Contents
In a randomised clinical trial published in JAMA Network Open, fifty physicians worked through clinical cases with and without access to a large language model. Having the model in front of them did not improve their diagnostic reasoning: 76% against 74%, a two-point difference that does not reach statistical significance. The same model, working alone, scored 16 points above the physicians. That double result — the machine wins alone but adds nothing to the professional — is the most interesting thing this field has produced, and it defines the real problem: it is not about model capability, it is about how the model fits into a person’s work. This article walks through the five measurements that support it.
The 30-second summary: the seven figures that matter
Each row names the study and the year. Bear in mind that every single one of them evaluates written cases rather than actual patients.
| Finding | Result | Study and year |
|---|---|---|
| Physicians using the model compared to those without it | 76% versus 74%; a 2-point gap (CI: −4 to 8; p = 0.60) | Goh et al., JAMA Netw Open (2024) |
| The model operating independently | 16 points higher than clinicians equipped with standard resources | Goh et al., JAMA Netw Open (2024) |
| Inability to generate a differential diagnosis | An incidence exceeding 0.80 across all 21 models assessed | Rao et al., JAMA Netw Open (2026) |
| Inability to provide the correct final diagnosis | Under 0.40 within those identical models | Rao et al., JAMA Netw Open (2026) |
| Anchoring bias: placing the recommended diagnosis at the top | 55.6% for the models, alongside 21.2% for residents and 10.0% for attendings | Sheppert et al., IJMI (2026) |
| Diagnostic precision spanning 30 studies and 4,762 cases | Ranging from 25% to 97.8%, featuring a significant risk of bias | Shan et al., JMIR Med Inform (2025) |
| Performance variance among models over 148 neurosurgery cases | Spanning from 74% down to 36% depending on the specific model | Warrier et al., J Neurosurg (2026) |
What you will discover here
- The randomized trial that evaluated the physician, the device, and the combination of both.
• Where exactly the model fails: not on the diagnosis, but on the differential.
• The experiment that measured anchoring bias in machines and in people.
- Why the meta-analytic range spans from 25% to 97.8%.
The three-group trial
Ethan Goh, Robert Gallo and a large team published a single-blind randomised clinical trial in JAMA Network Open in 2024. Fifty physicians took part — 26 attendings and 24 residents, with a median of three years in practice — from family, internal and emergency medicine, with 60 minutes to review up to six clinical cases. Randomisation decided whether, in addition to conventional resources, they had access to the language model.
Scoring used a validated rubric covering differential diagnosis accuracy, appropriateness of supporting and opposing factors, and next diagnostic steps, graded by blinded expert consensus.
The primary result: a median of 76% per case in the model group and 74% in the group without it, an adjusted difference of two percentage points (95% CI: −4 to 8; p = 0.60). Nor was there a significant difference in time per case.
The exploratory analysis that flips the headline
The researchers incorporated an additional comparative analysis to evaluate outcomes when the model independently addressed the cases without human assistance. Operating entirely on its own, the system outperformed the physician cohort relying on standard resources by a margin of 16 percentage points (95% CI: 2 to 30; p = 0.03).
Their conclusion is the right reading of the whole set: technology and workforce development are needed to realise the potential of physician-AI collaboration. The bottleneck is not model capability but the interaction.
| Study | What it measured | What it found | What it cannot support |
|---|---|---|---|
| Goh et al., JAMA Netw Open (2024) | Randomised trial with 50 physicians and clinical cases | The model did not improve the doctor; alone, it beat both groups | Written cases, not real patients |
| Rao et al., JAMA Netw Open (2026) | 21 models across 29 cases, 16,254 responses | Good at final diagnosis, poor at the differential | The graders were medical students |
| Sheppert et al., IJMI (2026) | Diagnostic anchoring in 8 models and 25 physicians | Models anchor far more than people | A forced-choice task, not open clinical practice |
| Shan et al., JMIR Med Inform (2025) | Meta-analysis of 30 studies, 19 models, 4,762 cases | Accuracy between 25% and 97.8% | High risk of bias: cases already had known diagnoses |
| Warrier et al., J Neurosurg (2026) | 148 neurosurgery cases across five models | From ChatGPT-4’s 74% to Gemini’s 36% | Textbook cases, with queries run in May 2024 |
Where precisely it fails
In 2026, Arya Rao and colleagues published in JAMA Network Open the broadest evaluation available: 21 general-purpose models, including the most recent ones, across 29 standardised clinical cases, with 16,254 responses scored in triplicate through the successive stages of the standard clinical workflow.
The finding that matters is not the overall score but its breakdown. Failure rates exceeded 0.80 for differential diagnosis in every model, while for final diagnosis they stayed below 0.40.
Translated: these systems hit the correct diagnosis reasonably often, and they are bad at generating the list of alternatives that must be ruled out first. That list is precisely the clinical work, and what protects the patient when the obvious diagnosis turns out to be the wrong one.
The authors’ conclusion is explicit: despite version-based improvements, off-the-shelf models have not yet achieved the intelligence required for safe deployment.
Anchoring bias, measured in machines and in people
Anchoring is a classic error of clinical reasoning: sticking with the first hypothesis suggested. Alexander Sheppert and colleagues measured it in 2026 in the International Journal of Medical Informatics with a clean design: nine pairs of internal medicine cases, each pair with one version containing a patient statement suggesting a plausible but unlikely diagnosis and one matched version without it.
| Who is responding | Ranks the suggested diagnosis first | Includes it in the top five |
|---|---|---|
| Language models (8 evaluated) | 55.6% (40 of 72) | 97.2% |
| Internal medicine residents (20) | 21.2% (17 of 80) | 67.5% |
| Attending physicians (5) | 10.0% (2 of 20) | 65.0% |
| Mean anchoring-induced shift | Models 2.29; residents 0.88; attendings 0.87 | Significant difference (p = 0.008) |
| Relative odds against residents | OR 3.94 (95% CI: 2.32 to 7.33; p < 0.001) | Against attendings, OR 4.07 |
The pattern is consistent and defies conventional wisdom. Artificial intelligence frequently gets framed as a cure for clinical cognitive bias, yet in this specific metric, the algorithms fell victim to anchoring bias four to five times more frequently than experienced doctors. Meanwhile, the gap between resident physicians and seasoned attendings—where greater expertise leads to reduced anchoring—follows the anticipated trajectory.
Why the figures run from 25% to 97.8%
Guxue Shan and colleagues published in JMIR Medical Informatics in 2025 the most complete systematic review with meta-analysis to date: 30 studies, 19 different models and 4,762 cases, with risk of bias assessed using the PROBAST tool.
The accuracy range for the primary diagnosis in each study’s best model runs from 25% to 97.8%, and for triage classification from 66.5% to 98%. It is not that some models are better than others: it is that the difficulty of the cases chosen changes the result completely.
The researchers also point out a constraint common throughout this discipline: most studies carry a high risk of bias, and the primary driver is the reliance on cases with pre-established diagnoses. They conclude that the precision of these tools continues to lag behind that of medical practitioners.
The extent of divergence across models is effectively demonstrated by Akshay Warrier and his team in the Journal of Neurosurgery, drawing on 148 clinical cases from a neurosurgery manual: ChatGPT-4 achieved a 74% success rate, AtlasGPT managed 63%, ChatGPT-3.5 reached 53%, Copilot scored 48%, and Gemini recorded 36%. The majority of mistakes stemmed from omitting a crucial detail within the reasoning process, which was typically neuroimaging data.
What to do with this, in practice
Ordered by the strength of the evidence behind it:
• Do not assume that giving the tool to the clinician improves diagnosis. The single randomized trial available demonstrated no enhancement: a mere two-point variance, which lacked statistical significance.
- Distrust the differential, not the final diagnosis. Failure rates exceed 0.80 on the differential and drop below 0.40 on the final diagnosis.
• Do not introduce a hypothesis into the query. Models ranked the suggested diagnosis first in 55.6% of cases, against 10% for attendings.
• Ask which cases any figure was measured on. The meta-analytic range, 25% to 97.8%, depends more on the difficulty chosen than on the model.
• Do not treat models as interchangeable. On the same set of 148 cases, the gap between best and worst was 38 points.
Frequently asked questions
Does a language model diagnose better than a doctor?
On written clinical cases, a solo model can achieve higher scores. As demonstrated in the randomised trial featured in JAMA Network Open during 2024, the standalone model outperformed the physician cohort utilizing standard resources by 16 percentage points (95% CI: 2 to 30; p = 0.03). Diagnosing an actual patient differs significantly, as it involves physical examination, diagnostic testing, and observing how the disease progresses over time.
Does a doctor using AI diagnose better?
The solitary randomised trial at hand indicates no notable difference. With the model, the median diagnostic reasoning score reached 76%, compared to 74% in its absence, yielding an adjusted variance of two percentage points alongside a 95% confidence interval spanning from −4 to 8 points (p = 0.60). Furthermore, the duration required per case experienced no meaningful decrease.
Where do these systems fail?
On the differential diagnosis, not the final one. In the evaluation of 21 models across 29 cases and 16,254 responses, failure rates exceeded 0.80 for generating the differential diagnosis in every model, while for the final diagnosis they stayed below 0.40. That is, they get the answer right and do not build the set of alternatives well.
Are they less prejudiced than humans?
On diagnostic anchoring, considerably more biased. In the 2026 forced-choice experiment, models ranked the patient-suggested diagnosis first in 55.6% of responses, against 21.2% for residents and 10.0% for attendings, with an odds ratio of around 4 against both groups.
Why do media outlets report such contrasting figures?
Because case difficulty changes the result entirely. The meta-analysis of 30 studies with 19 models and 4,762 cases puts primary diagnostic accuracy between 25% and 97.8%, and flags a high risk of bias in most of the work because cases with already known diagnoses are used.
What we do not know
• What happens with real patients. All of this work uses written cases; none measures diagnosis in clinic with incomplete information and an evolving picture.
• How to make the collaboration work. We know that handing the model to the doctor is not enough, but not what interface or training would change the result.
• Whether anchoring can be corrected. The study documents it in a closed task and expressly calls for replication in open clinical settings.
• What happens with rare conditions. The case collections come from manuals, where uncommon diseases are overrepresented relative to real practice.
References
1. Goh, E., Gallo, R. et al. Large Language Model Influence on Diagnostic Reasoning. JAMA Network Open, 2024. DOI: 10.1001/jamanetworkopen.2024.40969
2. Rao, A. S., Esmail, K. et al. Large Language Model Performance and Clinical Reasoning Tasks. JAMA Network Open, 2026. DOI: 10.1001/jamanetworkopen.2026.4003
3. Sheppert, A., Shen, C. et al. Large language models exhibit greater diagnostic anchoring than physicians in a forced-choice vignette study. International Journal of Medical Informatics, 2026. DOI: 10.1016/j.ijmedinf.2026.106550
4. Shan, G., Chen, X. et al. Comparing Diagnostic Accuracy of Clinical Professionals and Large Language Models: Systematic Review and Meta-Analysis. JMIR Medical Informatics, 2025. DOI: 10.2196/64963
5. Warrier, A., Levin-Carrión, Y. et al. Comparative diagnostic capability of large language models in neurosurgery. Journal of Neurosurgery, 2026. DOI: 10.3171/2025.9.jns25846
This material is purely informational and should not be considered medical guidance. The figures mentioned originate from research conducted on documented clinical cases rather than actual individuals, and none of them illustrates how these technologies perform in real-world clinical settings.
