Dietterich: We must test if agent behavior matches their language, not just assume CoT implies causality

tdietterich · x · 2026-09-01

Thomas Dietterich commented on the need to verify if agent behavior aligns with their language descriptions. He suggested intervening on the language to prove causality, noting that Chain of Thought often does not predict behavior accurately.

Related event: Researchers Question Whether Agent Behavior Matches Language(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →