Researchers have found that if you tone down a large language model's ability to lie, it's far more likely to claim that it's ...