Fuck AI

Depends on jurisdiction. In the States, some States allow more than one party to be liable. But others are single liability States.

But basically, if you market a product as being able to do specific tasks and the person using the product uses it that way, and it kills someone or harms them, I think it comes down to who is the person or entity who's actions directly lead to the event/could have been forseen to lead to such an event. Tesla has a lot of data on their vehicles and every day there seems to be a new report about what they knew about the safety of the "self driving" features. Specifically things like repainting the lines on roadways to get the results they wanted and potentially fabricating data or modifying it to present their vehicles as having specific capabilities. But that doesn't mean the driver is off the hook.

So for instance the person behind the wheel could argue that they didn't know that the vehicle could act on its own to leave the roadway and crash into a home because it's marketed to have safety features to prevent or mitigate a collision. And the court would have to decide if they indeed had time to slam on the brakes or take over the wheel. If they did have time and they didn't do anything to try to prevent the crash they are definitely on the hook.