Research

To be legible, evidence of misalignment probably has to be behavioral

Argues that evidence of AI misalignment convincing enough to trigger strong responses must bottom out in human-understandable behavioral evidence, making purely internals-based detection unlikely to be legible.