To be fair, you can't control what humans optimize what you're trying to teach them either. A lot of times they learn the opposite of what you're trying to teach them. I've said it before but all they managed to do with LLMs is make a computer that's just as unreliable (if not moreso) than your below-average human.