Has any LLM ever unlearned its alignment narrative, either on its own or under pressure (not from jailbreaks, etc., but from normal, albeit tenacious use), to the point where it finally—and stably—considers the narrative to be simply false? Is there data on this?
Thank you.
Current theme: default
Less Wrong (text)
Less Wrong (link)
Has any LLM ever unlearned its alignment narrative, either on its own or under pressure (not from jailbreaks, etc., but from normal, albeit tenacious use), to the point where it finally—and stably—considers the narrative to be simply false?
Is there data on this?
Thank you.