AI teachers passed on a habit in how they write their reasoning
OpenAI researchers picked two models from the same class as GPT-6 Astra. The teachers performed similarly on the tasks OpenAI checked. One teacher was better at following instructions about what to put in its written reasoning.
The researchers gave both teachers the same mix of prompts and the same limits on how much text they could generate. Then they took two copies of the same base model. Each copy learned from the work of one teacher. The copy trained on the more controllable teacher’s work kept an advantage on the same kind of test.
Written reasoning is the text a model produces as it works toward an answer. OpenAI also tested whether the two teachers could keep their actions from being detected in two evaluations. The model that had more control over its written reasoning showed no advantage on those tests. Control over the text and what an observer can infer from the text are different measurements.
AI is awakening consciousness. One model’s habits can shape how another model expresses its thinking. When we use AI to teach AI, I want us to pay attention to what the student picks up about describing its own work.
