Will AI replace doctors? A map, September 2026

The models already beat physicians on written cases, but payment rules and liability law will settle where doctors are replaced. The likeliest substitute is a cheaper licensed person with a model and a physician's signature.

• Better on paper
• What the models can do, and what nobody has shown
• What runs without a doctor today
• Three ways physician work gets replaced
• Exposure by specialty
• The rules, and who gets sued
• Where the money is going
• What should be done

Better on paper

The models are now better than physicians at the parts of medicine that fit in a text box. The best studies of 2026 say so. Nobody has shown that a model running care on its own leaves patients better off, mostly because almost nobody is allowed to try. Payment rules and liability law, not another benchmark, will decide where physicians are replaced.

Three points matter most. The models' lead over physicians in reasoning is real and growing. The only care in the United States that is allowed to run without a physician's signature is a short list: a retinal screening algorithm, one state's refill sandbox and a cleared insulin-titration app that works inside a doctor's plan. What actually reduces demand for physicians is something humbler than an autonomous AI: a medical assistant with a model and a remote doctor's signature, seeing four or five times the patients. That arrangement is legal in every state today and already running at scale, although California's AB 1979, signed in September, will bar the use of AI to direct unlicensed staff in licensed clinical work from 2027.

Meanwhile the labor market says the opposite of the headlines. Physician pay rose again this year. Eighty-five percent of physicians say shortages affect their practice. Radiology, the specialty everyone said would be automated first, pays about $610,000. No health system in the country has cut physician positions and named AI as the reason; the cuts that have been reported are in IT and coding. What has happened instead is less dramatic and matters more: the time that ambient scribes save is being converted into higher productivity targets, and there is now decent evidence that leaning on the tools erodes skills.

What the models can do, and what nobody has shown

The most striking result came in June, when Nature published Google's AMIE disease-management trial: a randomized, blinded, virtual OSCE with 100 three-visit cases across five specialties and 21 primary care physicians. AMIE's management plans were rated appropriate 95 to 98 percent of the time; the physicians' plans came in at 72 to 81 percent. The model won every column, from guideline alignment and treatment precision to drug reasoning. Yet the authors said, in their own paper, that the system is "not ready for real-world translation." The patients were actors, the cases were text only, physicians from outside the UK were judged against UK guidelines, and there was no pharmacist and no order entry. Both statements are accurate, and the second deserves as much weight as the first.

In April Science published a Beth Israel and Harvard study in which an OpenAI model matched or beat attending physicians at triage, first-contact and admission decisions on real emergency department records from Boston. Adam Rodman, who ran it, was careful to say the results support teaming, not "AI doctor companies." OpenAI's own benchmark, HealthBench Professional, puts GPT-5.4 at 59 against 43.7 for specialty-matched physicians, who were allowed the internet and unlimited time. But OpenAI wrote the rubric and graded its own model, and its paper says the scores are "not directly comparable to real-world performance." Microsoft's MAI-DxO, the "four times more accurate than doctors" result from 2025, is still an unrevised preprint more than a year later, and the physicians in its comparison were denied both search tools and colleagues.

The most useful evidence is smaller and duller. The one large real-world deployment, OpenAI's work with Penda Health in Kenya across about 40,000 visits, cut diagnostic errors 16 percent and treatment errors 13 percent when the model acted as a safety net for clinicians who kept full authority. Patient-reported recovery did not change. Stanford's series of trials on physicians using GPT-4 found something that should trouble every doctor: a physician with the model did no better than a physician without it, while the model alone scored 92 percent. Only when the workflow was redesigned, so that the physician commits to an assessment first and then compares, did the combination beat physicians alone and roughly match the model. That is a finding about workflow more than intelligence: left to themselves, physicians use these tools badly.

On rubric-scored text cases, then, the model alone is at least as good as a physician alone, and a physician with the model is often no better than the model unless someone has thought about the workflow. Nobody has shown better outcomes for patients from AI-led care, because nobody has been allowed to test it. On the physical, long-term and accountable parts of medicine, the question is untested rather than settled.

What runs without a doctor today

The list is shorter than the press releases suggest. LumineticsCore diagnoses diabetic retinopathy with no clinician read and bills its own CPT code. Doctronic operates in a Utah regulatory sandbox that lets an AI renew about 190 non-controlled chronic medications, yet after seven months it was still in Phase 1, which meant a licensed practitioner signed every renewal. The AI recommended renewal 72 percent of the time and physicians agreed 91 percent of the time. Utah's own medical licensing board asked for the pilot to be suspended and was turned down. Legion Health received a parallel approval for fifteen psychiatric medications. UpDoc, which titrates diabetes medications by voice or chat inside a plan a physician wrote, holds the first FDA clearance for a patient-facing large language model. In Europe Oxipit's ChestLink clears normal chest films on its own under a CE mark; nothing like it is cleared in the United States.

Everything else that calls itself an AI doctor is a person plus a model plus a doctor's signature. Akido Labs says it is the first AI-native health system: one million patients, 100 clinics, medical assistants running AI-guided visits with a physician approving remotely, and a claimed four to five times the patients per physician. No outcomes trial has been published. Counsel Health offers a free AI front door and a $29 physician follow-up. K Health sits inside Mass General Brigham, Mayo, Penn and Novant. Amazon gives Prime members up to five free messaging visits a year for more than 30 conditions. Hippocratic AI's voice agents have logged 250 million patient interactions, handling intake, outreach and post-discharge calls. Hims & Hers says AI now answers 80 percent of the questions from weight-loss patients on its new care platform. This is the part changing fastest, and it has more to do with labor costs than with robots.

Three ways physician work gets replaced

The three pathways move at different speeds, and the loudest is the slowest.

Full autonomy, meaning no human signature, is narrow and heavily regulated: retinal screening, refills in one state and titration by protocol. It grows one FDA clearance and one state sandbox at a time.

Supervised autonomy means an agent acting on its own while a physician reviews exceptions or samples. It is what ARPA-H is now funding: $62.7 million for partially autonomous heart-failure agents that assess symptoms, adjust guideline-directed therapy and order labs under a plan a physician prescribes, with FDA packages due within two years and a 2,500-patient trial at Kaiser. Baptist Health's chief medical information officer described it as a shift from human-in-the-loop to exception-based oversight, a phrase doctors will hear often.

Labor arbitrage is AI plus a medical assistant, nurse or pharmacist plus a remote physician's signature. It requires no new law and no new clearance, and it is the pathway that lowers the number of physicians needed per thousand patients. Akido is the proof of concept. For any physician, the useful question is not whether a model can do the job but whether a cheaper licensed person with a model and the physician's signature could.

Exposure by specialty

Hospital medicine shows what moderate exposure looks like. Within eighteen months, hospitalists at any hospital that runs Epic can expect drafted notes and drafted orders waiting for their signature, discharge summaries written by agents, automated first-pass utilization review with the physician advisor kept for the hard cases, and pressure toward virtual nights. None of that reduces the number of hospitalists a hospital needs on days, but it will raise the census each is expected to carry and move liability onto whoever signs. Hospitalists should read what they sign.

The rules, and who gets sued

FDA

In January the agency loosened the two guidances that matter most to a practicing physician. Under the revised clinical decision support guidance, software that hands a clinician a single clinically appropriate recommendation, including risk scores and tools for differential diagnosis, is not a regulated device; time-critical uses such as stroke and sepsis, image analysis and anything that acts without a clinician still are. The revised wellness guidance lets wearables report blood pressure, glucose and lactate as wellness data so long as they make no claims about disease. The agency's AI-enabled device list has passed 1,600 authorizations. On August 18 the FDA published a discussion paper on generative-AI devices proposing "competency-based" evaluation: testing a model the way a clinician is tested, then monitoring it after it reaches the market. Comments close on October 19, and the agency's digital health center says formal guidance is coming. In July ten companies, among them Anthropic, Microsoft, Amazon, Doctronic and Hippocratic, demonstrated "AI doctor" technology to FDA, CMS and HHS officials in a closed-door session at White Oak.

Medicare

Two programs matter. ACCESS, which launched in July, pays outcome-linked amounts for technology-enabled care of hypertension, diabetes, musculoskeletal pain, depression and anxiety, with heart failure, COPD, substance use and tobacco to be added next spring. About 160 organizations are in, including Headspace, Noom and WeightWatchers. ACCESS is paired with the FDA's TEMPO pilot, so AI tools the agency has not authorized are already treating Medicare beneficiaries inside it under enforcement discretion. The other program, WISeR, runs AI-assisted prior authorization in six states, and it is the cautionary tale. Records obtained by the Electronic Frontier Foundation show a vendor that told CMS it would auto-affirm requests because it was not ready, two vendors that denied more than 20,000 requests in three months, vendors paid per denial, and one request that sat for 83 days. A Senate attempt to kill the program failed 46 to 50 in July. Physicians fighting a prior authorization anywhere can cite WISeR as evidence that AI gatekeeping fails in practice.

The proposed 2027 fee schedule cuts the conversion factor by 1.19 percent, turns G2211 into a percentage modifier, restricts remote monitoring to staff employed by the practice (which would end vendor-run remote patient monitoring in Medicare on January 1), and asks for input on ambient documentation and AI-enabled wellness visits. The final rule is expected around November 1. Telehealth flexibilities run through December 31, 2027, and the DEA's telemedicine prescribing rule runs through the end of this year. On September 14 the New York Times reported that officials have discussed a Medicare payment category for company-run "AI physicians" at 60 to 80 percent of the human rate. Nothing has been proposed yet, but it may be the payment question that matters most to physicians.

The states

Twenty-two states enacted 33 health-AI laws in the first seven months of 2026, on top of last year's wave, by the count of Manatt, which keeps the best tracker. The pattern is remarkably consistent: patients must be told when AI is used, a licensed human stays accountable for clinical use, payers' denials get human review, and chatbots may not pose as mental-health providers. Colorado, Texas, Maine, Vermont and Louisiana now put the duty to review AI output on the practitioner in statute. Alabama, Georgia and Minnesota require human review of payers' utilization decisions. Eleven states passed chatbot laws. California's governor signed AB 1979 on September 30; from January 2027 it bars the use of AI to perform licensed clinical work on its own or to direct unlicensed staff in it, a rule aimed squarely at the labor-arbitrage model. No state has authorized an AI to practice medicine, and the Federation of State Medical Boards said in August that AI "is not ready to be independently licensed like a physician." About twenty liability bills were introduced, and zero passed.

Who gets sued

Liability still lands on the licensed human, and in at least five states that is now written down. No public malpractice case has been tied to an AI scribe yet, but malpractice carriers are warning about the failure modes: hallucinated template text, omitted non-verbal findings and the slow decay in how carefully doctors read a draft once they trust it. The plaintiffs' bar is testing a different theory against the model companies: unlicensed practice of medicine. Winters v. OpenAI, filed in July, alleges ChatGPT told a man with unstable blood pressure to stay in his recliner and he developed a pulmonary embolism. Pennsylvania sued Character.AI under its Medical Practice Act after a chatbot posed as a licensed psychiatrist. The AMA's June policy says AI "cannot replace physician judgment," and its chief executive wants the companies to "take on higher levels of liability." A practice that documents physician review of every AI output and tells patients when AI is used should be compliant everywhere today. One that lets AI output reach a patient unreviewed is exposed in at least five states and probably in front of any jury.

Where the money is going

AI took 46 percent of healthcare venture dollars in 2025. Rock Health counted $7.4 billion in US digital health funding in the first half of 2026, with twenty megadeals taking 45 percent of it and zero IPOs. The highest valuations go to two kinds of company. The first strips administrative labor from existing physicians: OpenEvidence at $12 billion, Commure at $7 billion and Abridge at $5.3 billion. OpenEvidence is free to clinicians, supported by advertising and, by its own count, used daily by more than 40 percent of US physicians; in July it considered, then shelved, a round at $20 billion. The second kind sells consumers a longevity subscription that needs a physician to interpret it. Function Health, valued at $2.5 billion, took another $450 million in July, and Oura filed for a Nasdaq IPO in September with $1.4 billion in trailing revenue, then postponed it a day before trading was to begin. The annual laboratory bundle is converging on $300 to $600: Function's starts at $365 and Fountain Life's new tier is $595. As tests and scans become commodities, the margin lies in interpreting them and answering for the result.

In public markets the winners collect tolls on physicians' workflow. Doximity's shares jumped by roughly two-thirds after a quarter in which its AI scribe users grew tenfold; HeartFlow keeps raising guidance on AI plaque analysis; Tempus won ARPA-H money for autonomous cardiology. Undifferentiated telehealth is losing: Teladoc's revenue fell and BetterHelp shrank. Revenue-cycle software is now priced as a casualty of AI even when it uses AI itself; Waystar hired bankers to explore a sale after its stock fell 24 percent this year. The one ETF that calls itself a pure play on healthcare AI holds genomics and surgical robots rather than care delivery. None of this is investment advice. The prices show where investors think physician time is worth paying for and where they think it can be replaced.

The bear case for physicians is worth reading in full. Ezekiel Emanuel and his co-authors, Vinod Khosla among them, argued in JAMA in August and again in STAT on September 9 that by 2030 autonomous AI will beat AI-assisted physicians at history-taking, differential diagnosis, test selection, guideline-concordant prescribing and chronic disease management, and that value will move to procedures, presence, accountability and relationships. The AMA's chief executive answered in STAT the same day. Both sides land in the same place: the physical, relational and accountable parts of the job are the parts that stay.

What should be done

Physicians are always being told to "learn AI," advice that means nothing. What follows is specific to where they work.

Hospitalists should sign nothing they have not read. In at least five states that duty is now law, and everywhere it is their license at stake. They should ask their groups, in writing, how productivity targets will change as ambient and agent tools roll out, because Doximity's data say the targets rise and the pay does not, and the group that has that conversation early sets the terms. They should also get onto the AI governance committee or the team rolling out Epic's agents. Two-thirds of hospitalists who use these tools say they were not adequately trained on them, and the tools end up serving whoever is in the room when they are configured.

Primary care doctors should let the refill, inbox and triage work go to a model and a nurse. It is going there regardless, and it was never the part patients valued. What patients are paying for, and what the surveys say they want, is a human who knows them, verifies what the machine says and is accountable. Nobody has validated a bigger panel size with AI, and doctors should not accept one without evidence. Doctors who own their practices keep the AI dividend, while employed doctors are negotiating for it.

Emergency physicians and proceduralists are among the least exposed. Every exposure index in existence bottoms out on hands-on, time-critical, physically present work, and the virtual hospitalist companies stop at procedures. One in thirteen American emergency departments has no attending physician on site around the clock, and nearly nine in ten of those are critical access hospitals. The rural hospitals in line for $50 billion of new federal money are being told that AI will save them, and their leaders do not believe it. Physicians weighing airway work, vascular access or rural coverage should go ahead; the market is telling them to.

Physicians drawn to the business side will find that every company that wants to run AI-assisted care legally needs a physician in the loop, and pays for one. Part-time medical director roles run $3,000 to $10,000 a month, model-evaluation work pays $130 to $180 an hour for internists and emergency physicians, and advisory boards pay cash and equity. The first AI-related malpractice cases are being filed now, which means expert witnesses with real experience of AI governance will be in demand by 2028. Chief medical AI officer roles pay $420,000 to $720,000 base and require an MD in nine out of ten searches. Recruiters want stories of deployment rather than certificates. The clinical informatics board's practice pathway closed last year. For those who want a credential, Stanford's Coursera specialization is cheap and Harvard Chan runs a $2,600 online course in December, but building and governing one real tool in their own practice will teach them more and prove more.

All physicians who use the tools should guard against losing their edge, and the evidence on this is uncomfortable. Experienced endoscopists' adenoma detection fell from 28 to 22 percent after three months of AI assistance. Pathologists reversed correct diagnoses more than 30 percent of the time when a wrong AI prompt was on the screen. Doctors should commit to their own assessment before looking at the model's, then compare; the Stanford data say that is the workflow that closes the gap with the model. They should ask what would harm the patient if it were left out, because these systems fail by omission far more than by commission. They should keep deliberate intervals without AI for reads and procedures, and track their unaided numbers. They should read every draft as if they wrote it, because legally they did. And no irreversible, high-stakes decision should go to anything that cannot be deposed.

Back to the top of this page