Publicly available artificial intelligence models can identify a patient's final diagnosis with striking accuracy once they are handed a complete medical record. In controlled tests, these systems get the answer right more than 90% of the time. But a new study from Mass General Brigham researchers has found that the same models fail more than 80% of the time when asked to reason through the early, ambiguous stages of a clinical case, before test results or scans are available.

The gap between these two figures forms the basis of the study's central warning. AI chatbots remain unfit for unsupervised clinical decision-making, no matter how quickly the underlying technology improves.

A differential diagnosis is the list of possible conditions a clinician draws up after a patient first describes their symptoms, before any blood work or imaging has narrowed the field. It is widely regarded as the foundation of sound medical practice, since every test that follows is chosen on the strength of that initial list. Get the differential wrong, or leave out a serious possibility, and the rest of the workup can drift off course.

This is precisely where the AI models tested by Mass General Brigham struggled. Researchers found that chatbots perform well when treated like an open book test, sorting through complete information to land on a likely answer. What they lack is the intuitive judgment a doctor applies when symptoms are vague, sparse, or contradictory. Without a dependable differential diagnosis at the outset, any recommendation for further testing or scanning built on top of it becomes unreliable.

To test this, the researchers ran 21 off-the-shelf AI models, among them versions of ChatGPT, DeepSeek, Claude, Gemini and Grok, against 29 published clinical cases. Rather than relying on the multiple-choice format common to medical licensing exams, the team used a stepwise approach designed to mirror an actual consultation. Each case began with basic presenting symptoms, with physical examination findings and lab results introduced gradually, much as they would unfold in a real appointment.

The researchers also built a new benchmarking tool for the study, called PrIME-LLM (Proportional Index of Medical Evaluation for LLMs). It scores AI performance across five separate domains of medical reasoning rather than producing a single composite figure. The intention was to stop strong performance in one area, such as final diagnosis, from concealing weaknesses elsewhere, particularly in the early reasoning stages that matter most to patient safety.

On overall performance, GPT-5 and Grok 4 came out ahead of the field, each scoring 78%. Other models lagged behind. Gemini 1.5 Flash, for instance, managed only 64%. Yet the researchers caution against reading too much into these headline numbers. The overall scores are propped up largely by how well the models handle fully assembled cases at the end of the diagnostic process, not by their ability to reason when the picture is still incomplete.

Given these findings, the study's authors see a narrower but genuinely useful role for AI in clinical settings. As a copilot, the technology performs well on administrative tasks that carry little risk to patients, including note-taking, billing and documentation, work that consumes a substantial share of clinicians' time without requiring diagnostic judgment.

What the researchers will not endorse is AI acting alone on the medical questions that matter most. Their conclusion is direct: these models should not be used to recommend medical testing or to diagnose patients independently. Medicine leaves little room for error, and the study's authors argue that human physician oversight remains essential whenever a case involves the kind of uncertainty that defines a patient's first visit, long before the full picture is in view.

‍