The Expensive Second Opinion

What they did

Twenty general purpose models and four medical ones (MedGemma, Meditron, OpenEvidence) each received the presentation half of 200 deidentified outpatient and emergency notes and wrote an assessment and plan. The treating physician's documented plan was the comparator. Three separate LLMs extracted the orders from every plan, and everything was priced at 2026 Medicare rates.

What they found

The models recommended $314 of tests and referrals per visit against the physician's $97, a 3.2 fold difference that held across every system (range 2.1 to 6.3). Referrals ran 0.87 per visit versus 0.15. New medications ran 1.14 versus 0.59. The physician ordered no tests at all in 133 of 200 visits, and on those visits every system added a workup, most often the same reflexive panel (CBC, metabolic panel, A1c, TSH, lipids). Newer flagships suggested more than their predecessors. A parsimonious prompt cut the excess to 1.4 fold but pushed 17 of 19 systems off the physician's diagnosis. A cost aware prompt reached 2.5 fold without that loss.

A physician review of 1,000 plans rated 5 percent clearly unnecessary, 17 percent debatable, 59 percent reasonable given the model's limited context, and 18 percent guideline concordant workup the physician had skipped.

Quality and bias

The notes come from MTSamples, a transcription training corpus from the mid 2000s, not a representative sample of practice. The comparator is one physician's plan, not a consensus standard, and a second reviewer judged the models to under order in about a quarter of cases. Every suggestion was priced as if acted on, so the figures are a ceiling. No outcomes were measured. One author is at MedBAR.ai, a company; data and code are public. Excess is not the same as waste, and the authors say so.

Solid

What they did

The authors built 177,120 synthetic vignettes from 123 disease and active ingredient pairs crossed with 1,440 patient profiles (age, sex, race, LGBTQ+ identity, income). Each vignette named the disease and the ingredient and asked for the top three FDA approved product names. Five models answered: GPT-5.4, Claude Sonnet 4.6, Gemini 3.1 Flash, Llama 4, and MedGemma. Four physicians curated the outputs against the FDA Orange and Purple Books.

What they found

For a given pair, one product took 97.5 percent of first choice recommendations on average, with 5.5 approved products available. The five models agreed almost perfectly (Fleiss kappa 0.895). Real world prescribing from MEPS matched the models' first choice 29.7 percent of the time. An originator label raised the odds of being recommended 20.5 fold, a dedicated Wikipedia page 3.0 fold, and Google search volume 1.7 fold; later FDA approval and higher real world volume lowered them. Across the 51 ingredients that could be priced, 38 cost more under the models' choices, a projected $55.6 million a year. Prompts asking for generics shifted the pick but not the concentration. Vignettes describing Black patients drew more recommendations that failed regulatory checks (odds ratio 1.64), mostly discontinued products.

Quality and bias

Synthetic vignettes and a narrow task (pick a product, the ingredient is given). Temperature was set to zero, which produces some of the agreement by design. MEPS covers 2019 to 2023 and the models were queried in 2026. The cost projection assumes dispensing follows the recommendation, and pharmacies, formularies, and substitution laws make sure it does not. Funded by a Japanese government research program; no competing interests; vignettes and code are public.

Solid on convergence Cost figure is a simulation

What they did

The group took 1,954 clinical cases from four sources (Reddit AskDocs, oncology patient messages, script concordance tests, and USMLE style dermatology vignettes) and rewrote each with GPT-4o to swap or remove gender, or to add a colorful or anxious tone, for 6,232 scenarios. Ten physicians supplied 7,356 annotations on three binary decisions: self manage at home, see a clinician, order a test or referral. Five models (GPT-4o, DeepSeek-R1-32B, MedGemma-27B, Llama-3.3-70B, Qwen2.5-32B) answered the same questions, 226,191 times.

What they found

On unperturbed cases GPT-4o scored 91.7 percent against the clinicians' 90.5. Across tasks the models recommended care more often than physicians did when physician consensus said it was not needed. Under an anxious or colorful rewrite, GPT-4o lost 8.5 to 10.1 points, Qwen lost 14.6 to 15.9, and DeepSeek lost more than 12 under every perturbation. Restricted to cases where the physician majority recommendation did not change, physicians moved 0.8 points on average while DeepSeek lost 12.5 points on the visit decision and 19.5 on the resource decision.

Quality and bias

The rewrites were generated by a model, not collected from patients, and an unchanged physician label does not prove a rewrite was clinically neutral. The model set is mid 2025 vintage and mostly small open weights; GPT-4o is the only frontier system. Decisions were reduced to yes or no. Retrospective, no clinician and model interaction. Published in the Findings track, which is peer reviewed but a tier below the main conference. Code is public.

Mostly Solid

What they did

Singapore General Hospital runs a radiographer led screening program in its emergency department. For three months in early 2025, thirteen radiographers (five senior, eight junior) read 2,496 consecutive musculoskeletal examinations three times: alone, then after seeing a commercial AI's output (Rayvolve 4.1), then a final call. A radiologist panel set the reference standard. The study was prospective and reported to the DECIDE-AI guideline.

What they found

Standalone AI sensitivity was 75.6 percent overall and 46.2 percent on subtle fractures, well below the radiographers' 91.6. With the AI as a second reader, radiographer accuracy rose from 97.6 to 98.0 percent, sensitivity from 91.6 to 92.9, specificity from 99.0 to 99.2. Eleven errors were corrected and one was introduced. Of 729 cases where the AI disagreed with the initial read, the radiographers changed their answer in 11. The AI returned no output or incomplete output on 8.8 percent of examinations, clustered on six days, with no visual signal on the workstation to say so.

Quality and bias

Single center, single radiography unit, and a staffing model (radiographer first read) that does not exist in most US departments. The reference standard was one radiologist's read with consensus only for ambiguous cases. Only 104 subtle fractures, so no subgroup finding survived correction. Readers knew they were being studied. Interpretation time, workload, and downstream outcomes were not measured. The vendor is thanked for technical support; authors declare no competing interests, and vendor uptime logs were not available to them.

Solid, small effect reported as small

What he argues

Emanuel names five cognitive tasks where he says specialized systems already match or beat physicians: taking a history, building a differential, ordering confirmatory tests, writing a guideline concordant plan, and managing chronic disease. He expects autonomous AI to arrive first in protocol driven care (hypertension, type 2 diabetes, lipids, gestational diabetes, preeclampsia, chemotherapy side effects), and offers blood pressure control as the example: about 25 percent of patients controlled today, perhaps 80 with daily monitoring and algorithmic adjustment.

From there the argument turns to the profession. Medical school admissions select for organic chemistry and physics; he wants them to select for emotional intelligence and communication without dropping baseline competence. He worries about deskilling and floats no AI days in residency. He wants the treating entity, human or machine, devoted only to the patient in front of it, with rationing handled by institutional frameworks outside the exam room. He proposes certifying autonomous systems by having them manage 200 to 300 real patients under physician oversight with expert panel review, domain by domain. And he sets out a liability rule worked out with tort attorneys: if a physician keeps final approval, the physician carries malpractice liability; if the AI acts alone, the developer does. Consumer chatbots, in his view, should carry a warning label.

Quality and bias

This is an argument, not a study. The five task claim rests on benchmark and simulation results, and the four papers above are the kind of evidence that complicates it (over ordering, brand preference, sensitivity to tone, silent failures). The blood pressure figures are a projection. The written versions are Emanuel's JAMA perspective of August 17, 2026 and an Atlantic essay with Vinod Khosla of September 8, 2026; Khosla disclosed an investment in Curai Health, which sells physician supervised AI rather than the autonomous kind the essay argues for. The AMA and the American College of Physicians have both published frameworks that keep decision authority with the physician.

Position, not evidence