection false positives. I've run into a couple of pages tagged/deleted as AI that was the consequence of using only one AI detector, or not being super knowledgeable about AI, etc. Deleting AI is good. Tagging real people as AI is not and decreases our retention rate. Short of custom developed AI detection tools, how can we improve our accuracy rate and cut down on false positives? CaptainEek ^{Edits Ho Cap'n!}⚓ 01:52, 13 July 2026 (UTC)

  Do you have specific examples of this? Gnomingstuff (talk) 06:38, 13 July 2026 (UTC)

• I feel like nonexistent categories and templates should be a G15 criterion. These are incredibly common and easy to identify, and rarely encountered in human-generated drafts. I recently encountered an instance at Udyot Chand. –LaundryPizza03 (dc̄) 22:31, 19 July 2026 (UTC)

Q2. Exempt code pages, templates, modules, etc. from G15?

As I've raised above, LLMs seem to be quite good at generating working code snippets and thus can be useful for writing templates, modules, and user scripts (in particular Codex, Claude, Copilot, etc.). Should LLM-generated templates, modules, user scripts, etc. be exempt from G15? Aasim (話す（はなす）) 18:39, 12 July 2026 (UTC)

• Support, as one who has previously used LLMs to generate code snippets for and fixing modules. G2 and other similar criteria may apply. Aasim (話す（はなす）) 18:39, 12 July 2026 (UTC)
• Support, Was this ever up for debate even in the context of the original criteria? My understanding for a while has been that the criteria does not apply to AI generated code. For what it's worth, my personal threshold/compass for AI-gen code is very different from AI-gen articles/content. I'm personally fine with folks AI-gening their own userscripts or even their own Toolforge tools for personal use. However, I personally follow (and encourage following) the "own the code you commit" principle when contributing code to templates/gadgets that will need to be maintained by other folks (i.e. regardless of whether you used AI tools, you should be intimately familiar with the code, should have rigorously dry run/tested the code to the point where it is indistinguishable from code you have written -- and this is very different from my stance of AI-gen in articles at the moment). -- Sohom (talk) 19:09, 12 July 2026 (UTC)
• Support in principle pending exact wording. The primary reasons G15 was supported do not apply to template code, modules, and similar pages. G1-G12 will remain available to speedily delete these pages if required, and TfD will also remain an option for anything that would be better deleted but doesn't fit a speedy criterion. Regarding the exact wording I suggest something like: This criterion does not apply to templates or modules, nor to other pages consisting exclusively or primarily of computer code (including python, javascript, css and lua but excluding wikitext and HTML) and associated comments. I expect that to need wordsmithing though. Thryduulf (talk) 19:17, 12 July 2026 (UTC)
• Support Should be a common sense exception, but it would be harmless to codify it. One of the userscripts (User:Ca/RSNsearch.js) I use was generated by ChatGPT all the way back in mid-2023. Ca ^{talk to me!} 12:14, 13 July 2026 (UTC)
• They already don't qualify, unless the generated code is so horrendously malformed that it has (nonsensical) references or placeholders that would make it not work anyway. G15 is not defined by the title "LLM-generated pages without human review", which is why the title needs to be accurate, and not made more inaccurate by the proposal above. —Cryptic 14:44, 13 July 2026 (UTC)
• Support The graph on my userpage is coded by ChatGPT as I don't know how the language works. Maybe adding what Thryduulf said Warm Regards, Miminity (Talk?) (me contribs) 00:07, 14 July 2026 (UTC)
• Oppose as we still want human review of code produced this way. Also G15 has strict requirements which a piece of code is unlikely to meet. So as Cryptic said code would not be deleted under this criterion. Graeme Bartlett (talk) 07:48, 14 July 2026 (UTC)
• Oppose. We do want humans to review and test any LLM-assisted modules and templates they create, which means that any reasonable LLM use in coding templates is already exempt from G15. —Kusma (talk) 07:57, 14 July 2026 (UTC)
• Oppose. If a template or module contains "communication intended for the user" or "non-existent or nonsensical references", then it should absolutely be eligible. It would make sense to exempt them only if the requirements were vague or broad, which they are not. FaviFake (talk) 08:42, 14 July 2026 (UTC)

  @FaviFake: Excessive comments containing "communication intended for the user" in Lua can typically be left in even after human review without harming the integrity of the program though. Sohom (talk) 09:30, 14 July 2026 (UTC)

  Looking at User:Ca/RSNsearch.js it contains the text: Added relative position for #rsn-content which is communication intended for the user, which would make it eligible under G15 even though I'm fairly certain that Ca has tested the script. Sohom (talk) 09:33, 14 July 2026 (UTC)

  No it wouldn't. The G15 criterion only applies to any page that exhibits one or more of the following signs which indicate that the page could only plausibly have been generated by large language models (LLM) and would have been removed by any reasonable human review (emphasis supplied)

  A reasonable human reviewing that line could not possibly think that // Added relative position for #rsn-content could only plausibly have been generated by a LLM. FaviFake (talk) 10:58, 14 July 2026 (UTC)

  @FaviFake: I don't know how to explain this, but there is no reason for a human programmer to write that comment at that line. That line of code is self-documenting. A programmer (even somebody without that is a beginner) would either put a comment in line 23 to explain at a high level why the styles were necessary or that the subsequent block would add styles or put no comment at all. Commenting specifically that lines makes no sense and would to me immediately make me suspect LLM use. Sohom (talk) 11:56, 14 July 2026 (UTC)

  It's the use of past tense, a human programmer would have written something like "set relative position for ..." or just "relative position for/of ..." using the present tense (and indicative verbal mood) instead. It's still a pointless comment that most programmers wouldn't leave. ‑‑gurkubondinn 12:13, 14 July 2026 (UTC)

  I agree it is not necessary. But could that comment only plausibly have been generated by a LLM? Of course not. I could write that comment myself. Maybe I'd want to highlight something I changed in my code so that i can quickly jump to it and revisit it later, or something else. The point is that you can't be sure it was 100% AI-generated, and I don't even think it would qualify as "Communication intended for the user" in the first place:

  This may include collaborative communication (e.g., "Here is your Wikipedia article on..."), knowledge-cutoff disclaimers (e.g., "Up to my last training update ..."), self-insertion (e.g., "as a large language model"), and unfilled placeholders (e.g., "Smith was born on [Birth Date]" rather than the person's actual birth date).

  I couldn't write any of these myself, they'd have to be generated by an AI model. But I could, plausibly, write // Added relative position for #rsn-content. FaviFake (talk) 12:14, 14 July 2026 (UTC)

  Are you talking about stuff like {{#if:YOUR CONDITION HERE|PLACE YOUR TEMPLATE CODE HERE}}? I would apply G2 to this kind of stuff, it shouldn't be eligible for G15. Aasim (話す（はなす）) 18:47, 14 July 2026 (UTC)

• Oppose, user scripts can cause a lot of damage. If someone prompts an LLM to generate a script that they don't have the knowledge or ability to understand themselves, the risk of it doing damage is even higher. Besides, you can't edit .js files in someone else's user space (I think that only interface admins can do this?) so how would you tag it with {{db-g15}} in the first place (but you can probably CSD the files in some other way though)? Also this is pointless policy cruft in my opinion, nobody was talking about this, and it's not really relevant to the problem at hand. Let's not make the deletion criteria more complicated than it needs to be. ‑‑gurkubondinn 11:10, 14 July 2026 (UTC)

  I think that only interface admins can do this?) so how would you tag it with {{db-g15}} in the first place - Folks can put a {{IPER}} template on a talk page requesting the deletion of a page. I remember actioning/responding to one deletion request done in this manner for user css pages. Sohom (talk) 12:02, 14 July 2026 (UTC)

  Last time I checked, I as an ordinary admin could delete .js files, but not undelete them. A tag on the talk page plus a note explaining what the admin should do will do the trick. —Kusma (talk) 18:32, 14 July 2026 (UTC)

  Right, I was thinking about who can edit .js files. ‑‑gurkubondinn 18:38, 14 July 2026 (UTC)

• Clarification: This is largely applicable if the text and would have been removed by any reasonable human review is removed from the speedy deletion policy. However, since there can be comments which are intended for collaboration (although not necessarily LLM generated, for example // @TODO: Insert your name here), G15 not applying to those pages would let users experiment with user scripts prior to listing on WP:USL. Aasim (話す（はなす）) 15:23, 14 July 2026 (UTC)

  Just to reiterate, the fact that there is communication intended for the user does not mean that the requirement that the page could only plausibly have been generated by LLMs doesn't have to be satisfied. That comment alone cannot plausibly only be generated by an artificial intelligence model. FaviFake (talk) 15:34, 14 July 2026 (UTC)

• Oppose Many of the same fundamental problems with LLM-generated article text are present in LLM-generated code. —pythoncoder (talk | contribs) 12:53, 17 July 2026 (UTC)
• Oppose per above. Unchecked code can cause catastrophic damage, especially when developed by an LLM. It's got the exact same issues with code as it does text, and exempting it is only going to cause problems down the line. in solidarity, Chloe (talk) 02:03, 20 July 2026 (UTC)

  If you could give an example? Because poorly-written human code can also do much the same. Speaking from experiencing countless Lua errors and whatnot because of typos, undeclared variables, etc. Aasim (話す（はなす）) 04:30, 20 July 2026 (UTC)

  This is entirely offwiki, but it's the first thing I thought of when I read this proposal. In my first semester of uni last year, I was in an introductory engineering class and we had group projects. Someone in another group had the idea to get an LLM to write code for some database thing for their group's project (I never learned the specifics for it unfortunately) and the first time they tested it something went wrong with the code and corrupted pretty much all the data they had collected over the previous week, which, in their words, was a "metric crapton". I've been wary of LLM-code ever since, because unless you absolutely know what you're doing, it can do something very unexpected. I'm worried that an entirely LLM-generated userscript or bot or something could do something that takes a headache and a half to clean up, and I think if editors are required to review their code before deploying it, that's much less likely to happen. To be frank, I think there should be some sort of user right requirement to make these things, but that's a different discussion. in solidarity, Chloe (talk) 04:52, 20 July 2026 (UTC)

  We already have policies re: WP:MEATBOT or the more expansive WP:BOTPOL to deal with those specific cases. Sohom (talk) 05:04, 20 July 2026 (UTC)

  A human not actually knowing what they are doing can accidentally blow up a database as well, even before LLMs. And then, back in March a malicious user script mass deleted pages on Meta-Wiki. I doubt that script uses an LLM, but it probably would qualify for WP:CSD#G3 were it on English Wikipedia. Aasim (話す（はなす）) 14:31, 20 July 2026 (UTC)

  And LLMs makes it easier for a human who doesn't know what they're doing to blow up a database. jolielover♥talk 14:32, 20 July 2026 (UTC)

  But a human who knows what they are doing is unlikely to blow up a database, even with an LLM. Aasim (話す（はなす）) 14:36, 20 July 2026 (UTC)

• Oppose AI-generated code is a bigger danger than AI-generated articles. We want humans who know what they're doing to be making code and doing a thorough review of any LLM output, lest there be any horrible repercussions. Why should they be exempt?? jolielover♥talk 14:38, 20 July 2026 (UTC)
• Oppose If you're not willing to write the code yourself, you shouldn't be contributing it to a volunteer-built encyclopedia. Stepwise Continuous Dysfunction (talk) 22:07, 22 July 2026 (UTC)
• Oppose I remember an ANI where a user sent out over 900 notifications for an AfD, including to nonexistent users, using an LLM-generated script that blanked existing content on the user talk pages. Unfortunately, I was not able to find it in search, but I also remember that the closure was headed with something like "Using an AI-generated script didn't go so well". –LaundryPizza03 (dc̄) 04:46, 25 July 2026 (UTC)

  It was Wikipedia:Administrators'_noticeboard/IncidentArchive1211#TattooedLeprechaun. –LaundryPizza03 (dc̄) 04:58, 25 July 2026 (UTC)

  Still, the existing G15 is not fully applicable to computer code, since computer code doesn't have references. I have an idea what "communication intended for the user" would look like: messages that ask the programmer to add their own code or make manual edits in specific places. –LaundryPizza03 (dc̄) 05:00, 25 July 2026 (UTC)

  I don't know if that was through the MW API or if that was using some script clicking buttons on behalf of the user. But in that case the user did not seem to know what they were doing or even how their code worked.

  Also, what is to stop someone from using AI code in the F12 developer console? Or with certain browser extensions one can inject their own scripts into any website. Since that is all stored locally on the machine. Publishing at least gives a chance for someone more competent to review the code and perhaps fix it up. Aasim (話す（はなす）) 16:08, 25 July 2026 (UTC)

  I personally doubt disallowing that kind of code onwiki is going to help with mitigating that specific incident. Bot code is rarely if ever uploaded onwiki, it is run on a user's computer and at best released on Github neither of which the community has jurisdiction over. Sohom (talk) 20:41, 25 July 2026 (UTC)

  if ever – User:AnomieBOT/source; there are others. Also, and more relevantly to the discussion here, user scripts are typically saved on-wiki. Anomie⚔ 00:41, 26 July 2026 (UTC)

  Fair, and I think leeky and a few others do to, and yes userscripts are onwiki, but I think my point that a ban on putting them onwiki doesn't translate to a ban on using them period still kinda stands, scripts can be hosted elsewhere or be transpiled onwiki, (or even just loaded in the browser by copy and pasting), bots are run on Toolforge or on other machines. Sohom (talk) 03:03, 26 July 2026 (UTC)

  BadRfC Question: (with apologies) Seems like both support and oppose comments are indicating that good LLM written code can stay. I respectfully ask this question be withdrawn by the originator so we can do a follow up if needed. Does support mean slop can stay? Does oppose mean even perfect code must be removed if the author admits to LLM usage? I’ll try to reread and change my vote later but would appreciate confirmation from anyone else seeing the same issue Dw31415 (talk) 17:13, 29 July 2026 (UTC)

  "Support" means "we should add an exception so that templates, modules, user scripts, etc cannot be deleted under G15". "Oppose" means "they should remain speedily deletable under G15". I don't see a problem with the question. FaviFake (talk) 10:57, 12 August 2026 (UTC)

• Status quo: per Cryptic, G15 already doesn't apply to most code, and per Thryduulf, G15 is the wrong tool to deal with the kinds of LLM code we want to get rid of. If we feel like we need to clarify that admins shouldn't be G15ing things like User:Ca/RSNsearch.js, then I guess we can do so, but I'd prefer not to policycruft or open the door for arguing about edge cases. -- LWG ^{talk (VOPOV)} 17:01, 29 July 2026 (UTC)

  Ah thanks, status quo would be a good !vote. I don’t always want to add cruft ;-) Dw31415 (talk) 17:15, 29 July 2026 (UTC)

  Just to clarify, G15 does apply to every type of page because there are no exclusions listed for [the] specific criteria. FaviFake (talk) 11:04, 12 August 2026 (UTC)

• Irrelevant per Cryptic. Nearly any code which clearly exhibits the signs mentioned just... wouldn't work. To do what many opposes seem to want amounts to a new criterion, but I don't see a reason that, in principle, someone using this for terribly malformed code would be a travesty. LittlePuppers (talk) 04:13, 6 August 2026 (UTC)
• Oppose, as templates can affect a large number of pages and deserve at least as much scrutiny as articles do. — Newslinger talk 00:32, 12 August 2026 (UTC)

  The issue is that G15 doesn't provide meaningful scrutiny for the types of pages being discussed here, because the criterion is exclusively focused on what is relevant to articles and drafts not what is relevant to templates and code. For example there is no scrutiny of whether the code works in any form, whether it matches the intent, whether it does unwanted things, etc. only whether if has communication intended for the user (because templates don't have or need references).

  If there is a need for scrutiny of those pages in the form of speedy deletion, that isn't already provided for by other criteria, then it should be a separate criterion that evaluates objective factors relevant that separates what should be deleted from what should not be. Thryduulf (talk) 11:15, 12 August 2026 (UTC)

  Even if the criterion only applies to a very small percentage of AI-generated template-related pages, which I agree with, why should those pages be exempt from it as well? Why should we not speedily delete templates that could only plausibly have been generated by LLMs, even if they are very few? Everyone is saying that these criteria were made with articles in mind, but if they're even marginally useful for deleting templates, why should they be restricted to articles?

  As i have explained in my conversation with Sohom above, G15 cannot possibly produce false positives in the way that it is currently written. FaviFake (talk) 12:54, 12 August 2026 (UTC)

  The question is whether we want to speedily delete templates, modules, etc. that are LLM-generated just for being LLM-generated, and based on the comments here and in other discussions the general view seems to be somewhere between "no" and "only sometimes" (and the more code-like a page the more the tends towards not deleting them just for being LLM-generated) which means the criterion as written fails WP:NEWCSD point 2. Thryduulf (talk) 13:06, 12 August 2026 (UTC)

  The question is whether we want to speedily delete templates, modules, etc. that are LLM-generated just for being LLM-generated
  No, that's not the question at all. I would support excluding templates from G15 if that were the RfC question.

  The actual question is whether we want to speedily delete templates that exhibit one of the signs below which indicate that the page could only plausibly have been generated by LLMs

• collaborative communication (e.g., "Here is your Wikipedia template on..."), knowledge-cutoff disclaimers (e.g., "Up to my last training update ..."), self-insertion (e.g., "as a large language model")...
• technical indicators like ```wikitext markdown code blocks, etc.

If a template is completely AI-generated, the author themselves said they used AI and they didn't write any code themselves, but there are no obvious signs, then it cannot be deleted under this criterion, even if Option B passes.

I cannot stress this enough. As Graeme Bartlett said above, G15 has strict requirements which a piece of code is unlikely to meet. If a template has been reviewed and there are no signs, then it can't be G15'd. If it has NOT been reviewed and there is text like "Here is your Wikipedia template on..." on the template, then of course we shouldn't keep that on the template namespace for anyone to use. The requirements are strict for a reason. FaviFake (talk) 13:23, 12 August 2026 (UTC)

And of course when G15 is misused it will be declined, right? Techie3 (talk) 16:40, 12 August 2026 (UTC)

if a page is wrongly deleted, the admin can be notified and admonished if it happens too frequently. If a user tags the page incorrectly, the admin will usually decline the request. FaviFake (talk) 19:16, 12 August 2026 (UTC)

Looks at U5 abuse Sohom (talk) 19:24, 12 August 2026 (UTC)