AI Outperforms Doctors in Emergency Triage Trial

In an emergency-triage trial involving Harvard, AI's judgment accuracy surpassed that of doctors. The conclusion grabs eyeballs, but the trial's boundary conditions and the distance to deployment are worth spelling out just as much.

What the Trial Did

Triage is the first gate of the emergency room: judging who is critical and who can wait, where a wrong call costs lives. The trial compared AI's triage suggestions with medical staff's actual judgments, and AI won on the accuracy metrics. This isn't absurd—triage is essentially pattern recognition under limited information, precisely machine learning's comfort zone, while real human decisions in the ER are also disrupted by fatigue, crowding, and differences in experience. Similar results have appeared repeatedly in imaging diagnosis, and triage is just the next.

How Far from Paper to the ER

But between "won in the trial" and "should be handed to AI" lie several hurdles. The data quality of a trial environment is worlds apart from the chaos of a real ER; the model's performance on edge populations and its sensitivity to rare emergencies both need longer-term validation; and who bears responsibility when it errs has no legal answer yet. A more realistic deployment form is assistance rather than replacement: AI gives a suggestion and reasons, and the triage nurse keeps veto power. Progress in medical AI is real; it's just that its pace is usually measured in years—no rushing it.

via: Hacker News