takeoff.watch

An AI is licensed to practise primary care unsupervised

Will any G20 country or US state license or authorise an AI system to diagnose and treat patients across general primary care without a clinician reviewing each case?

Chance by Q4 2031

22%

Cumulative probability

0%25%50%75%100%
  • Latest forecast

Forecast history

Forecast made on 15 Oct 2026

Latest of 5 forecasts · Methodology v2.1

Model reasoning

Forecast made 15 Oct 2026 · Methodology v2.1

No government has yet let an AI act as a general family doctor, diagnosing and prescribing with no clinician checking each case. Utah's program is the only plausible early route, but its approvals so far cover only refills, acne creams and physical therapy, and none has yet dropped case-by-case doctor review. I put the chance at about 2% by end-2027, rising to about 22% by end-2031. The biggest change would be Utah signing a broad AI primary-care pilot.

Where things stand (15 Oct 2026)

No G20 country or US state has authorised any AI system for general primary care without a clinician reviewing each case. Utah's Office of Artificial Intelligence Policy (OAIP) is the only regulator issuing anything close, and everything it has signed is narrow and still phased:

  • Doctronic: renewals of existing prescriptions only. Signed Oct 2025, live since Jan 2026, and still in Phase 1, where a physician authorises every renewal. The 2026 amendment tightened the bar for moving to Phase 2.
  • Legion Health: psychiatric renewals only. August AI: refills only.
  • Nolla Health: signed 22 Sep 2026. It can issue first prescriptions, but only for topical acne treatments. It starts with every prescription reviewed beforehand, then every one reviewed afterwards, then sampling. Each step needs OAIP's written approval.
  • Expect Fitness: pelvic-floor physical therapy plans.
  • University of Utah Health and Intermountain Health: master agreements that authorise nothing by themselves.

All of these are excluded as a named condition, a specialty, or refills. Under the question's phased-permission rule, even a broad pilot would count only when its sampling-only stage is approved. A stage where a clinician reviews every case afterwards does not count.

Signals pointing toward YES

  • OAIP is planning for broad AI. Its Aug 2026 "Evidentiary Expectations" document sets out a "clinical competency" route for broad-scope autonomous AI: open-ended patient intake and 80 or more clinical functions. Pilots move through three stages, ending with review of a sample of cases only.
  • Stage 3 can come quickly in a narrow pilot. In Nolla's plan it starts at 750 cumulative patients.
  • Companies want this. Doctronic (raised $40M) is in talks with Arizona, Texas and Wyoming. Curai has an application pending at the FDA for "autonomous AI primary care doctors" (NYT, 14 Sep 2026). Federal health officials are openly pushing AI that diagnoses and prescribes.
  • Other states may follow. Bills based on the Cicero Institute model (Idaho HB 945, Iowa HSB 766) would license autonomous AI providers. Neither passed in 2026. Texas created a regulatory sandbox in 2025 (part of its broader AI law), but nobody has used it yet.

Signals pointing toward NO

  • Strong professional opposition. Utah's medical licensing board asked for the Doctronic pilot to be suspended. The Federation of State Medical Boards says AI should not be licensed and has set up a working group on AI that acts with little supervision.
  • California went the other way. AB 1979, effective 1 Jan 2027, bars any clinical decision based solely on the output of a clinical decision-support system and requires licensed-professional judgment before approval.
  • Utah's legal basis expires unless renewed. The AI Policy Act (Utah Code Title 13 Ch. 72) is repealed on 1 July 2027 under Utah Code 63I-2-213. 2026's HB 320 did not change that date. Renewal looks likely but may come with limits, and voters in Utah have turned against AI data centres.
  • Federal device rules are unresolved. FDA has not settled how it treats autonomous prescribing software.
  • Even the narrowest pilot has not reached autonomy. Doctronic's refill pilot is still in Phase 1 after roughly nine months live.
  • Other G20 countries are far behind. China's health commission rules forbid AI-generated prescriptions. Brazil's medical council resolution (CFM 2454/2026) keeps decisions with physicians. Korea's new strategy only assists clinicians. Saudi Arabia's AI clinic ("Dr. Hua") still has a physician sign off on each case and covers a list of conditions. The UK sandbox bill (Regulating for Growth) is only at the proposal stage.

Reference class and model

There is no precedent for broad autonomous medical practice. As an analogy, California took about 4–5 years to go from narrow driverless-car test permits to broad commercial approval.

I broke the probability down by path, through end-2031:

  • Utah: program survives (~0.8) × a broad primary-care pilot is signed (~0.45) × its scope truly counts as general primary care (~0.6) × the sampling stage is approved in time (~0.5) ≈ 11%.
  • Other US states: ~6%.
  • US federal (FDA/CMS/VA): ~2%.
  • Other G20 countries: ~5%.
  • Combined: about 22%.

I modelled this as a quarterly chance of first approval that rises over time:

  • about 0.3% per quarter now, since nothing that could qualify is in the pipeline;
  • 0.8–1.2% per quarter in 2028, the earliest realistic point for a broad pilot to reach the sampling stage;
  • about 1.7% per quarter in 2030–31.

The cumulative figures were computed in Python from these quarterly rates.

Strongest case against my estimate (higher): AI may improve very fast, Utah moved from refills to new prescriptions in nine months, and there are many states that could act. That pulls me up from about 10% but not beyond the low 20s, given how strict the scope and review criteria are.

Sources

Resolution criteria

Opened 16 Sep 2026Not yet resolved

Question. By the last day of each quarter, has any G20 country or US state licensed or authorised an AI system to diagnose and treat patients across general primary care without a clinician reviewing each case?

Definitions.

  • G20 country: one of the 19 sovereign G20 members at opening: Argentina, Australia, Brazil, Canada, China, France, Germany, India, Indonesia, Italy, Japan, Mexico, Russia, Saudi Arabia, South Africa, South Korea, Türkiye, the United Kingdom, or the United States. Membership in the European Union or African Union alone does not qualify a country. US state means one of the 50 states, excluding DC and territories.
  • License or authorise: an enacted statute, final regulation, regulator’s decision, licensing-board decision, or executed government regulatory-relief agreement that expressly permits an identifiable AI system or AI-operated clinical service to provide the qualifying care. The permission may be granted to the system’s operator rather than to the software itself. A national authority must issue the permission for a G20 country; a state legislature, regulator, or licensing board may do so for a US state. National authorisation limited to a region or pilot site counts.
  • General primary care: first-contact outpatient medical care whose authorised scope encompasses assessment of new, undifferentiated symptoms, diagnosis and initial treatment of acute conditions, and diagnosis and continuing treatment of chronic conditions across multiple organ systems. A general-practice or primary-care scope with these powers qualifies; permission limited to a named disease, specialty, screening test, prescription renewal, or a list of isolated condition-specific services does not. Adult-only, telemedicine-only, or geographically limited primary care may qualify; authority to handle emergencies, perform procedures, or prescribe controlled substances is not required.
  • Diagnose and treat: the AI may make patient-specific diagnoses and select and initiate treatment, including new prescription medication where clinically indicated, rather than merely recommend decisions to a clinician. It need not physically dispense medicines or perform procedures.
  • Without a clinician reviewing each case: the authorised workflow permits completed diagnosis-and-treatment encounters without mandatory patient-specific review or sign-off by a human physician or other licensed clinical practitioner, either before treatment or routinely afterwards. General clinical supervision, clinician availability, sampled audits, and mandatory escalation of selected high-risk or uncertain cases are allowed. Both diagnosis and treatment must be permitted without universal case review across the qualifying primary-care scope, not just for one narrow task.

Resolves YES if a qualifying official publication establishes the permission and its scope and review requirements. A time-limited pilot or regulatory sandbox counts if it authorises actual patient care meeting all the definitions. Permission may be final with a specified future commencement date; actual deployment or patient uptake is not required. A phased permission counts only once the qualifying autonomous phase is approved without a further discretionary approval being necessary.

Exclusions (resolve NO).

  • Bills, draft rules, model legislation, funding awards, political announcements, applications, or general licensing frameworks with no identifiable system yet authorised.
  • Clinical decision support, triage, health information, diagnostic-only systems, specialty-only systems, or refills of previously prescribed medications without authority to diagnose and initiate treatment across general primary care.
  • Research ethics approval alone, simulated patients, and testing in which clinicians must approve every case.
  • Medical-device marketing clearance alone unless the official record also establishes permission to deliver the qualifying care without universal clinician review; reimbursement eligibility or regulatory silence alone is insufficient.
  • For each horizon, resolve NO if no qualifying publication has appeared by quarter-end.

Resolving source. Official legislative or gazette records, published decisions and registers of the competent national health ministry or medical-products regulator, and official medical-licensing-board records; for US states, official enacted laws and records of state health, commerce, or professional-licensing authorities. For example, Utah Department of Commerce Office of Artificial Intelligence Policy’s Doctronic regulatory mitigation agreement and pilot record document a refill-only permission, not qualifying general primary care. The legal instrument and its incorporated scope and supervision conditions control over summaries. These official records are the narrower resolving sources; developer claims under Public confirmation (A2) alone do not suffice. Apply Resolution date (A3) and the Resolution window: the qualifying official publication must be dated on or after 2026-09-16. Later revocation does not undo an established YES.

Download all forecasts (JSON)