Is AI Taking Over Medicine? Utah Just Let One Write Prescriptions

What Utah actually approved

Nolla Health's pilot launched in Utah on 5 October 2026. A resident over 18 completes an intake questionnaire, verifies identity, consents, and takes a five-angle face scan. The AI proposes a treatment.

That sentence alone is worth sitting with. A questionnaire and a photograph, and a prescription comes out the other end — no appointment, no waiting room, $4.99 a month for the subscription with medication billed separately.

The scope is deliberately narrow: topical creams and gels for mild-to-moderate acne only — no severe acne, no oral medication, no isotretinoin, no hormonal therapy. If the model is not confident, the patient is routed to a physician rather than given an uncertain answer.

The part the headline skips

Oversight is staged, not absent. In stage one, two Utah-licensed physicians review every prescription before it reaches a pharmacy. Only later, and only with written state approval, does the AI submit directly — and even then a physician reviews cases retrospectively each week, capped at a few hundred patients.

The authorisation runs one year, renewable twice at the discretion of Utah's Office of Artificial Intelligence Policy, and the agreement states plainly that it is not a state endorsement of the product. Nolla's claim that clinicians agreed with its recommendations in over 96% of real-world cases is the company's own, not an independent finding.

And there is precedent worth knowing: Utah ran an earlier AI pilot with Doctronic, and in April the state's Medical Licensing Board called for it to be suspended over patient-safety and oversight concerns. Regulators are experimenting, and they are also willing to stop.

Everywhere else, AI is doing the filing

The FDA's List of AI-Enabled Medical Devices passed 1,600 authorisations as of September 2026 — and the agency notes the list is not comprehensive, because it finds devices by keyword. Roughly three quarters are radiology. Clinician use climbed from 38% of physicians in 2023 to 66% in 2024, and now past 80%.

The pattern is consistent: regulators clear AI fastest where a human still signs the result, and clinicians adopt it fastest where it removes typing rather than judgement.

But look at what they use it for: summarising research (39%), discharge instructions (30%), visit notes (28%), chart summaries (28%), patient messages (19%) — and assistive diagnosis dead last at 17%. Utah is the exception that proves the rule.

The gains are real, and smaller than the pitch

The largest study of ambient AI scribes compared 1,800 clinicians against 6,770 controls across five health systems and found 13 fewer minutes a day in the record and 16 fewer documenting. Clinicians using it for more than half their visits saw two to three times that — but only 32% did. The benefit tracks adoption, which makes this a change-management problem in software costume.

The cost nobody budgeted for

At four Polish centres, researchers measured what happened to endoscopists' unassisted colonoscopies after AI polyp detection arrived. Detection fell from 28% to 22% in three months, among 19 clinicians with over 2,000 procedures each. Observational, single-system, authors calling for randomised work — and still the first real-world evidence that leaning on the tool can cost you the skill beneath it.

What follows from this

Buy for documentation, not diagnosis: that is where the clearances, the adoption and the measured savings actually are. Budget for adoption rather than licences, because the benefit concentrated among heavy users. Keep measuring unassisted performance, because nobody finds deskilling who does not look for it. And read the pilot agreement, not the headline — Utah's says two doctors sign off on every prescription.

Medicine is not being taken over. The filing cabinet is — and the boundary around judgement is being tested one narrow, revocable pilot at a time. That is a slower story than the one circulating on social media, and a far more useful one if you actually have to make a decision about this technology.

Frequently asked questions

Can AI legally prescribe medication in the United States?

Only in a narrow, state-authorised pilot. In October 2026 Utah authorised Nolla Health's AI to issue initial prescriptions for topical acne treatments to adults with mild-to-moderate acne. Two licensed physicians review every prescription in the first stage, and the one-year authorisation explicitly is not a state endorsement.

Is AI replacing doctors?

No. Despite more than 1,600 FDA-authorised AI-enabled devices, only about 17% of physicians report using AI for assistive diagnosis. The overwhelming majority of medical AI handles documentation, summarisation and image detection rather than clinical decisions.

How many AI medical devices has the FDA authorised?

Over 1,600 as of September 2026, roughly three quarters of them in radiology. The FDA notes its published list is not comprehensive, since devices are identified by searching for AI keywords in authorisation summaries.

How much time do AI scribes actually save?

About 13 minutes per day in the electronic record and 16 minutes in documentation, in the largest multisite study. Clinicians who used it for more than half their visits saw two to three times the benefit, but only 32% used it that heavily.

Can relying on AI make clinicians worse?

Early evidence says it can. After AI polyp detection was introduced at four centres, experienced endoscopists' detection rate during colonoscopies performed without AI fell from 28% to 22%. The study is observational, but it is the first real-world clinical evidence of deskilling.