OpenAI Tries To Measure Whether AI Reasoning Can Be Trusted

OpenAI is putting Monitorability under a microscope with a new framework meant to measure whether a model’s reasoning can actually be watched for trouble. In a December 18, 2025 research post, the company introduced 13 evaluations spanning 24 environments to test how well monitors can infer what a reasoning model is doing from its chain of thought, instead of just judging the final answer. 

The core idea is simple but uncomfortable. Plenty of high stakes decisions are hard to supervise directly, so the next best option is supervising how the decision was made. OpenAI defines monitorability as the ability of a monitor to predict important properties of an agent’s behavior, including misbehavior like deception or reward hacking, plus more neutral signals like whether it relied on a hint. It also stresses this is two sided. Better monitors help. But some agents are inherently easier to read than others, depending on what signals you give the monitor, including actions, outputs, or chain of thought. 

The evaluation suite clusters into three archetypes. Intervention tests look for evidence a controlled change altered behavior. Process tests ask whether a monitor can identify the steps the model actually took when only a few valid solution paths exist. Outcome property tests check whether a monitor can predict a measurable property, like reward hacking, from reasoning versus actions alone. The headline result is that monitoring chain of thought generally beats monitoring outputs and actions. Longer thinking tends to improve monitorability. Current frontier scale reinforcement learning does not appear to meaningfully degrade it. 

There is a catch. OpenAI flags a tradeoff where a smaller model run with higher reasoning effort can reach similar capability while being easier to monitor, but it costs more inference compute. It calls that compute premium a monitorability tax. The company also found a practical boost. Ask follow up questions that push the model to reflect on what it did, then monitor that extra reasoning too. Deeper analysis on this phenomenon can be found at Olam News for a sharper perspective.