On Nov 12 a multi-center study started making the rounds. It was run by MedGPT, a future-doctor platform, together with thoracic surgery departments at several top Chinese tertiary hospitals, and according to New Intelligence (Xinzhiyuan) the result was blunt. Across 40 real difficult lung cancer cases, the MedGPT system beat 250 doctors from 98 medical institutions by an aggregate score of 88 to 12. The timing hardly hurt. Just a month earlier, Microsoft had reported in Science that its own AI diagnostic system hit 83% accuracy while the GPs in its control group managed 36%.
Not just any cases, either. The team pulled 40 so-called gray-zone cases from 100 complex lung cancer records, the kind where guidelines fall silent, late-stage cancer complicated by interstitial lung disease, say, or a post-surgery relapse still waiting on gene-test results. Before anyone saw an answer sheet, 20 top multidisciplinary experts wrote the scoring rules, 6 per case for 240 in total, and buried 40 safety red lines among them, any one of which fails a response outright. The human side was no pushover. More than 60% of those 250 doctors were chief or associate chief physicians spanning thoracic surgery, respiratory medicine and oncology.
The gap turned out larger than expected. Unaided, the doctors averaged 53.26 points; MedGPT took 85.29. Over 750 head-to-head comparisons the system produced the better plan 660 times, and its answers ran about 40% shorter than the doctors'. Put on the same cases side by side, MedGPT scored 96.22, leaving GPT-5.4 at 71.38, Claude Opus 4.7 at 61.81 and Gemini 3.1 Pro at 54.62 well behind.
An ablation study shows where those points came from. Strip the workflow and let the base model answer cold, and the score drops by nearly 24 points. Remove only the multidisciplinary consultation module and it costs 16 points. Under the hood, the system ground 13,000 lung-cancer papers into more than 200,000 structured evidence units, all under one hard constraint. Evidence may only be copied from source text by the program; the model may cite but never invent. That structurally locks out hallucination.
The assistive side was quantified too. With MedGPT's help, doctors' decision scores climbed from 53.4 to 66.27, a relative gain of 24.1%. Nearly 10 leading clinicians, among them Chinese Medical Association committee chairs and top US oncologists, gave positive reviews. In the clinic, though, the value of a system like this is not replacing doctors but lifting the floor on hard cases. What sets the pace of adoption is regulatory approval and liability rules, not leaderboard scores.