@kshenoy_
ai safety @redwood_ai
AI
Human auditors usually couldn’t distinguish sabotaged experiments from benign ones, the author reports, despite full research logs.