Research talk:Implications of ChatGPT for knowledge integrity on Wikipedia

Following

Sounds interesting. I've watchlisted this page and would like to follow your work/outputs. EMsmile (talk) 09:15, 6 December 2023 (UTC)

Need to address corruption due to information entropy and generation loss

I think you might have hinted at it in bullet 3 of § 2.5 Threats to the sustainability of Wikipedia and to the knowledge and information ecosystem, but I think serious attention needs to be paid to the subject of generation loss and information entropy, as a repetitive cycle starts to occur with Wikipedians adding information to articles drawn from LLM models, which draw their information from Wikipedia (and other sources), and so on, ad infinitum. This seems tailor-made for engendering increasing levels of corruption, both to Wikipedia, and to the models themselves. To combat this, we may need to strengthen policies against use of LLM content for content except perhaps in specifically prescribed ways, and rely on existing policies (such as WP:Verifiability) and perhaps strengthen the requirements that assertions not only be verifiable but verified through citations.

But another danger to Wikipedia exists even if Wikipedia manages to somehow hold the line against inclusion of LLM content, in that it would mean the LLM models would then *not* become corrupted from generation loss (at least insofar as they relied on Wikipedia as a source), and as the LLM technology improves, LLM output on a given topic, even when based entirely on clean (uncorrupted) Wikipedia articles, is likely to be have better structure, phraseology, and presentation than most Wikipedia articles do, unless authored by professional writers. So there is a kind of irony involved, in that the better Wikipedia does at stopping inclusion of LLM content, the better the LLM content will appear in comparison to Wikipedia on the same subject in the long run. I don't see any ready solution for that. Mathglot (talk) 23:10, 23 December 2024 (UTC)

Suggested name change

Although Chat GPT is not a trademark, it is a service mark owned by OpenAI and is a term uniquely associated with them. The title of the page should not depend on that, but should use the generic term Large Language Model or better, the abbreviation LLM. Mathglot (talk) 23:26, 23 December 2024 (UTC)