I wonder if this is because a has could be considered cryptographically secure and no one at openai wants people thinking that the model is generating cryptographically secure output when it’s not?