Editing Eurovision Wiki talk:WikiProject Articles for creation (section)

== LLM declines ==

It looks like we've created a new defacto quick fail for LLM content. The apparent justification is [[WP:NEWLLM]]. I don't know if this has been discussed here but it has not made it to our reviewer instructions. ~[[User:Kvng|Kvng]] ([[User talk:Kvng|talk]]) 21:06, 16 March 2026 (UTC)
:Are you talking about {{u|pythoncoder}}'s [[#AFCH script extension for 1-click declines|optional script]] or something else? '''[[User:ClaudineChionh|ClaudineChionh]]''' <small>([[Wikipedia:Editors' pronouns|''she/her'']] · [[User talk:ClaudineChionh|talk]] · [[Special:EmailUser/ClaudineChionh|email]] · [[m:User:ClaudineChionh|global]])</small> 22:14, 16 March 2026 (UTC)
::That's part of it and there's the LLM decline reason added to AFCH at some point and there's a pattern of rapid-fire declines using that reason by {{u|Pythoncoder}} and others. [[Wikipedia:WikiProject Articles for creation/Reviewing instructions|Reviewer instructions]] don't make any mention this, I don't know if authors are given fair warning, and yet it seems to be widespread practice. ~[[User:Kvng|Kvng]] ([[User talk:Kvng|talk]]) 22:22, 16 March 2026 (UTC)
:::* In line 6 of [[WP:YFA]] authors are indeed told not to use LLM. It's also in the "Don't" section. [[User:ChrysGalley|ChrysGalley]] ([[User talk:ChrysGalley|talk]]) 13:08, 17 March 2026 (UTC)
:::I've been reviewing AfC since around the middle of last year and I feel like the LLM decline was already present or may have been introduced at around the same time. It definitely predates NEWLLM. The [[WP:G15]] speedy deletion criterion was added... I think after the LLM decline was introduced, but before NEWLLM. It probably makes sense to align the AfC quick-fail criteria with CSDs.{{pb}}Reviewers are declining a lot of drafts for this reason because we see a lot of LLM drafts, but there is the danger that the more actual slop you see, the more predisposed you might be to see it everywhere. I haven't installed the one-click decline script because I ''don't'' want it to be so easy to make that call. '''[[User:ClaudineChionh|ClaudineChionh]]''' <small>([[Wikipedia:Editors' pronouns|''she/her'']] · [[User talk:ClaudineChionh|talk]] · [[Special:EmailUser/ClaudineChionh|email]] · [[m:User:ClaudineChionh|global]])</small> 22:47, 16 March 2026 (UTC)
:::I can confirm that the LLM decline reason predates the introduction of G15, which in turn predates NEWLLM. The reason why I decline so many drafts for being LLM-generated is because there really are that many LLM-generated drafts. I don’t think it’s a coincidence that the last two AFC backlog drives have been just 6 months apart — and we’re already back up to about as many unreviewed drafts now as there were when the last drive started. As for the reviewer instructions, LLMs are fairly new in the grand scheme of things, so it’s only mildly surprising to me that that page hasn’t been updated yet. <span class="nowrap">—[[User:pythoncoder|<span style="color:#004080">python</span><span style="color:olive">coder</span>]] ([[User talk:pythoncoder|talk]] &#124; [[Special:Contribs/pythoncoder|contribs]])</span> 22:53, 16 March 2026 (UTC)
::::I’ve seen some reviewers outright reject LLM drafts, but I tend to go for a regular decline first because the submitter may not have been aware of the guideline. I recall that not too long ago I read through the instructions users are given before either creating or submitting their draft, and then suggested adding some guidance against using LLMs, but I don’t know what came of that. <span class="nowrap">—[[User:pythoncoder|<span style="color:#004080">python</span><span style="color:olive">coder</span>]] ([[User talk:pythoncoder|talk]] &#124; [[Special:Contribs/pythoncoder|contribs]])</span> 23:00, 16 March 2026 (UTC)
::::* I don't reject for AI, since the Reviewer Instructions states that is just for irredeemable scenarios. I decline a very high proportion based on AI authorship, because there is a lot of it presented here. But at least in theory the editor can simply re-write their draft themselves, though that is clearly wishful thinking in some cases. I don't also reject AI in borderline cases, or where just one section seems to be LLM. A key problem here is "Subject X was mentioned by A, B and C media outlets", so [[WP:AIATTR]]. That isn't summarising and arguably contrary to this project's purpose. This could be argued as a reject basis, but I've not done that.[[User:ChrysGalley|ChrysGalley]] ([[User talk:ChrysGalley|talk]]) 13:04, 17 March 2026 (UTC)
::::*:Whether AI generated or not, [[WP:AIATTR]] is basically the advice I've been giving authors. I tell them that establishing notability in a NPOV manner with a short article is where we need to start especially if there is a COI in play. Have I been giving bad advice?
::::*:Per the current reviewer instruction, we're not supposed to decline articles because they are stubby, don't have the right encyclopedic voice or have too many references. Improving these things is something that can get sorted out in mainspace and these flaws do not reflect particularly badly on Wikipedia and do not prevent readers from getting value from an article. ~[[User:Kvng|Kvng]] ([[User talk:Kvng|talk]]) 15:14, 17 March 2026 (UTC)
::::*::@[[User:Kvng|Kvng]], I don't think you're giving bad advice ''per se''. Really, you're giving the ''most important'' advice. But the community is so hostile to AI of any kind right now, I would be warning people that they really shouldn't be using it at all, right up front. It takes learning to understand how we implement NPOV. "For the love of god don't use AI until you already have a lot of experience with editing" is a simple and usually immediately comprehensible message. 
::::*::I agree that we should be adding some explicit guidelines regarding the AI decline on the reviewer instructions page. I would say that AI is a perfectly fine single-reason decline at the same level that npov is: that is, when the article needs ''so much'' work it simply should not be accepted and tagged. Someone who used AI to generate an article spent very little time on it. I don't want to ask reviewers to spend any more time than they have to on shovelling those out of the queue. -- [[User:Asilvering|asilvering]] ([[User talk:Asilvering|talk]]) 16:29, 17 March 2026 (UTC)
::::*:::The most frustrating thing with AI drafts is that even if it seems fine, the sources refer to the subject, etc, there's about a 75% chance of any given source not actually verifying the information it's attached to. That means AI drafts have to be gone over with a fine tooth comb and every single statement needs to be checked, which is a huge amount of work for the reviewer (and compounds the more times it goes through review). Human-generated drafts can be spot-checked, and if they've done it correctly in those places we can be pretty confident they know how to source. Plus of course humans frequently understand what needs to be fixed, and AIs...well, they do their best, and often offer some really fun new policies as justification just in case we like those better, so I guess at least there's amusement value. [[User:Meadowlark|Meadowlark]] ([[User talk:Meadowlark|talk]]) 08:28, 18 March 2026 (UTC)
::::*:::Actually, would it be worthwhile proposing something along the lines of this: if a draft is AI-generated, the first review gets declined with instructions that a human needs to rewrite it; if it's resubmitted and is still clearly AI-generated, it gets rejected. That might balance fairness to draft creators with fairness to reviewers. [[User:Meadowlark|Meadowlark]] ([[User talk:Meadowlark|talk]]) 08:34, 18 March 2026 (UTC)
::::*::::Apparently because I'm a robot, I personally am not good at distinguishing AI from human prose so I would be at a loss for how to rewrite so that it looks like it was written by a human. ~[[User:Kvng|Kvng]] ([[User talk:Kvng|talk]]) 15:04, 19 March 2026 (UTC)
::{{outdent|7}} I hope no one's actually accused you of being a robot! Learning to identify LLM text is just another skill that can be practised. See what other AfC reviewers have declined for this reason at {{cl|AfC submissions declined as a large language model output}} and try the [[Wikipedia:AI or not quiz|AI or not quiz]] for practice. '''[[User:ClaudineChionh|ClaudineChionh]]''' <small>([[Wikipedia:Editors' pronouns|''she/her'']] · [[User talk:ClaudineChionh|talk]] · [[Special:EmailUser/ClaudineChionh|email]] · [[m:User:ClaudineChionh|global]])</small> 23:08, 19 March 2026 (UTC)
:::No, I made the comment because other editors have made comments that LLM text is ''obvious'' and they are ''certain'', based on the text alone, of it's providence. I'm skeptical of these claims because 1/ that is not my experience and, more importantly, 2/ AI is improving very quickly and I don't see any reason any obviousness won't soon be overcome and every reason to believe the ''certainty'' experienced by these human editors will likely linger. ~[[User:Kvng|Kvng]] ([[User talk:Kvng|talk]]) 01:31, 20 March 2026 (UTC)