A new benchmark built from 5,620 real patient-doctor consultations across 64 hospital departments finds that frontier AI models can match human physicians on positive clinical steps, yet trigger more unsafe, unsupported, or contradictory responses in the process [S1]. The gap that separates AI from doctors is not what the models do right. It is what they fail to avoid.

Why existing medical AI benchmarks fall short

Large language models are increasingly deployed in online medical consultation, particularly in China's internet hospital system [S1]. But the benchmarks used to evaluate them have a structural problem: many rely on synthetic conversations or patient simulators rather than real interactions [S1]. They tend to omit patient-uploaded medical images, the photos of skin conditions, wound scans, and test results that real consultations hinge on [S1]. And they grade open-ended clinical responses using multiple-choice or lexical-overlap metrics, which measure whether the model's words overlap with a reference answer rather than whether the advice is clinically sound [S1].

A model can score well on a multiple-choice medical exam while giving dangerous advice in an actual consultation. The gap between test performance and real-world competence is exactly what MedRealMM, released July 13 on arXiv, sets out to measure [S1].

How MedRealMM works

Researchers constructed the benchmark using anonymized interactions between patients and doctors from a national internet hospital in China [S1]. Instead of utilizing all interactions, the team employed a Multimodal Clinical Challenge Point (MCCP) extraction framework to pinpoint the moments in a consultation where clinical reasoning becomes highly complex [S1]. These moments are transformed into a standardized next-response generation task, retaining the prior text and image context so the AI receives the same information the doctor did [S1].

Physicians refined a rubric for each case, which rewards clinically beneficial actions and penalizes unsafe, unsupported, or contradictory replies [S1]. This marks a shift from traditional scoring: rather than checking if the model generated the correct answer, it evaluates whether the model took the right actions and avoided harmful ones.

The present release includes 5,620 cases across 64 clinical departments [S1]. The research assesses 19 general and medical-specialized LLMs, encompassing both text-only and multimodal architectures [S1]. The dataset is scheduled for public release on Hugging Face, although the repository page currently displays an empty README [S1][P4].

The finding buried in the results

Two results stand out, and the second one matters more.

First, visual data is essential for dependable clinical performance [S1]. Models that can process images alongside text do better on cases where patients have uploaded photos. Text-only models are working with one hand tied behind their back.

Second, current leading models still fall short of the overall online physician response [S1]. However, the nuance is significant. Certain leading models meet as many or more positive clinical criteria than doctors [S1]. They pose the correct questions and explore the proper diagnostic territory. Yet they also activate more negative criteria than doctors, resulting in more unsafe replies, unsupported assertions, and contradictions within a single response [S1].

Avoiding safety-sensitive errors remains the primary obstacle [S1]. The models are not failing because they lack knowledge. They are failing because they do not know when to stop.

What it means

For those monitoring AI in healthcare, MedRealMM shifts the focus. Models can already equal a doctor's knowledge in certain instances, at least regarding positive actions. Whether they can match a doctor's restraint is another matter, and the evidence indicates they cannot yet. A doctor who provides five correct pieces of advice and one dangerous one has still delivered dangerous advice. The rubric reflects this by scoring both aspects simultaneously.

This aligns with a wider trend in AI evaluation. MedRealMM identifies the failure points in actual consultations, rather than the simple successes in artificial ones.

The benchmark also disputes the idea that multimodal capability is optional in medical AI. If visual data is essential for dependable performance, any deployment that removes images from the process accepts a lower limit on clinical quality.

What it means for business

For health-tech operators in China and elsewhere, MedRealMM provides what current benchmarks lack: a test that mirrors what truly occurs when a patient sends a doctor a photo and a complaint. A small telehealth startup assessing which LLM to use for its consultation assistant can now examine performance on cases with patient-uploaded images, evaluated against physician-refined rubrics that penalize unsafe replies, rather than just multiple-choice accuracy.

The finding regarding negative criteria has a direct operational consequence. If a model equals doctors on positive steps but triggers more unsafe replies, the deployment question is: what guardrails intercept the unsafe answers? A local medical practice testing an AI triage tool would need to prioritize error avoidance over the coverage of correct advice. The benchmark indicates that current models require human-in-the-loop review specifically for safety, not for completeness.

For model developers, the 5,620-case dataset, once publicly accessible on Hugging Face [S1], provides a training and evaluation surface that includes actual clinical images and real consultation context. Teams developing medical AI can evaluate against physician-level rubrics instead of lexical overlap, meaning the feedback loop finally mirrors clinical quality.

What we don't know yet

The paper is an arXiv preprint and has not been peer-reviewed [S1]. Findings should be considered provisional.

The national Chinese internet hospital that provided the data is not identified in the paper, which restricts transparency and reproducibility [S1]. The de-identification process for real patient-doctor interactions cannot be independently verified from the source.

The Hugging Face repository exists but the README is empty at the time of writing [P4], and the paper states the dataset "will be" publicly available [S1]. It is not yet clear when the full dataset will be downloadable or under what license.

The 19 evaluated models are not all named in the available source material, so specific model-by-model comparisons cannot be reported here. The performance comparisons depend on the authors' own physician-refined rubric, which may contain framework-specific bias. A different rubric might yield different rankings.

The next concrete event to monitor is the public release of the dataset on Hugging Face, which will enable independent researchers to run the benchmark and verify the findings. Until then, the central claim, that AI matches doctors on doing right but fails at avoiding wrong, is a signal worth tracking, not a settled verdict.

If this kind of evidence-first reporting is what you want in your feed, subscribe to keep reading.

Sources

More from Not A Tech Guy

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.