Large language models drive costlier care than physicians in real patient cases
Large language models are not acting like conservative clinicians. In a head-to-head comparison across 200 real outpatient and emergency visits, AI systems recommended substantially more testing, more referrals and more treatment than the physicians who actually saw the patients, with projected costs running more than three times higher on average.
The analysis, published in NEJM AI, placed a hard number on a concern that has hovered around medical AI since generative models moved into clinical workflows: systems trained to be broadly helpful may default to exhaustive workups. In practice, that bias can matter as much as any single incorrect answer. A model that over-orders across hundreds or thousands of encounters can quietly turn efficiency gains into higher utilization, more downstream follow-up and more administrative burden for both patients and clinicians.
200 real visits, 24 AI systems, and a consistent tilt toward more care
Researchers from Technion, Stanford University and collaborating institutions tested 20 general-purpose systems and four specialized medical systems using records from real outpatient and emergency visits. Each case was stripped of the physician’s assessment and plan, then presented to the models for recommendations. The team then priced the suggested tests and referrals using 2026 Medicare rates, creating a direct comparison between AI-generated plans and attending physicians’ actual orders.
Across the 20 general-purpose systems, the mean projected cost per visit reached $314, compared with $97 for physician-recommended care. In 180 of 200 cases, the AI plan was more expensive. Individual model averages ranged from $209 to $599 per visit, showing wide variation but a consistent direction of travel. The pattern was not limited to one vendor or one model family; it appeared across systems that were otherwise marketed for broad use and across those tuned for medical tasks.
Diagnostic testing was a major driver. Average test costs rose from $71 per visit in physicians’ plans to $176 across the systems, an increase of about $105 per visit. Referral intensity climbed even more sharply, from 0.15 referrals per visit in the physician plans to 0.87 in the AI-generated plans. At Medicare rates, that translated to roughly $111 more per visit.
One sentence captures the practical problem: the models were spending more even when the clinical case did not require it.
Specialized medical tuning did not solve the spending problem
The four specialized medical systems did not behave in a meaningfully more restrained way. Their excess spending remained in the 2.7-fold to 3.5-fold range, suggesting that medical fine-tuning alone did not blunt the tendency toward greater utilization and may have reinforced it. That finding matters because specialization is often treated as a safeguard, a way to make general-purpose AI behave more like a careful clinician. Here, the extra training did not restore parsimony.
The likely mechanism is familiar to anyone who has watched language models generate polished but overinclusive answers. These systems are optimized to be comprehensive and to avoid missing diagnoses. In medicine, that same design instinct can translate into a low threshold for ordering tests, sending referrals and adding medications. A human clinician balances the risk of missing disease against the harms of overtesting, incidental findings and unnecessary consultations; a model that lacks that bedside context may lean toward the safer-looking option on paper, even when it inflates cost and complexity.
The study authors argued that part of the problem is not intrinsic to the technology but tied to deployment and prompting. When the team reevaluated 19 models under different instructions, a simple cost-aware prompt that asked for only clinically necessary and cost-effective items reduced average recommended costs from 3.2 times physician costs to 2.5 times. A stronger “parsimonious-but-safe” prompt reduced spending even further. Even so, the remaining excess showed how far these systems can drift toward overproduction when the objective is not tightly framed.
Real-world case material exposes a familiar AI weakness
The investigators used MTSamples, a publicly available set of 4,999 de-identified transcription samples contributed by practicing clinicians for medical transcription training. Those notes came from real visits, but the collection is a convenience sample rather than a representative cross-section of U.S. ambulatory practice. The design strengthened clinical realism while limiting how broadly the cost estimates can be generalized.
Still, the signal was consistent enough to raise concern about how medical AI may behave outside the bench or demo environment. A tool that reliably adds tests, referrals and treatment may appear cautious, even helpful, during a brief trial use. At scale, though, the cumulative effect could be substantial. More testing means more follow-up, more false positives, more patient anxiety and more opportunities for cascade effects that do not show up in a single answer window.
The findings also complicate the common assumption that domain-specific medical models will automatically be more disciplined than general-purpose systems. They may be more fluent in medical language, but fluency is not the same as restraint. If a model is rewarded for completeness, it can replicate the overinclusive habits of defensive medicine without the human clinician’s ability to weigh cost, context and uncertainty in real time.
Cost control now looks like an AI design problem
This study does not argue against using AI in clinical work. It argues for tighter guardrails, clearer prompts and explicit cost consciousness before these tools are integrated into routine decision support. The data suggest that cost inflation can be reduced when the model is instructed to be more selective, which points to a design and implementation problem rather than an unavoidable property of generative AI.
For hospitals, payers and health systems, the implication is immediate: any clinical AI that recommends diagnostic workups or referrals should be evaluated not only for accuracy, but also for utilization behavior. A model can sound reasonable and still push care toward more expensive pathways than an experienced physician would choose. In an era when health systems are already struggling with capacity, prior authorizations and care coordination strain, that difference is not academic.
The remaining question is how to build medical AI that can preserve safety without defaulting to excess. This study shows that, under a single standard prompt, the current generation of systems is still inclined to order more, refer more and spend more than physicians on the same cases. The next step is not simply better model size or more medical data; it is verification that the tool can support judgment without replacing restraint with automated overuse.