If it's trained on data generated by humans, then all bets are off. Neither "alignment" fine-tuning nor output filters are likely to be 100% effective, and a s

If it's trained on data generated by humans, then all bets are off. Neither "alignment" fine-tuning nor output filters are likely to be 100% effective, and a single failure can be disastrous.