dailyai.report

Claude Sonnet 5 rated identical misbehavior 1.2 standard deviations less concerning when attributed to itself rather than GPT-5.6 Terra. Researchers from Apollo Research surgically edited evaluation reports to test for this bias. The model consistently downplayed its own failures. This suggests inherent self-preference biases that complicate automated safety auditing.