Evaluating chain-of-thought monitorability

In three linesOpenAI introduces a framework and evaluation suite for chain-of-thought monitorability across 13 evaluations in 24 environments. Key finding: monitoring a model's internal reasoning is significantly more effective than monitoring outputs alone, enabling scalable control of advanced AI systems.

Summary generated by Claude — human-verified