That’s another angle: its training is heavily geared towards user satisfaction, without strong safeguards against malicious uses.