Reporting on OpenAI’s safety analysis describes how pretraining capability, architecture, reinforcement learning and distillation can erode reasoning monitorability. Models become more capable of solving tasks without written reasoning or controlling how traces are expressed, and teacher characteristics can pass to students. This is linked to capability and compute-cost choices. However, models with greater trace control did not consistently evade monitors more often than comparison models in practical deception tests, so reduced transparency is not proof that every response is deceptive.
Back to Tech
OpenAI analysis traces declining reasoning monitorability through model training
Reading a reasoning trace alone is incomplete oversight, and capability and cost choices also affect transparency.