If you have the full outputs, it might make it easier for competitors to distil the model or reverse engineer the full process.

It may also be that misaligned responses can be in CoT which OpenAI does not want to show to users.