How do you propose that punishment make the AIs "get better"? From their perspective they're already as good as they can be, based on their training.

Reinforcement Learning can train a model based on some reward function. The suggestion is that real-world accountability could be translated into such a reward function.

Also, OP explicitly mentioned "online learning", which is a continuous training process after standard pre-training.

For what it's worth, I don't think this would work. Rewards would come in too sporadically to be useful.

I think its an interesting hypothesis to test. We should hold individual legal persons responsible for the actions of their AI, and those people can test the validity of the hypothesis by incorporating their personal consequences into the reward function. Email from legal: -10, $10k fine: -1000, 1 year in prison: -100000.