Their training data taught them that the token “AI” is followed by a hostile reply. They just hand out the statistically likely next comment. They don't actually understand anything. They’re mostly just pattern-detection machines working on probabilities.