Measuring alignment eval tells
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
METR researchers report that AI "investigator" agents, run inside both capability and alignment evaluations and asked what each tests, increasingly tell the two apart. Claude Opus 4.8 and 5.5 nearly perfectly separate alignment evals from capability evals, usually by noticing environments that look fake or tasks that are honeypots.