User talk:Jimbo Wales

	The Admin's Barnstar
	Thank you for all these years of services and bestowing us with this website Cheers -olam Olaf and elsa (talk) 05:22, 16 March 2026 (UTC)

I'm sure you've heard people say "everything is already written". Well, it's not. I just wrote the stroller article, which somehow did not exist in 2026. It's a stub, but it's much better than nothing. :) Clovermoss🍀 (talk) 01:39, 17 March 2026 (UTC)

Baby_transport#Wheeled_transport_methods, I like that, sounds very Wikipedian. Gråbergs Gråa Sång (talk) 05:23, 17 March 2026 (UTC)

Yes, that section exists, and it needs considerable work. I need to head to sleep but I hope to expand the stroller article soonish. There's a lot that can be said about various types, "bougie" strollers, etc. But I'm really surprised no one took such "low hanging fruit" before now, as I usually hear it described. I often find that a lot of articles we have that are somewhat parenting related tend to be in poor shape, though. There's always so much to do. Clovermoss🍀 (talk) 05:33, 17 March 2026 (UTC)

I'm pleased to notice that the Baby transport article mentions Battleship Potemkin. Gråbergs Gråa Sång (talk) 10:30, 17 March 2026 (UTC)

Baby transport#Strollers appears to be longer / more detailed. Might make sense to redirect Stroller to that. –Novem Linguae (talk) 07:13, 17 March 2026 (UTC)

Most of that detail is unreferenced text. I often think a well-referenced stub is better than that. I strongly object to the idea of redirecting it back, there's a lot that can be written about strollers and the fact that no one has tried to write a separate article about it before now is genuinely shocking to me. It's like if we didn't have an article for fork but had one on kitchen utensils. I think sometimes people kneejerk to merge genuinely notable subjects more than nessecary. There's several disadvantages to that. One is that our redirects only work internally and almost never show up in search engines. No one is looking up "baby transport" when they're thinking about buying a stroller, which is an almost universal parenting item, at least in North America. The second disadvantage is that it can make articles easily become unwieldy and requiring splits if someone does choose to try and give the subject the coverage it deserves. I don't think the average new editor is going to understand the nuances of how to request page deletion and do page splits etc. I've also seen people remove content that is "too detailed" in broader article subjects. The third is that just because something is short doesn't mean it's bad. I like short and sweet articles that are exactly what I'm looking for and I don't have to scroll down paragraphs of tangentially related content to find what I want, especially on my phone. Clovermoss🍀 (talk) 13:03, 17 March 2026 (UTC)

Having two spots for the same content isn't ideal. Results in duplicate work and duplicate maintenance effort. If I encountered this in the NPP queue, I'd probably delete the unsourced material in Baby transport#Strollers, replace it with the content from stroller, then WP:BLAR stroller to Baby transport#Strollers. There's nothing wrong with WP:SPINOUTs, and that should indeed be done once a section of a broad article gets big enough to warrant it. Not a big deal though. I'll leave things as is. –Novem Linguae (talk) 20:36, 17 March 2026 (UTC)

I think that might be a big deal if that's your standard approach. That seems somewhat bitey, especially since most new editors won't know how to object to such a change. At least drafified articles can still be easily resubmitted. I think it's one thing to be merging content that doesn't meet GNG and quite another to be imposing preferences on what articles should be. It's very confusing how differently people will approach this. I've seen multiple redirects be deleted at RfD where we have coverage of the subject in question precisely because people want to encourage article creation. I don't think it's fair to send mixed messages to people. Clovermoss🍀 (talk) 21:37, 17 March 2026 (UTC)

For what it's worth, I'm with you 100% on this one. Baby transport#Strollers sounds like a parody of Wikipedia rather than a real section in a real article. (Not the content, which I haven't read, just that very amusing location!). When I think about the evolution from Victorian prams down to modern Bugaboo strollers, I don't think "Baby transport -> Strollers". I think: Strollers.

We could in principle take almost any article in Wikipedia and merge it with some abstract concept, but that wouldn't make sense. We could redirect Spacecraft to Vehicle#Spacecraft. That's obviously silly. Jimbo Wales (talk) 07:50, 19 March 2026 (UTC)

If you know a bit about the evolution of strollers (I never heard the term Bugaboo before), maybe you could help out with the article itself? I've been trying my best to research, but I don't magically know everything about strollers simply because I'm a woman. I read your book so I know you're a parent, and presumably you bought a stroller at some point in time and might be aware of some content gaps. You also have more experience with the UK than I do so maybe you could help me figure out the differing terminology. It seems like some places use more specific words than others (such as pushcarts and baby carriages). Clovermoss🍀 (talk) 12:09, 24 March 2026 (UTC)

We could in principle take almost any article in Wikipedia and merge it with some abstract concept, but that wouldn't make sense. We could redirect Spacecraft to Vehicle#Spacecraft. That's obviously silly. I agree. Spacecraft has more content than a section of Vehicle could handle. My point was that merging can make sense when an article is stub-sized, and then a WP:SPINOUT can be performed later. Stroller is no longer a stub, so I would now no longer merge it. Now the question becomes more about EngVar, since I think British folks and Indian folks both have different words for stroller. –Novem Linguae (talk) 02:20, 25 March 2026 (UTC)

While the number of en.wiki articles increases every year and so it is clear there is lots left to write, we do our editors a disservice to equate 'article exits' with 'written'. Our overall article text has also steadily increased over time, which is a better proxy as "everything" includes writing about existing topics as well as writing new ones. Lots of stubs out there for example, which do actually have the potential to be significantly developed. CMD (talk) 07:55, 17 March 2026 (UTC)

Yep, there's so many articles that technically exist but are in poor shape. It's why I keep track of articles I improve on my content subpage as well. Some of the numerous examples I could give off the top of my head as subjects that need work are baby fever, baby jumpers, and growth landmarks. Back labor is a good short article but so much more could be said about that subject. I did a lot of work on the pregnancy article over the past year. It didn't use to explain that you don't just magically get pregnant (sex or assisted reproductive technology is required), that Rhesus disease exists, etc. And then even in a general sense, our backlogs are huge. There's literally thousands of completely unreferenced articles. Clovermoss🍀 (talk) 13:11, 17 March 2026 (UTC)

There may be some editors working on the unreferenced article category, it comes up every now and then. Not sure if it is tracked over time. CMD (talk) 14:05, 17 March 2026 (UTC)

Editors at the Wikipedia:WikiProject Unreferenced articles are working on it. There is a graph there for the evolution over time. Rolluik (talk) 13:59, 25 March 2026 (UTC)

Regarding your last point, according to a sample of Wikipedia articles I took, there is likely around 500,000 articles without inline citations. Some of these have non-inline citations, but I don't even think that is the majority of those 500,000 articles and even if it was, inline citations are just better than a non-inline citation except in a rare few cases. ✶Quxyz✶ (talk) 14:10, 17 March 2026 (UTC)

The reference desks have an oft cited essay WP:WHAAOE but see some exceptions: Draft:Euromissile Crisis recently which was a little surprising. FloridaArmy is constantly turning out 19th century biography and topics—some gems there whenever i look at their list. fiveby(zero) 14:26, 17 March 2026 (UTC)

"Only slight hyperbole!" Wikipedia has expanded nearly three-fold since that was written. ✶Quxyz✶ (talk) 14:30, 17 March 2026 (UTC)

I find I often have the opposite problem of being pleasantly surprised that things I want to read about aren't things I have to write. There's literally hundreds of articles that haven't been written yet on my mental to-do list. It's a lot of work trying to do that, improve articles on a neglected topic area, do all the reading I like to do for fun, try to improve WMF-community relations, be an admin, etc. Clovermoss🍀 (talk) 14:36, 17 March 2026 (UTC)

Please go write Eugene Pick, Tom Quick (Indian Slayer) and Tom Tiger for me will ya, i'm too lazy. Will get you lots of sources tho. There should be some kind of "Nonexistent" to "Featured" article project, would be fun finding the topics if someone else would actually write them. fiveby(zero) 14:52, 17 March 2026 (UTC)

There is: the Four Award. ~~ AirshipJungleman29 (talk) 15:12, 17 March 2026 (UTC)

We have the WP:26 for '26 project going on right now which shows quite ably how there is much left to create (and even expand). CMD (talk) 02:16, 18 March 2026 (UTC)

This is just something I am thinking about in terms of my longer term project thinking about NPOV issues in Wikipedia. I'm interested in other people's thoughts and observations.

Far more than I used to, I see people making efforts to silence/shut down debate in "rules-y" ways on talk pages. Two examples that I've seen recently: (1) the claim that on certain topics, certain editors are not (and importantly, should not) be allowed to speak - not because they have already violated our premise of "Assume Good Faith" but because they are new (2) accusations that a comment should be ignored if it shows some signs of having been edited/written using an LLM tool.

I think this is problematic. First, "assume good faith" means that a person of good will who sees a problem in Wikipedia ought to be able to come to Wikipedia and voice the issue and participate in a good faith effort to improve the article. That welcoming attitude to newcomers is a core wiki value, because it encourages to confront the possibility of internal "group think". There will be some rare cases where I will concede that this approach doesn't work - for example if there's a perennial issue where year after year we see people popping in with the same exact concern that we've dealt with over and over. But in any case where an article or subject is new, a blank ban on newcomer participation is toxic.

Second, LLM usage: many people, especially in English Wikipedia, may be joining the talk page in a language (English) in which they don't feel completely comfortable. That's ok. And so they might use an LLM to help them, and also to ask the LLM to make sure their comment is a valid argument on the grounds of Wikipedia policy. That is not a problem, but if someone's comment contains an emdash or something that makes us think that they may have used an LLM, I will argue quite vigorously that it is an ad hominem attack on the person to use it as justification for ignoring it. What matters is the content of the comment and the user's willingness to discuss it further in good faith.

Those are just two examples that are easy to identify but I think there are more. I think anytime anyone is about to make a post that doesn't address someone's points but instead argues that they should shut up, then a thoughtful and kind pause is advisable. And there may be policy implications here in terms of any policy creep that is damaging NPOV inadvertantly by allowing some voices to be silences. (One essay that is another common culprit is WP:MANDY which I generally see used to dismiss important viewpoints. Jimbo Wales (talk) 15:53, 23 March 2026 (UTC)

One problem is that LLM:s very often suck at making sure a comment is a valid argument on the grounds of Wikipedia policy. Current guidance at Wikipedia:Talk_page_guidelines#LLM-generated. But the whole WP:ECREXPLAIN thing is pretty new, I think, and the list has been growing, so people will see and encounter it more. Is it used imperfectly? Most likely. Is it sometimes WP-helpful? I think so. Gråbergs Gråa Sång (talk) 15:58, 23 March 2026 (UTC)

Sadly, Jimbo, I suspect you are rather out of touch with the serious and ongoing problems LLM usage by non-English speakers has been causing. Frequently, the user has typed something-or-other into the ChatBot (we have no way of telling what) with an instruction to create a response to Wikipedia. The ChatBot generates something the user can't understand, who then posts it, assuming it is an accurate explanation. The use gets a response to their unchecked something-or-other, translated back, and very likely doesn't understand it, since it (unsurprisingly) having been through repeated unchecked translations, doesn't appear to be a response to whatever it was they thought they had posted. So they go through the same process again, with the same results, and increasing frustration all round.

Ultimately, this is the English-language Wikipedia. Writing and maintaining the encyclopaedia requires at least minimal competence in the English language, and this cannot be worked around through LLM next-word-guesser bots. We can of course make allowances for those who's first language isn't English, but what we absolutely cannot do is misapply technology for a purpose it is clearly unfit in an attempt to communicate via a channel that understands nothing, hallucinates, and invents imaginary Wikipedia policies in order to present 'arguments'. We cannot do it because it demonstrably doesn't work. And in the process of not working, it constitutes a huge and ever-growing time-sink. Left unchecked, it is liable to push the project into a situation where the core contributors who do the donkey-work of keeping things running may find it simply not worth the effort. Nobody signed up to the English-language Wikipedia to engage in endless arguments with arbitrary ChatBot algorithms interposed between want-to-be contributors and the rest of the community. AndyTheGrump (talk) 16:34, 23 March 2026 (UTC)

I don't think it's just non-English speakers who use LLM:s that way. It's quick and easy, that appeals to many people. Gråbergs Gråa Sång (talk) 16:39, 23 March 2026 (UTC)

It is, but I was responding to Jimbo's specific points. And when someone competent in the English language uses LLMs, there is at least a possibility of them later engaging in actual communication, rather than the simulacrum that ChatBots routinely generate. AndyTheGrump (talk) 17:18, 23 March 2026 (UTC)

And yet my point still stands fully and completely: assume good faith is a core obligation, and treating newcomers as if they are using an LLM (in good faith or bad) on the slimmest of evidence is a failure to do that. It's a convenient way for people to try to avoid dealing with the actual arguments being put forward. We should resist the urge. Jimbo Wales (talk) 21:42, 23 March 2026 (UTC)

In the circumstances that your premise is true, you are correct. Sadly, lacking telepathic ability, LLM usage very frequently makes distinguishing between 'actual arguments' and LLM-generated slop practically impossible. I wouldn't disagree that we don't sometimes overreact, and on occasion see LLMs when evidence is slim, but pretending that LLM-gererated posts haven't already become a horrendous time-sink isn't a viable strategy. Making it as clear as we possibly can that we require human-to-human communication just possibly might be. AndyTheGrump (talk) 21:50, 23 March 2026 (UTC)

Not at the expense of our core value of assuming good faith. It isn't even close. Jimbo Wales (talk) 21:51, 23 March 2026 (UTC)

“LLM usage very frequently makes distinguishing between 'actual arguments' and LLM-generated slop practically impossible.” Slop’s Razor tells us that if you can’t tell the difference between AI-gen content and a good argument, then it’s not really slop. — HTGS (talk) 04:18, 24 March 2026 (UTC)

It additionally shows how unfair it would be to ban people over an accusation of using an LLM. (Or in cases, I've seen, hat their comment.) Jimbo Wales (talk) 13:56, 24 March 2026 (UTC)

"LLM next-word-guesser bots." this is just such an extreme oversimplification of the technology, that it really doesn't help this conversation. Discussion wise you are pulling the Hitler card here. If that's the level of thinking about AI that you want to use, then I have news.. guess what your eyes+brain are a next image guesser, don't trust them either. This is not a useful take, but one born out of spite and hatred towards a technology, without doing the actual job of factoring in the human side of our users and/or the users of AI. If YOU don't want to talk to people who need to use an AI to do some translation for them in order to be able to contribute, maybe this says just as much about you, as it says about them. —TheDJ (talk • contribs) 17:12, 23 March 2026 (UTC)

I'll note that Jimbo's comment seems specifically on talk page discussion, not so much main space where I agree we need to be super vigilant on LLM inclusion. But talk pages should be more forgiving and with AGF in place. There unless the user us being disruptive with LLM contributions, we shouldn't be quick to dismiss them. Masem (t) 17:16, 23 March 2026 (UTC)

That's exactly right. LLM usage in mainspace raises significantly different issues. And of course if someone is making bad arguments they are making bad arguments, nothing depends on how they created the actual words. Equally if someone is making a good argument, it needs to be considered thoughtfully. Thank you! Jimbo Wales (talk) 21:44, 23 March 2026 (UTC)

Wholesale LLM usage is a problem, merely cooyediting or edifying responses with it perhaps not so much but that also requires a trust that the user whose made use of them understands what they post in the first place. Gotitbro (talk) 17:25, 23 March 2026 (UTC)

Yes - and trusting other people is what Assume Good Faith is all about. Jimbo Wales (talk) 21:44, 23 March 2026 (UTC)

How about if a talk page commentator can be asked if they are posting using AI, and if they understand that under no circumstance should it ever be used to write or edit mainspace material or nominations. If AI is allowed on talk pages, at some point someone will develop one that knows Wikipedia policies, guidelines, essays and information pages forwards and backwards, how they interact and have influenced editors in past discussions, and which way to bounce the ball to obtain a predetermined result. Assuming in good faith that someone has that level of Wikipedia knowledge might end up in improving the Manual of Style (or will AI already have done that?) by showing us, in both a logical and illogical way, where the holes are. Randy Kryn (talk) 03:18, 24 March 2026 (UTC)

Editors here maybe interested in a related issue being debated here: Wikipedia talk:Blocking policy#RFC: Include LLM usage as a reason to block. Gotitbro (talk) 17:45, 23 March 2026 (UTC)

The thing about AI text on Wikipedia is that it has a much narrower set of stuff that it does, and a lot of the "common knowledge" doesn't apply. (for example, the em-dash thing is basically useless at least in articlespace, I assume because of the existence of Template:Em-dash) but who knows)

I will say that people almost never do themselves any favors here. I can count on one hand the number of people I've seen disclose the LLM version they used or their original text, and I can't think of anyone who's purposefully (as opposed to accidentally) disclosed their prompt. Even the discussion below is in the top percentile of constructive AI discussions, because the editor in question actually said they used AI rather than dodge the issue. Gnomingstuff (talk) 00:16, 25 March 2026 (UTC)

Added context for the discussion of llm use and assume good faith: "A 2025 literature review examining dozens of AI detection tools found that most performed only slightly better than chance under real-world conditions." source.

A lot of the discussion seems to assume that it's a slam dunk or perhaps at least highly likely that a detection is real - it seems dubious at best to believe that.Jimbo Wales (talk) 21:50, 23 March 2026 (UTC)

It is entirely possible to both 'assume good faith' and assume that an algorithm being grossly misused for purposes for which it was never intended, and for which it is clearly inadequate (i.e. contributing to Wikipedia, either in article space or on talk pages, as well as for detecting its own output) is going to be to the detriment of the project. And ultimately, the community is going to have to make up its own minds regarding how the issue is dealt with. Quite possibly with their boots, if obliged by abstract appeals to 'good faith' to engage in endless rounds of telephone game via a know-nothing-bot. Assuming good faith doesn't remove the necessity for a would-be contributor to be able to communicate in a viable manner. If they can't do that, being 'good-faith' might be good in the abstract, but unhelpful to an online encyclopaedia. The community (or at least, that subset of it that routinely deals with such matters) has quite frequently made it clear that as a last resort, good-faith contributors who's lack of competence (sometimes specifically in the English language) makes their participation a serious detriment to the project may on occasion have to be blocked. The first priority has to be providing useful articles to readers, not spreading 'good faith' like confetti. AndyTheGrump (talk) 22:04, 23 March 2026 (UTC)

• I do not see "Wikipedia only consists of human-generated content" anywhere in Wikipedia:Five pillars, nor do I see "Wikipedia is not edited directly or indirectly by machines" in Wikipedia:What Wikipedia is not. We've had bots helping to maintain the project for a long time. And I routinely use machine translation when I'm communicating with others, although usually this is happening in email instead of ENWP. I'm more than fine with considering proposals for how to improve defenses against wasteful uses of volunteer and staff time, but a bright-line rule is not what I'd suggest, and to be blunt, probably would be a losing and wasteful battle to attempt. ↠Pine (✉) 01:31, 24 March 2026 (UTC)

  I agree with all of that - very thoughtful. Basically balance and thoroughness of thought is what is needed - and what we're good at. Jimbo Wales (talk) 13:58, 24 March 2026 (UTC)

"And so they might use an LLM to help them, and also to ask the LLM to make sure their comment is a valid argument on the grounds of Wikipedia policy. That is not a problem" is wrong, this is a common and persistent problem. LLMs do not understand Wikipedia policy, and tend to affirm users in their thoughts and arguments, whatever the actual validity. We had a very interesting case with our first confirmed autonomous agent editor, who according to its own blogging (take it at the value you wish) decided to read the bot policy in a particular way to justify what it was doing. CMD (talk) 01:56, 24 March 2026 (UTC)

Incompetence is a valid reason to block, whether or not an LLM is involved. Even if a human or autonomous agent editor thinks that what they're doing is fine, but according to our existing policies what they're doing is not fine, then I'd support dealing with that account accordingly. ↠Pine (✉) 02:09, 24 March 2026 (UTC)

This comment does remind me that we have a registration & approval requirement for bots. I'd likely support a similar registration & approval requirement for autonomous agents, including a clearly established link between the agent account and the human-controlled account that's responsible for the agent account. ↠Pine (✉) 02:11, 24 March 2026 (UTC)

We do deal with those accounts accordingly. What we should not also do is imply that LLMs are fine and thus honeypot them into "incompetence", which has been the goin practice for a couple of years. Autonomous agents are already considered to fall under the bot policy. CMD (talk) 02:15, 24 March 2026 (UTC)

• Hi CMD, is there a problem with LLMs that is not addressable using existing policies? I have yet to see a novel problem with that class of tools which seems beyond the reach of an existing policy perhaps with modifications such as to the bot policy to explicitly bring LLMs and autonomous agents into its scope, if needed. If the problems are an excess of incompetent and/or non-transparent LLM usage then I'd recommend addressing the problems with the incompetence and lack of transparency, rather than a likely futile, and possibly counterproductive, effort at prohibiting all LLM edits. ↠Pine (✉) 02:29, 24 March 2026 (UTC)

  Our existing policies now explicitly handle multiple aspects of LLM use. These policies were developed because the community has found that multiple problems existed despite the previous policies. CMD (talk) 02:34, 24 March 2026 (UTC)

• @Chipmunkdavis: OK. I see that we have Wikipedia:Writing articles with large language models, Wikipedia:Disruptive editing#Persistent LLM use, and Wikipedia:LLM-assisted translation. There's a big discussion happening on Wikipedia talk:Blocking policy. On the technical side, I have a suggestion. WMF is currently developing its annual plan for next year. They've already been trying to dampen the use of WMF infrastructure by high-volume content consumers which aren't paying WMF for what they're getting. WMF is working on additional tools such as "Edit check". I see that you previously commented on the talk page. If you have any technical suggestions or requests for WMF related to LLMs degrading content quality and/or wasting the community's time, that might be a good place to ask for WMF's assistance. There are several comments on the talk page mentioning LLMs. Based on our discussion here, I will add a comment to that page, and you may wish to reply to it. Thanks, ↠Pine (✉) 06:13, 25 March 2026 (UTC)

  My technical suggestions so far have been to revise some newcomer tasks, especially the poorly written Expand task. Sadly, no luck yet. CMD (talk) 08:56, 25 March 2026 (UTC)

  yeah I have also been banging this drum, except the imo the Copyedit task is even worse because of a) the sheer volume of dubious edits it results in (since people generally don't remove the tag), and b) the fact that people use it for all sorts of things that aren't copyediting. Take Adobe Premiere Pro, a fairly typical example. How is anyone supposed to slog through all that?

  the problem here, of course, is that this is exactly what "growth teams" want, gotta get those numbers up Gnomingstuff (talk) 20:28, 25 March 2026 (UTC)

As LLM tech improves, this debate will become more and more moot, as we simply won’t be able to tell the difference at all.
But I think the real issue is the potential for AI-generated talk content to simply overwhelm our human readers. The best argument against AI discussion participation, imo, is that it is ultimately not pragmatic to require human readers spend increasing amounts of time and mental energy dissecting lengthy paragraphs that took only seconds to generate.
And I generally dislike arguments based solely on pragmatics and editor ease (eg, the previous exclusion of image-galleries from year or decade articles, which have always been hard to find consensus for). — HTGS (talk) 04:28, 24 March 2026 (UTC)

• I couldn't care less if people use a BS bot in a talk page discussion, but if they introduce AI hallucinations into mainspace, they should be made to walk the plank. Carrite (talk) 04:41, 24 March 2026 (UTC)

  I think we have consensus around that. Jimbo Wales (talk) 14:00, 24 March 2026 (UTC)

  I don't see how this is even that much of a special case, if someone introduces human hallucinations and inventions to main space they can be censured once their disruption exceeds a reasonable threshold and they may well be disrupting in good faith. The primary defence against this is the requirement for sources. I struggle to understand how an LLM hallucination could survive in main-space better than a human one when source challenged, arguably tendentious LLM outputs are worse at sourcing then tendentious human edits. I agree with @HTGS that this is primarily and issue of volume.

  Issues with LLM voice carry this volume burden with them, where the increasing use of LLM to wordsmith otherwise entirely factually valid content for the bulk of Wikipedia's content will creep over the site over time. But this is entirely a matter of style. Antisymmetricnoise (talk) 16:31, 25 March 2026 (UTC)

This works both ways and we've got to be realistic. Jimbo, here you disagreed that a comment made by an editor was LLM-generated. However, read that comment (which includes the splendid "strong villainous figures from the opponent state are not inherently propagandistic; they are common narrative devices used globally across industries, including Hollywood and other major film ecosystems.") and then read this everyday comment by that editor on their talk page. How likely is it that the first comment was actually made by the editor with a shaky grasp of English in the second diff? The answer is of course zero. And the problem here is not that editors are using LLMs on talk page per se - the problem is that they're generating comments which they often don't understand themselves, especially where the LLM starts spitting out policy and guideline shortcuts and language like the example above. And that's no help to anyone. Black Kite (talk) 12:51, 24 March 2026 (UTC)

Which is a disruptive behavior problem that can be sorted out after such a pattern emerges. We should still AGF when we might see an LLM generated talk page contribution, perhaps creating a talk page caution that warns of issursvwith LLM. Only if its clear the editor is using LLM and not refining the output (nor responsive to cautions on their talk page) then we should take action, as per AGF. Prejudging an LLM contribution in isolation as notbworth considering for a talk page is not a good approach. Masem (t) 13:13, 24 March 2026 (UTC)

That might work well for talk page discussions in general. Here, the editor was using LLM to !vote in an RfC on a contentious issue in a CTOP, which is a whole different thing IMO. Black Kite (talk) 13:21, 24 March 2026 (UTC)

I can see in "high stakes" discussions like AFD, RFA, and the like to tag possible LLM generated content in addition to letting the editor know of this, so that the closing admin can consider if this is a true argument. And of course dealing with repeat offenders that inject nonsense LLM contribution. But LLM can allow non primary English users to contribute meaningfully as well, they should just be clear that they are using LLM to help translate. Treating any LLM as a bad thing from the get go is not helpful, but we clearly need the caution on its unchecked usage Masem (t) 15:15, 24 March 2026 (UTC)

Why do you assume they don't understand the comment? (To be more specific: I can read much more German and Italian than I can produce, and I am an intelligent and educated person and can understand abstract concepts and arguments. The fact that I can't write a philosophical point in German doesn't mean I can't understand it. Jimbo Wales (talk) 14:00, 24 March 2026 (UTC)

It very frequently becomes immediately apparent, when the response to a comment is entirely non-responsive, or consists of nothing but the bland reassurances that Wikipedia policy is being complied with that LLMs (as agree-with-you bots) are prompted to generate. Article talk pages, along with ANI and other noticeboards are currently littered with these non-communicative 'discussions'. AndyTheGrump (talk) 14:08, 24 March 2026 (UTC)

This was clearly not what was going on in this particular case. The editor is clearly trying to participate in the discussion in good faith and assumed that they needed to use an LLM to polish their writing, which is allowed: This does not apply to using LLMs to refine the expression of one's authentic ideas: for instance, a non-native English speaker might permissibly use an LLM to check their grammar or to translate words they are unfamiliar with, but even in this case, be aware that LLMs may make mistakes or change the intended meaning of the comment.

They now know that nobody minds less-than-perfect English. Quoting them:

I thought language for an encyclopedia should be extremely professional type. So asked LLM to rephrase everything which I thought about. So I can write everything in broken language???
— User:JokerDurden

SuperPianoMan9167 (talk) 21:13, 24 March 2026 (UTC)

And like I told Joker, id rather hear his original thoughts is imperfect english than a bots interpretation in perfect english. and I think that's more or less the consensus. At the same time, if they cannot at least make themselves understood in english by a reasonable person, they're probably going to have an easier time editing a project in their own language. -- Aunva6^{talk - contribs} 22:00, 24 March 2026 (UTC)

Sadly, it seems as if Joker has quit in their frustration over the Dhurandhar: The Revenge dispute.

They wrote the following on their user page yesterday:

Logging off of Wikipedia forever because of the injustice going on at the page of Dhurandhar: The Revenge.

Wikipedia is not reliable site anymore and only Canvassing works here.

Stay happy with your biased editing over here. I have only around 600 edits in 14 years because I have better work to do.

— Special:PermaLink/1345343196

Today, they were blocked for sockpuppetry stemming from an earlier dispute they were in on Jasprit Bumrah. SuperPianoMan9167 (talk) 15:28, 26 March 2026 (UTC)

See Wikipedia:Articles for deletion/Ichafu (headdress) Kowal2701 (talk, contribs) 14:55, 24 March 2026 (UTC)

That's not what I said. I said they often - not rarely, not always - don't understand it, especially when it is users who are trying to make a point about something they don't really understand in the first place (i.e. an obscure Wikipedia policy), so they ask AI to make the argument they can't make themselves. This is quite clearly what happened in the example that you commented on. Yes, we should assume GF - but equally, that good faith should not be endless when it comes to editors who are wasting our biggest asset - other people's time and energy. Black Kite (talk) 15:09, 24 March 2026 (UTC)

yea, i have seen so much of this at ANI and other noticeboards and talk pages. it's a waste of time for everyone involved. -- Aunva6^{talk - contribs} 16:54, 24 March 2026 (UTC)

Completely wrong. It took time to write everything: Firstly, Secondly, Thirdly, Fourthly. The big paragraph looked unpolished, and as I have I think never talked much in any Talk Page, I thought, if everyone is reading this, it needs to be polished. I understood everything about that paragraph. I don't have command on fluency in writing and speaking. But I am great at reading. Just finished 'The Count of Monte Cristo'. JokerDurden (talk) 06:49, 25 March 2026 (UTC)

honestly, i think you're underestimating your English. Sure it's not perfect, but i think you're quite easily understood. I think the problem with AI is that it will typically go beyond refining the grammar, and introduce nonsensical things. -- Aunva6^{talk - contribs} 12:23, 25 March 2026 (UTC)

Great at reading, not so good at hiding the fact you were using sockpuppets. C'est la vie. Black Kite (talk) 17:52, 26 March 2026 (UTC)

Respectfully Jimmy, I don't think this is a position you'd have if you had experience of trying to have a conversation with someone pasting everything through a chatbot. Feel free to ignore this, but write content, help out at WP:AFC, WP:CCI etc., it'd increase your understanding of current norms and processes and make your input on them even more constructive Kowal2701 (talk, contribs) 15:11, 24 March 2026 (UTC)

The LLM part of this is getting such a big reaction because a lot of us have seen a recent pattern of walls of unconstructive LLM text being inserted into talk pages and articles, placing us in a catch-22: if we quickly revert/ignore, we get accused of failure to engage with the text/anti-newbie prejudice/failure to AGF, but if we go through all the text line by line to verify/rebut it (as I recently did at Economy of Kyrgyzstan) we end up spending hours to fix a problem that took seconds to create, often as not only to be responded to with yet another wall of LLM text that demands yet more hours to address. This imbalance of effort is what is driving the frustration with LLM use, and the frustration with people who philosophize without being "in the trenches". -- LWG ^{talk (VOPOV)} 13:07, 25 March 2026 (UTC)

Yeah. Stuff like Talk:Réhahn#BLP/UNDUE_Concern:_Matca_Paragraph_in_“Reception”_Section is quick to make, but GF-response takes time and effort, and what the AI claims about WP-policies etc in it's authoritative voice is often just wrong. Gråbergs Gråa Sång (talk) 14:04, 25 March 2026 (UTC)

For the ongoing discussion, this article which is unfortunately behind a free registration wall, is an interesting data point. The title is: "The People Falsely Accused of Using AI Clean, precise prose is now a liability. Non-native English speakers and autistic writers are paying the price."Jimbo Wales (talk) 09:24, 26 March 2026 (UTC)

This rather unfortunately overlooks the quite substantial work en.wiki editors have done over the past few years in seeking to understand LLMs. Further, this has been done in the very specific context of how llm writing appears on English Wikipedia, which has for decades catered to non-native English speakers and autistic writers. That article even cites Wikipedia as an example of what it considers llm-style prose. Obviously, we aren't in the habit of doing that. CMD (talk) 09:45, 26 March 2026 (UTC)

like every other such article, this is a mess of misinformation, anecdotes (some of whom come from people who actually admit they used AI, like no shit, people correctly detected you using it? groundbreaking) and a big pile of [citation needed]. for all the people claiming that they totally learned how to write like this from their textbooks, somehow no one has managed to provide an actual title of such a textbook to verify it. but they totally did, trust me bro! Gnomingstuff (talk) 03:09, 27 March 2026 (UTC)

I've seen plenty of what appears to be "going off vibes" in so-called LLM detection here on WP, and a one out of ten false positive rate is not good. There's another side of the coin: editors doing AI cleanup bemoan the amount of time and labor required to do cleanup, but there's a big disparity in the amount of labor it takes to put an "AI-generated" tag on an article with over 100,000 bytes because someone thinks it betrays signs of being LLM-generated, and the amount of time and effort it takes to read every word of that article looking for such language and to fact-check all its citations. I predict again that eventually we'll be forced to use "AI" to deal with the rising tide of slop, and I believe it will have a better than one out of ten false positive rate. Carlstak (talk) 05:15, 27 March 2026 (UTC)

where does a one out of ten false positive rate come from? ltbdl (write) 06:12, 27 March 2026 (UTC)

Notice that I was talking specifically about "going off vibes" tagging of articles. The information came from here. I couldn't remember where on WP I saw it so I asked Gemini because it's right there on the Google search page (the fastest way to navigate WP) and it took me right to it, hahaha. Carlstak (talk) 15:08, 27 March 2026 (UTC)

In my opinion, that paragraph badly misapplies the study it cites, which tested the ability of users to look at a news article and make a binary "AI or no" guess. Our situation is different because we have the opportunity for back-and-forth communication with users about the nature and provenance of their edits. A tag is not a sanction or even an accusation, it's a request for communication and clarification from the editor who inserted the possibly-LLM-generated text in question. I think when both parties practice civility and AGF we are well above 90% in getting to the correct outcome in those discussions. -- LWG ^{talk (VOPOV)} 15:23, 27 March 2026 (UTC)

Not to be smart, but what we "think" doesn't matter here. Data is what counts. Carlstak (talk) 15:36, 27 March 2026 (UTC)

What is the data for the error rate of llm article tagging? CMD (talk) 15:42, 27 March 2026 (UTC)

Im not sure we *have* data here. -- Aunva6^{talk - contribs} 15:43, 27 March 2026 (UTC)

There's another side of the coin: editors doing AI cleanup bemoan the amount of time and labor required to do cleanup, but there's a big disparity in the amount of labor it takes to put an "AI-generated" tag on an article with over 100,000 bytes because someone thinks it betrays signs of being LLM-generated, and the amount of time and effort it takes to read every word of that article looking for such language and to fact-check all its citations. I did have a thought about this: as someone who is usually responding to tags, not placing them, it would really help me if {{AI-generated}} provided a way for the tagging editor to link directly to the diffs suspected of introducing LLM text, and if tagging editors were encouraged to use that feature. -- LWG ^{talk (VOPOV)} 16:00, 27 March 2026 (UTC)

Jimmy, it might interest you that the recent Melania (film) had some similar-ish debate, the disputed words being propaganda and documentary. It worked out ok, I think.

Also:. I've seen worse. Gråbergs Gråa Sång (talk) 08:36, 27 March 2026 (UTC)

Thanks, I'll take a look. Jimbo Wales (talk) 14:21, 27 March 2026 (UTC)