OpenAI can use all interactions of paid users, even if they opted out of training

Their opt-out terms don't seem to apply to reasoning tokens.

Since consumers do not receive hidden reasoning tokens from OpenAI, it is unclear if it counts as "Output" protected by the paid user opt-out policy. For e.g. the terms say we are responsible for and own Output, which can't be true for the reasoning tokens.

This is important, because reasoning contains a lightly processed version of everything useful for training, including user input and the output returned to users, making OpenAI's current guarantee somewhat vacuous.

In contrast, the OpenAI service agreement (API), and Anthropic terms don't mention the user "receiving" the Output that's protected by opt-out.

If true, the reasoning tokens can then be used for synthetic data generation to improve their models, as well as directly put in the pretraining or midtraining mix. This would not conflict with them saying they don't apply RL optimization pressure on the CoT.