You seem to argue that, on the unproven premise that current AI is better than human drivers, we should let corporations test it out in the real world even if they are not criminally liable ever

I'm arguing on the assumption that it is proven.

Until it's proven, the driver takes the responsibility if the corporation doesn't, and insurance costs should reflect that. There are reasons I don't own a car equipped with self-driving features, and this is one of the big ones, it's unproven.

But as long as our approach to it continues to be profit over people, socialise the risk

We've gotten really far with prioritizing profit, but I agree that socializing the risk is a big problem. However, criminal acts generally require motive, so we're unlikely to see actual jail time without provable, malicious intent.

So I think we should do the next best thing: fine them. Increase the fines for each infraction in a given year until the problem is fixed. Force them to continue to improve.