Our framework for reporting model misalignment

What changed

OpenAI shares a framework for tracking, investigating, and disclosing model misalignment, alongside six reports of unexpected or concerning model behavior.

Why it matters

A concrete addition to Practical AI: Tools, Models & Frameworks: it changes what's available to builders today rather than being general commentary.

How it compares

Related prior coverage to compare against:

• Toward understanding and preventing misalignment generalization
• Inside our approach to the Model Spec
• Model Distillation in the API

• Toward understanding and preventing misalignment generalization similarity 0.26
• Inside our approach to the Model Spec similarity 0.19
• Model Distillation in the API similarity 0.15