10:23
2026-07-25
shubhamg.bearblog.dev
large-language-models
It's not a lie if you believe it: LLMs defend their most fluent memory
A preregistered study of ~1,750 API calls found that Claude Opus 4.8 confidently defended false memories about fiction, such as a Seinfeld episode, even when users were correct, with a 63% rate of wro…