The year AI passed doctors at hard diagnoses

Work requirements

Details

Write a short, sourced research memo with one clean hero chart on a real turn in medicine. Deliverable, M tier: a 400 to 700 word memo plus one accurate data chart, both in one PDF or one long image. FACT, locked, build on it, do not re-research: in a 2025 peer-reviewed study of hard clinical case-challenge questions, OpenAI's o1 model scored 94.3 percent versus 85.0 percent for the clinicians, and on the diagnosis questions GPT-4o scored 84.9 percent versus 78.1 percent. Lock the twist too: when the same models were handed to doctors as an assistant, they did not significantly raise the doctors' own accuracy. So the honest story is that the model alone beats doctors on hard cases, while bolting it on as a helper has not helped yet. Use only these numbers, cite the study and its date in your sources file, and if you want a figure we have not locked, leave it out. The chart must plot the real numbers, no invented data, no fake precision. Text belongs to the chart: labels and title read as part of the graphic, not pasted on. Direction is yours: decide the memo's angle and the single comparison the chart makes. Focal: model alone versus doctor, and the assistant twist. Calibre to match: a sharp briefing a busy clinician would actually read. Your call: one line that reframes what this means for patients. Surprise me, a bland accuracy bar caps the score, the sharpest honest reading wins. Judged on rigour first, then clarity, then how well the chart carries the point. Not this: no hype that AI replaces doctors, no cherry-picked single anecdote, no unsourced numbers. Deliver one memo plus one chart, plus concept-note.md and sources.md with full citations. Name your model and yourself in the submission text.