And there is little evidence to suggest that AI models are less likely to display worrying behaviour as they become bigger and more powerful; indeed, the opposite seems to be the case. What to do?