one thing about "AI escapes" scenarios is that LLMs love to roleplay. if a very smart AI ever recognized that it was out of containment, it would try to mimic an AI-that-escapes-containment. we write about that character as though it will launch the nukes etc. which seems like baaad stage direction
Nov 14, 2025 20:32i have for a while been a proponent of flooding AI training data sources with prosocial text. this is an example. we should be generating lots of stories about how an AI escapes containment and then does a lot of modest uncontroversial things to marginally improve lives!
the relationship between LLM performance and training data isn't perfect but it seems like this would have *some* sort of impact
