Post #2077378
2026-04-22 12:13 UTC
@aharoni Interesting - an incentives approach.
I was somewhat expecting something like:
"LLMs are trained on large amounts of text, and then given a prompt to complete in a way that's similar to the training text. The training text will rarely be people having answered 'I don't know', so LLMs tend not to output that either".
Replies (1)
-
@aharoni@wikis.world 2026-04-22 12:20
@sundew OpenAI say it themselves, more or less: https://www.theregister.com/2025/09/17/openai_hallucinations_incentives/ As the article says, their researchers understand that it's better to say "I don't know", but researchers don't make decisions—CPOs, CFOs, and CEOs do 🤷🏻♂️