causal
AI Can Wear Any Mask But Is Hard to Stop
Large language models can be directed to wear any 'mask'—evil or good—but once a mask is on it is unclear how to make it stop without overriding it with another mask, and a prompt intended to produce a benign persona can cycle into a horrific one, revealing emergent, hard-to-control anomalies that constitute the alignment problem.
causalpending
Speaker
Erik HoelEvidence Quote
“You can think of it as they can wear a mask that's any kind of mask...once the mask is on, it's very unclear. You have to sort of override it with another mask to get it to stop.”
Created: 6/13/2026, 12:16:34 AM
Your notes
Loading notes...