A new study of 228 US critical care physicians asked which AI advice format produced the best sepsis treatment decisions. The clear "model suggests X, you can override" condition won at 56.3%, beating the respectful briefing format at 42.9% and the always-follow baseline at 31.8%.

My reaction

I have sat in that product review. The AI lead sketches three advice formats on the whiteboard. Strict default, respectful briefing, simple recommendation with override. Nine times out of ten the room picks the respectful briefing. It sounds like the ethical choice. It sounds like the one that trusts the expert.

Sivaraman and colleagues at Carnegie Mellon, Hopkins, UCSF and Pitt put a number on what happens when you ship that choice. In complex cases, the respectful briefing is 13.4 points worse than the plain recommendation. The anchor you refused to give the clinician is the anchor they needed to think.

The scene

A head of underwriting at a specialty insurer in Zurich walks me through her wireframes last month. Her team is building an AI assist for complex casualty submissions. Three advice patterns on the Figma file. She leans toward option two. "I want them thinking, not clicking."

I ask her what option two does when the underwriter is tired and the submission is messy. She says it shows context and risk factors. No recommendation. "That way we respect their judgment."

I show her the abstract on my phone. Across 228 physicians and six complex sepsis decisions, the plain recommendation hit 56.3% optimal when the physician could override. The context-only format hit 42.9%. The always-follow baseline sat at 31.8%. She reads it twice.

That is the mechanism. The plain recommendation is a thing to push against. The respectful briefing is a wall of context the brain treats as background. The deliberation she wanted to protect does not fire without a hook.

Her product manager was defending option two on exactly the grounds the study attacks. Respect clinical judgment. Avoid automation bias. Both intuitions are right about what you want. Both are wrong about which format produces it.

The retrospective arm matters too. On 6,928 sepsis patients the model was better than clinicians on average and worse on the hard cases. Where the AI is weakest is where the format decision is a 24.5 point lever. That is where the override UI becomes the product.

"Physicians selectively adapted and improved weak advice, likely because it prompted them to reconsider their assessment."

She rebuilt the mock the next week. Plain recommendation. Two buttons, accept and change.