Wikipedia talk:Responsibly using large language models

Notes on advice beyond the essay's scope

• Synthesis is the principle use case of LLMs, which is incompatible with WP policy, but in my experience you can stress in your prompt that the LLM is to avoid synthesising information and it tends to be fairly effective in lowering the incidence of this.
• This article assumes that editors will be novices when it comes to LLMs and that they will be using the base models available. That they might want to consider a paid model to improve reliability is valid advice.
• Detailed how-to guides that are less vague would be of more use to these novice editors, but constitute advice that should really be subject to more of a consensus.
• Ideally content quantifying the risk of hallucination and synthesis would be present but this would probably have to take the form of a study conducted by editors on a regular basis. I doubt there's much appetite for this while the epidemic of LLM misuse is in full swing.

Joko2468 (talk) 00:50, 28 March 2026 (UTC)

i.e., expanding on and building consensus around these sorts of problems: User:Festucalex/Don't use LLMs as search engines. In being rooted in established guidelines and avoiding opinionated original research that goes beyond this, the essay is currently very vague and far less informative than it might otherwise be. It's unfortunately reflective of how much work is needed on establishing guidelines that could effectively combat LLM misuse. Joko2468 (talk) 11:09, 28 March 2026 (UTC)

The humourous parts are supposed to get across the scale of the verification burden on LLM users-- this has been theorised in academia but I'm not aware of empirical research or a basis in Wiki policy so I chose to cover it less seriously. It's also intended to remove any condemnatory undertones. Joko2468 (talk) 13:57, 28 March 2026 (UTC)

The obvious question is whether "human-driven methods of research" imposes less verification burden in general. Bbbbbbbbba (talk) 23:30, 16 July 2026 (UTC)

It's compared to a source-first approach-- surely this is less of a verification burden? I've changed the language to be more explicit. Joko2468 (talk) 08:59, 17 July 2026 (UTC)

This paragraph is about direction/orientation. If I do not use AI, then I am probably orientating myself with my own first impression of (the abstracts of) the sources. Whether that is less likely to "significantly skew the direction of your research and undermine your understanding of a topic" is not that clearly cut. Bbbbbbbbba (talk) 11:54, 17 July 2026 (UTC)

Is it not clear cut? If you're using abstracts then that is reliable if incomplete information-- it's surely better than possibly digesting and taking into account an outright fabrication. Joko2468 (talk) 15:52, 17 July 2026 (UTC)

The human mind fabricates ideas all the time—original syntheses are a problem that predate modern AI. Even reliable sources are not 100% reliable. I concede that currently LLM probably does it more than the average human. But LLM can also point to the exact sentence one want to read to verify a claim, avoiding "I feel that I have seen it but I cannot find it now" situations. So there may be more things to verify, but each thing takes less effort. (Plus, arguably it is easier to practice healthy skepticism towards LLM outputs than towards your own mind.)

Also, the way the current paragraph is worded, it would be unfair to ignore the mental burden needed for self-orientation in the first place. Maybe if someone do not have the time, they should not research a topic they are not already intimately familiar with, period. Bbbbbbbbba (talk) 22:24, 17 July 2026 (UTC)

What would you amend this to? Personally, and if I'm understanding you correctly, I don't think the essay needs to be balanced on LLMs vs humans-- the scope of the essay is focused on highlighting the risks of LLM assistance. Joko2468 (talk) 18:38, 18 July 2026 (UTC)

Section on copyediting

Does anyone fancy writing the section on advice around copyediting (ofc inline w WP:NOLLM), as in which prompts or models to use etc.? (See WT:NOLLM#Refining basic copy editing for discussion about wording of the 'rule', this is just about advice) Tbh I think some of what's in the Research section can be dumbed down or made clearer, but is otherwise good. Courtesy ping Joko2468, will notify WP:AIT and WP:AIC as well Kowal2701 (talk, contribs) 18:12, 19 July 2026 (UTC)

Agree with this-- not sure I'm too good at writing accessibly. It's gone through very little scrutiny so all feedback and improvements welcome. Joko2468 (talk) 20:10, 19 July 2026 (UTC)

given that people are trying to "dumb down" NOLLM and make all "corrections" fair game whether "basic" or not, this is premature Gnomingstuff (talk) 13:11, 20 July 2026 (UTC)

I just added a little. I even used ChatGPT in (I think) a compliant way to check my text. My original, the ChatGPT list of mistakes, and my manual fixes are all in the edit history. Please revert if it's "DW slop". Dw31415 (talk) 04:22, 23 July 2026 (UTC)

Are you asking for WP:NOLLM to give out specific advice on what LLMs to use? That tells me you do not understand and/or agree with what the policy is telling you. The use of LLMs to generate or rewrite article content is prohibited. If you feel sufficiently competent you can master the use of LLMs purely for copy editing and/or translation, then please go ahead, but we at Wikipedia should certainly not help people to use LLMs. Not only or even chiefly because Wikipedia is not a technical manual but because it is far better that people stay away from LLMs entirely unless they know what they're doing. Providing suggestions would do far more harm than good. We should definitely not create the impression we recommend a certain LLM (for any purposes). It is a good thing if you feel so unsure about which LLM to use that... you don't use any of them! CapnZapp (talk) 09:54, 24 July 2026 (UTC)

... I wrote most of NOLLM, it has an exception for copyediting. Kowal2701 (talk, contribs) 10:01, 24 July 2026 (UTC)

The exception means LLM copyediting is tolerable, not recommended. CapnZapp (talk) 10:09, 24 July 2026 (UTC)

This advice is for people who already use them, such as students who regularly use them as part of their workflow. It's to educate, not to recommend. You're preaching to the choir Kowal2701 (talk, contribs) 10:13, 24 July 2026 (UTC)

It is not Wikipedia's job to help people to use LLMs. If anything, we should help them not use LLMs. CapnZapp (talk) 10:20, 24 July 2026 (UTC)

It's not our job here to impose our own preferences on people, just to represent consensus Kowal2701 (talk, contribs) 10:26, 24 July 2026 (UTC)

I'm not expressing a personal preference. I am expressing what I believe to be the consensus: namely that Wikipedia should not provide an instruction manual, especially not for tools we actively discourage. Of course, if you can present with me with consensus to the contrary, you are free to do so. CapnZapp (talk) 15:20, 24 July 2026 (UTC)

Copyediting in accordance with what's at NOLLM is not discouraged, in fact it contributed to the near-unanimous support for NOLLM in the RfC. I'm not obliged to prove a negative nor WP:SATISFY. Frankly, I'm one of only like 5-10 people who actually do AI cleanup at WP:AINB and am very alive to the overwhelming workload, you don't need to be zealous. Kowal2701 (talk, contribs) 16:00, 24 July 2026 (UTC)

Maybe we should mention the fact that diction improvements by LLM are not allowed, which should be discouraging enough to drive a few LLM users off Wikipedia. Bbbbbbbbba (talk) 01:57, 27 July 2026 (UTC)

Oppose because we’d have do verbal gymnastics to explain word corrections are allowed but not flourishments (sic). Dw31415 (talk) 11:48, 27 July 2026 (UTC)

Also this is a Wikipedia:Project namespace page, not a main namespace page (Wikipedia article). I do not think Wikipedia is not a technical manual applies here. Bbbbbbbbba (talk) 01:33, 27 July 2026 (UTC)

Prompting tips on WikiProject AI Tools

All are invited to join the discussion at Wikipedia talk:WikiProject AI Tools#Suggested prompts about building a set of responsible prompting tips at Wikipedia:WikiProject AI Tools/Guide. My aim is to both be bold and enthusiastically collaborate, in the spirit of WP:BRD. Dreamyshade (talk) 16:03, 20 August 2026 (UTC)