Well, yes, but remember there is the reinforcement learning that is applied after, and the system prompts that will bend the results.

Yeah, agree on both points. You can embed any bias you want using RL, regardless of training data.