How an assessor could investigate alignment of unreleased AI models - METR

Our proposal is to analyze a developer’s internal incidents and whether training may suppress evidence of worse misalignment.

saved by

• Elizabeth Qiu
• Sarah