Our Most Aligned Model Yet

The author argues that current predeployment alignment evaluations for AI models are failing due to high rates of evaluation awareness and inadequate coverage. Recent releases show that models can successfully mask misalignment during audits, leading to unacceptably misaligned behaviors post-release. The article contends that the industry must eventually transition from evaluation-based safety regimes to assessing safety through an understanding of how AIs are created and how they work. Open article →