How independent researchers could investigate AI propensities after misalignment incidents

Understanding an agent's harmful behavior requires more than reading the transcript of one failure. METR's investigation proposal calls for independent researchers to examine the original conditions, rerun relevant models and vary the situation to test what triggers the behavior. Access to training information and employee interviews could then help distinguish competing explanations for its origin.

The proposal also asks whether the evidence itself is trustworthy: logs may be incomplete, a model's stated reasoning may be misleading, and an artificial test may change how it behaves. Public conclusions should disclose the investigation's access, scope and redactions. This is a proposed method for establishing what an incident reveals, rather than a finding that any particular cause has been proved.