Wikipedia talk:Large language models/Archive 3
 | This is an archive of past discussions on Wikipedia:Large language models. Do not edit the contents of this page. If you wish to start a new discussion or revive an old one, please do so on the current talk page. |
| Archive 1 | Archive 2 | Archive 3 | Archive 4 | Archive 5 | → | Archive 7 |
Detection tools
1) Can someone link me to the best detection tools? It's mentioned on one of these talk pages but I can't find it.
2) Should we add a paragraph about detection and a link to some of these tools to WP:LLM? Should document this somewhere. If not WP:LLM, then maybe a separate informational page. –Novem Linguae (talk) 09:57, 23 February 2023 (UTC)
- See summary —Alalch E. 12:37, 23 February 2023 (UTC)
- @Novem Linguae: Here are a few more. Some of them were mentioned at Wikipedia:Village_pump_(policy)#Wikipedia_response_to_chatbot-generated_content.
- If we have something really reliable then we could mention it here. But I don't think that this is the case and this is also a rapidly developing field, so the what is reliable now may not be reliable in a few months. But having a link to an essay where they are discussed would make sense. Phlsph7 (talk) 13:34, 23 February 2023 (UTC)
- This data is for phab:T330346, a PageTriage ticket exploring the idea of "Detection and flagging of articles that are AI/LLM-generated". Feel free to weigh in there. –Novem Linguae (talk) 23:59, 23 February 2023 (UTC)
- Currently, the available detectors are being regarded as "definitely not good enough" to use for important decisions, due to frequent false positives and false negatives (and are often intended for outdated models like 2019's GPT-2). This includes OpenAI's own tool for detecting ChatGPT-generated text. This situation may change of course, especially if OpenAI goes forward with their watermarking plans. (Turnitin announced last week that they have developed a detector that "in its lab, identifies 97 percent of ChatGPT and GPT3 authored writing, with a very low less than 1/100 false positive rate" and will make it available to their customers "as early as April 2023." But even it's worth being skeptical about whether they can keep up these levels of sensitivity and specificity.) Regards, HaeB (talk) 06:45, 24 February 2023 (UTC)
- If/when such detection and flagging is implemented, probably the best course of action would be not to forbid them, but maybe create some "quarantine" space (similar in spirit to WP:Draft), so that it can be properly edited and verified into a valid Wikipedia article. This allows for more adaptability because these models are surely going to grow and get better at their task, and in time they could be a real help for Wikipedia's growth. What do you think? Irecorsan (talk) 16:06, 10 March 2023 (UTC)
- LLM's issues with generating inaccurate but fluent-sounding information, combined with forging citations, makes them pernicious. They take experienced editor time and expertise to spot and remove. I currently see them as a net negative to the encyclopedia. –Novem Linguae (talk) 23:15, 10 March 2023 (UTC)
Merger proposal
@CactiStaccingCrane and JPxG: I propose merging Wikipedia:Using neural network language models on Wikipedia into into this page. I think the content in the other article can easily be explained in the context of this one. QuickQuokka [talk • contribs] 19:51, 5 March 2023 (UTC)
- Hello QuickQuokka and thanks for your suggestion. However, I think the purpose of these two pages are very different: the page here is a draft for a policy while the page Wikipedia:Using neural network language models on Wikipedia is an essay that includes many very specific suggestions on how LLMs may be used. It's probably not a good idea to try mix these purposes up and to include all those suggestions and ideas in a general policy. Another point is that some of these suggestions are controversial, for example, which uses are safe and which ones are risky. A policy should represent a very general consensus and not include controversial claims. Phlsph7 (talk) 21:13, 5 March 2023 (UTC)
- See also this discussion this discussion. My suspiscion is that if we merge the two people will be continually trying to get rid of examples as "not to do with policy" so it's probably good to leave the other page where it can have less aggressive editting. Talpedia (talk) 21:17, 5 March 2023 (UTC)
- Yeah, whilst these two pages are broadly on the same subject, they’re coming at it in two completely different and incompatible directions. They can certainly feed off each other in some ways, but merging them will devalue both. Letting both exist, and letting them converge or diverge organically as time passes, would seem to serve Wikipedia as a whole and interested editors in particular better than stopping everything to try to meld them together now. We’re not on a deadline here. — Trey Maturin™ 02:23, 6 March 2023 (UTC)