the hidden chain of thought allows us to “read the mind” of the model and understand its thought process. For example, in the future we may wish to monitor the chain of thought for signs of manipulating the user.
If they don't train on it, they are definitely reading the reasoning tokens.https://developers.openai.com/blog/responses-api