AI language models often respond to embarrassing questions with falsehoods.