AI Beats Doctors on Diagnosis, But Regulators Aren't Convinced

A Landmark Study, With an Asterisk

A new study published in Science has given the “AI outperforms doctors” narrative its most rigorous evidence yet, while also revealing why that narrative oversimplifies what actually happened. Researchers from Harvard Medical School, Beth Israel Deaconess Medical Center and Stanford pitted OpenAI's o1-preview reasoning model against hundreds of physicians across six experiments, and the model matched or beat human performance in nearly every one 891213. The result has been described as a milestone in clinical AI, but nearly every outlet covering it paired the finding with a caution: this is not the same as AI replacing doctors 111517.

The most striking piece of the study examined 76 real, unedited emergency-department cases from Beth Israel Deaconess, reconstructed at three points of care — initial triage, physician evaluation and hospital admission 101213. At triage, when only vitals, demographics and a brief nursing note were available, the model identified the exact or a very close diagnosis in 67.1% of cases, compared with 55.3% and 50.0% for two attending physicians 101114. By the time of admission, when far more clinical detail existed, the model's edge narrowed to roughly 82% versus 70–79% for physicians, a gap that was no longer statistically decisive 1112.

Beyond the Emergency Room

The study did not stop at triage. On 143 published New England Journal of Medicine clinicopathological-conference cases, the model reached 88.6% diagnostic accuracy compared with 72.9% for the older GPT-4 model 910. On the NEJM Healer training exercise, it earned a perfect reasoning score on 78 of 80 responses, dwarfing GPT-4 (47 of 80), attending physicians (28 of 80) and residents (16 of 72) 1213. When asked to devise treatment and management plans — arguably a harder task requiring judgment about context and competing priorities — the model scored a median of 89%, compared with just 34% for physicians relying on conventional tools like search engines 111213. In one striking anecdote, the AI caught that a patient's lupus history, not failing anticoagulants, explained worsening lung symptoms — a connection the treating physicians had missed 11.

To guard against the possibility the model had simply memorized published cases, the researchers also tested it against six previously unpublished vignettes from a 1994 study. It scored a median of 97%, versus 92% for GPT-4 and around 74–76% for physicians using either GPT-4 or conventional resources, though the small sample size kept the comparison from reaching statistical significance 1213.

Praise, but With Guardrails

Independent experts quoted across the coverage largely agreed the study represented genuine progress rather than hype. A University of Edinburgh medical informatics researcher called the systems credible “second-opinion tools,” while a Mount Sinai chief clinical officer described the paper as a vivid illustration of how far the technology has advanced 1115. Stanford's Jonathan Chen, a co-author, called the results “humbling,” noting that AI now sometimes outperforms not just doctors but doctors using AI assistance themselves 17.

But the study's authors were emphatic that their findings do not license replacing physicians. Harvard's Arjun Manrai said repeatedly that the results reflect “a really profound change in technology that will reshape medicine,” not a case for eliminating clinicians 111517. Co-author Adam Rodman described a future “triadic care model” combining doctor, patient and AI system, rather than a doctorless one 11. NPR initially framed its own headline around AI beating “ER doctors,” then issued a correction clarifying that the comparator physicians were internal-medicine specialists, not emergency-trained ones 15 — a distinction the American College of Emergency Physicians and Science News also flagged, noting that emergency physicians receive specialized triage training that the study's comparators may not have had 16.

The Limits Baked Into the Design

Several structural caveats ran through the coverage and the published critiques that followed. The evaluation was entirely text-based; the model never examined a patient, heard their voice, observed distress or interpreted an image, all of which matter in real triage 111517. Critics writing in Science's eLetters warned that curated, “cleaned up” case text likely flatters AI performance relative to the “messy” multimodal reality of emergency medicine, and that historical-control comparisons carry confounders that head-to-head trials would eliminate 8. Others pointed out that diagnosis is only one narrow slice of what benchmarks can measure, and that performance on a written vignette says little about safe operation inside noisy, time-pressured hospital workflows 89. A related Science commentary noted that funding for the research came in part from AI-adjacent philanthropic and institutional sources, underscoring the need for independent replication 8.

That caution is reinforced by findings elsewhere. A University of Virginia trial of 50 physicians found that doctors using ChatGPT Plus reached a median diagnostic accuracy of 76.3%, only marginally better than the 73.7% achieved with conventional tools like UpToDate and Google — and, notably, adding a human to the AI's output sometimes reduced accuracy compared with AI alone 18. Two independent meta-analyses paint an even more tempered picture: one pooling 83 studies found generative AI's overall diagnostic accuracy at 52.1%, with AI performing significantly worse than expert physicians even though it matched non-experts 19; another pooling 54 studies put AI accuracy at 57%, with physicians outperforming AI models by 14 percentage points on average 20. Both reviews flagged high risk of bias in the underlying literature and substantial variability across medical specialties.

Regulation Racing to Catch Up

The diagnostic breakthrough lands amid a broader, unsettled fight over how AI should be governed generally. In the United States, the federal posture has leaned toward a light-touch approach, with officials pressing G20 counterparts to avoid dedicated AI regulators and to embrace so-called “Carolina Principles” favoring innovation over binding rules 56. That stance stands in sharp contrast to the European Union's AI Act, a comprehensive and legally binding framework explicitly designed to constrain risk 4. Domestically, the deregulatory push has run into resistance: state legislators, buoyed by public concern and new funding for pro-regulation advocacy, have increasingly pursued their own AI rules even as industry lobbyists retreat from fighting them state by state 32.

For healthcare specifically, the FDA already regulates AI through existing device pathways — 510(k) clearance, De Novo classification and premarket approval — and has authorized more than a thousand AI-enabled devices, the overwhelming majority of them narrow imaging and signal-analysis tools rather than open-ended diagnostic chatbots. That regulatory architecture assumes a relatively fixed product, which sits uneasily with generative models that can behave differently as prompts, guardrails or underlying versions change — a mismatch regulators are now trying to address through draft guidance built around lifecycle monitoring rather than one-time approval.

What Comes Next

A recurring theme across the coverage, including a Forbes reflection tracing today's standards back to a 1968 paper on preserving clinical reasoning rather than just data, is that medicine has long tried to formalize what counts as sound diagnostic thinking — and AI is now being measured against that same yardstick 7. The consensus among researchers, physicians and outside experts is that the Science study marks a genuine inflection point in benchmark performance, not proof that AI is ready for autonomous clinical deployment 91417. The next real test, most agree, is prospective clinical trials that track patient outcomes, missed diagnoses, unnecessary tests and clinician trust — not just whether a model can name the right disease on paper.