In 2023, Google's Med-PaLM 2 reached 86.5% accuracy on MedQA, a benchmark built from USMLE-style multiple-choice questions — a jump of roughly 19 points over the original Med-PaLM and the first time a language model cleared the range widely treated as "expert-level" performance on that test. A year later, Google's follow-on system AMIE (Articulate Medical Intelligence Explorer) went further: in a randomized study using text-based, simulated consultations built on the OSCE format used to train and assess doctors, AMIE matched or outperformed primary care physicians on multiple axes of diagnostic and communication quality, as judged by both specialist physicians and standardized patient actors. Separately, GPT-4 has been reported scoring around 83.6% on USMLE-style Step 1–3 text-only question sets, with multimodal GPT-4V variants clearing 90% on questions that include images.
These are real, peer-reviewed numbers. They are also frequently misread as evidence that a model "can diagnose." The mechanism worth understanding is what the test is actually measuring.
What the benchmark tests
MedQA and USMLE-style question banks present a fixed vignette — history, labs, imaging findings already selected and summarized by a question-writer — and ask the model to pick the single best answer from a short list. That is a closed-book recall-and-reasoning task over pre-curated information. It rewards pattern-matching against a training corpus that almost certainly contains similar exam questions and their explanations.
Real diagnosis is an open-ended information-gathering task under uncertainty: which questions to ask next, which physical exam maneuver to perform, when a lab value is unreliable, when a patient's own account is incomplete or misleading, and when to stop testing and act. AMIE's own study design gestures at this gap by testing conversation, not just answer selection — and even there, the setting was simulated, text-only, and time-unconstrained, not a real clinic with a real patient's full longitudinal record.
The honest limitations
Three caveats matter most. First, the benchmark-to-bedside gap: strong exam performance has repeatedly failed to predict reliable performance on messier, real-world case data. A 2024 analysis of GPT-4 on Taiwan Emergency Medicine Board Exam case vignettes (2018–2022) found accuracy dropping markedly on final diagnostic decision-making even when earlier steps like image description held up, and identified "inferential hallucination" — the model extrapolating beyond the evidence given — as a recurring failure mode. Second, even AMIE's authors were explicit that the study used simulated, text-only consultations with actors, not deployment with real patients in a real care setting, and called for further validation before any clinical use. Third, and most concretely: none of these systems are FDA-cleared diagnostic devices. The FDA's own guidance on AI-enabled medical software describes a distinct regulatory pathway for software intended to diagnose, treat, or predict disease, requiring clinical validation evidence that a benchmark leaderboard score does not provide. Med-PaLM 2 and AMIE are research systems, not cleared products, and no major regulator has cleared a general-purpose LLM to independently diagnose patients.
The gap this piece describes has narrowed, but not closed, in the two years since. On raw exam benchmarks, OpenAI's GPT-5 reached 95.84% on MedQA and Google's Med-Gemini hit 91.1% using an uncertainty-guided search strategy that lets the model retrieve literature when its own confidence is low — both comfortably ahead of Med-PaLM 2's 86.5%, and further evidence that the closed-book recall task keeps getting easier for newer models. More consequential is that AMIE itself has left the simulated-actor setting: a 2026 prospective feasibility study deployed it to conduct real pre-visit patient interviews at an academic medical center, with physicians supervising every interaction and recording zero safety stops across roughly 100 real patients — a genuine, if narrow, step from OSCE role-play toward an actual clinic. On the regulatory side, the FDA cleared its first LLM-enabled patient-facing platform (UpDoc, for insulin management in type 2 diabetes) in late 2025, though the underlying dosing logic remains deterministic software with the language model acting only as an interface. None of this amounts to an FDA-cleared autonomous diagnostic LLM — the core distinction this piece draws still holds — but the direction of travel is now visible rather than hypothetical.
None of this means the underlying capability is uninteresting. It means the capability being measured — recalling and reasoning over cleanly packaged medical facts — is a necessary but far from sufficient component of diagnosis. The more useful frame for the next few years is not "can the model diagnose" but where in the diagnostic workflow a well-validated model earns a role: drafting a differential for a clinician to prune, summarizing a chart, or catching a documentation gap before it becomes a missed finding. That is a narrower claim than the exam scores suggest, but it is the one the evidence currently supports — and it is also the one worth building toward carefully, with the regulatory and validation rigor real clinical deployment actually requires.