I think the solution is to ask a different question, namely: Where, in the pipeline of causal events, are we best poised to intervene, such that we most effectively mitigate future damage?

Even today, society actually reaps no benefit from punishing people because they "deserved" it. But society does derive some benefit from the fact that we quarantine (i.e. imprison) people who are statistically more likely to commit crimes (e.g. by virtue of them having committed a crime previously).

We should stop thinking of crimes as deeds to be punished, and instead, look at them as potential sources of problems that might warrant some form of intervention.

This is a problem that exists even without conscious AI. But I agree that the emergence of sentient AI would shine a light on this issue, and force us to start thinking of crime and punishment more in the terms I've described above.