The 'Selfie Prescription': How Utah's AI Experiment Could Get You Treated in Minutes Before a Doctor Ever Sees You

Picture a teenager in rural Utah with a face full of painful acne. The nearest dermatologist is hours away, and appointments can be about two months out. Now picture that same teenager opening an app, answering a structured set of medical questions, snapping a few photos of their skin, and walking away with a prescription, written not by a doctor but by an algorithm.

That is no longer a thought experiment. On October 5, 2026, Utah's Office of Artificial Intelligence Policy announced that it is expanding its health-care "AI sandbox" with three new clinical pilots. One of them lets an AI system assess mild-to-moderate acne and issue first-time prescriptions from a short list of low-risk treatments, with physician review that starts on every case and is designed to taper off as the pilot proves itself. Reports describe it as the first time a U.S. regulator has cleared AI to start a new prescription, rather than simply renewing one a doctor already wrote.

It sounds small. Acne cream is not open-heart surgery. But the precedent is enormous, because it answers a question health systems around the world have been circling for years: when, if ever, should a machine be allowed to make a clinical decision on its own?

What Utah Actually Approved

Utah created its Office of Artificial Intelligence Policy under a law championed by Senate Majority Leader Kirk Cullimore. The office runs a "regulatory sandbox": companies can apply for limited, temporary relief from specific state rules, including some rules on the practice of medicine, in exchange for close supervision and data reporting. Before any pilot starts, a company must submit a plan covering patient safety, consumer benefit and how it will report results. Pilots begin with 12-month agreements, and the state keeps the final say on who gets in and who stays.

The October announcement added three health pilots, each run by a young digital-health company:

• AI-led acne care. An app-based service evaluates patients through structured intake questions and analysis of skin photos, then chooses among pre-approved, low-risk treatment plans. Utah dermatologists oversee the program, and more serious cases, including severe acne and pregnancy, are routed to a licensed clinician or in-person care.
• AI-assisted prescription renewals for chronic conditions. A second company will renew existing, non-controlled prescriptions from an initial formulary of roughly 148 medications, starting with lower-risk drugs. The AI cannot write new prescriptions, change doses or substitute medications, and a Utah-licensed physician reviews each renewal.
• AI-supported pelvic-health physical therapy. A third pilot uses AI to help with intake, treatment planning and follow-up so that a small number of remote physical therapists can serve more women during pregnancy and after childbirth.

Alongside the pilots, the state signed "master agreements" with two of its largest health systems, Intermountain Health and University of Utah Health. These cover data privacy and AI governance and let individual clinical departments propose future pilots, which still require state review.

The safety ladder

The acne pilot does not hand an algorithm the keys on day one. Published accounts agree on the basic shape, though they differ on the exact numbers. In the first stage, a licensed physician reviews and signs off on every AI-generated prescription for the first 100 patients before anything reaches a pharmacy. In later stages, review shifts to after-the-fact checks, and eventually to a physician auditing a sample of prescriptions on a regular schedule, plus any case involving an escalation or a side effect. The company says clinicians have agreed with its AI's treatment recommendations in more than 96% of real-world cases. That figure comes from the company, not from an independent study.

That gradual approach mirrors how we train new doctors: supervised at first, then trusted with more independence as they show they can handle it.

The New Twist: Outside Referees

The most important part of the announcement may not be the pilots themselves but who will judge them. Utah named a third-party evaluation network to independently test whether products meet the state's standards of evidence: the expected benefit, the ways a tool can fail, and what evidence would change an approval decision. The first evaluators include the Coalition for Health AI, a non-profit that has published frameworks for responsible health AI, and Stanford's Clinical Excellence Research Center, along with several independent assurance firms.

This responds directly to the loudest criticism of Utah's first experiment. In January 2026, the state let a different company's AI renew certain routine prescriptions. The Utah Medical Licensing Board objected that it had not been consulted, and in April it urged the state to shut the pilot down, citing risks around side effects, drug interactions, contraindications and whether a treatment was still appropriate for a patient whose condition had changed. Physician and pharmacist groups nationally echoed those worries.

The state answered with data. Five-month results released in late May showed the AI recommended renewal in 72% of cases, and reviewing physicians agreed that renewal was appropriate 91% of the time. In the other 28%, the AI escalated to a human. Physicians agreed with 69% of those escalations and judged 31% overly cautious. In other words, the system tended to err on the side of caution rather than recklessness. No specific patient harm was reported, but critics note that five months of data from one state cannot settle the question.

The licensing board has since formed a working group to support the AI office's reviews. Its members have stressed that this is not a blanket endorsement, and that their focus remains physician-led oversight, phased safety review, clear patient disclosures and data collection.

Why This Is Happening Now

The driving force is simple arithmetic. Demand for care is rising faster than the supply of clinicians. At the Utah announcement, a senior Intermountain Health executive warned that up to 40% of today's health-care providers may retire before 2035. Specialties like dermatology already have long waits: the acne company cites an average Utah wait of 61 days and says eleven Utah counties have no dermatologist at all.

Officials framed the pilots as a way to direct scarce specialist time where it matters most. If an AI can safely handle a routine, well-defined condition, the dermatologist is free for the patient with a suspicious mole. Utah Commerce Commissioner Margaret Woolley Busse argued that, done right, the approach could improve access and lower costs while keeping human responsibility for key medical decisions. University of Utah Health's chief innovation officer, Dr. Jim Hotaling, put it plainly: "The goal is not to take the human out of the loop."

The Collision Course: States vs. the FDA

Here is where things get legally interesting. In the United States, the practice of medicine, including who may prescribe, is regulated by states. Medical devices, including software that diagnoses or recommends treatment, are regulated by the federal Food and Drug Administration. Utah's position is that prescribing authority is a state matter. Legal scholars and commentators have asked whether software that effectively makes a diagnosis and picks a drug is itself a medical device that should go through FDA review first.

The FDA is moving on its own track. On August 18, 2026, it released a discussion paper on regulating generative AI-enabled medical devices, floating a "competency assessment" approach loosely modeled on how physicians are trained and tested, plus risk-proportionate monitoring after a product reaches the market. The paper explicitly addresses foundation models and agentic AI, and public comments are due October 19, 2026. Meanwhile, a federal bill, the Healthy Technology Act (H.R. 238), would let AI count as a prescriber only if it is both authorized by a state and cleared by the FDA. It was referred to the House Energy and Commerce Committee in 2025 and has not advanced.

Utah's leaders have said openly that they hope the model will be adopted elsewhere, including at the federal level. Whether Washington embraces it, preempts it or ignores it will shape how quickly AI prescribing spreads.

How the Rest of the World Is Approaching It

Canada

Canada splits authority in a similar way. Health Canada regulates software that qualifies as a medical device, while prescribing rights are set by the provinces and territories and enforced by their regulatory colleges. No Canadian jurisdiction currently lets AI prescribe on its own, but the same pressures exist: long specialist waits, many people without a family doctor, and enormous distances in rural and northern communities. Any Canadian equivalent would also have to satisfy privacy law, including PIPEDA federally and provincial health-information statutes, which govern how patient photos and health data are collected and used.

United States

Beyond Utah, U.S. states are moving in different directions, some writing rules to restrict AI in clinical care, others experimenting. Privacy is a live question too. HIPAA governs covered health-care entities in the U.S., much as PIPEDA and provincial laws do in Canada, but consumer health apps do not always fall neatly under either, which is why the sandbox's data-reporting and privacy terms matter.

Europe

Europe has chosen a more rules-first route. Under the EU AI Act, AI used in medical devices is classed as high-risk. After the Digital Omnibus amendments took effect in July 2026, the high-risk obligations for AI embedded in regulated products such as medical devices now apply from August 2, 2028, and those for standalone high-risk systems from December 2, 2027. The United Kingdom, outside the EU, runs its own sandbox. The MHRA's "AI Airlock" has tested AI medical devices in two phases, and the government has committed £1.2 million a year from 2026 to 2029 to continue it. Unlike Utah's approach, the Airlock focuses on how to regulate devices rather than waiving rules on who may practise medicine.

Asia and the Middle East

Several governments in Asia and the Gulf have used sandbox-style programs to trial digital health under supervision, often paired with national strategies to build AI capacity. Where clinician shortages are acute and digital adoption is high, a Utah-style model for routine, low-risk conditions could be attractive. But these regulators also tend to keep licensed professionals legally accountable for clinical decisions, which makes fully autonomous prescribing a harder sell.

The Risks Nobody Should Ignore

• A photo is not an exam. Lighting, camera quality and angle can all distort an image. Dermatology AI has a well-documented history of performing worse on darker skin tones when training data is unbalanced. A system that works well for some patients can quietly fail others.
• Missed context. A human clinician might notice that "acne" is actually a different condition, or that a patient's mood or medication history raises a red flag. Structured questionnaires only catch what they are designed to ask.
• Accountability. If an AI-issued prescription injures someone, who is responsible: the company, the supervising physician, or the state that waived the rules? Researchers have warned that this remains unclear, and Utah's leaders say pilot results will inform future laws on liability.
• Mission creep. Today it is acne creams and routine refills. State leaders have not yet defined which clinical decisions AI should never make. Without bright lines, the boundary could shift gradually without enough public debate.
• Evidence quality. Many performance claims come from the companies themselves. Independent evaluators are a big step forward, but their findings need to be published openly to build trust.

What This Means for You

• Faster care for simple problems may be coming. If these pilots succeed, routine conditions such as mild acne or stable chronic prescriptions could be handled in minutes instead of weeks, and often at lower cost.
• Ask who, or what, made the decision. You have the right to know whether a person or an algorithm reviewed your case. Utah's oversight board has pushed for clear patient disclosures; look for them wherever you get care.
• Know when to escalate. If a treatment is not working, causes side effects, or your symptoms change, ask to speak with a licensed clinician. Good AI services make that path easy.
• Protect your images and data. Before uploading photos of your body to any health app, read how they are stored, who can see them, and whether they are used to train the AI.
• Don't stretch the tool. If you are pregnant, have a complex medical history or take several medications, an AI designed for simple cases may not be right for you.

For Health-Care Professionals and Leaders

Utah is building a template that others may copy: phased autonomy, a narrow formulary, mandatory escalation pathways, and independent evaluators. Clinicians should engage early, through licensing boards, professional associations and comment periods such as the FDA's open docket, rather than reacting after policy is set. Health-system leaders should ask vendors for independent, subgroup-level performance data, including results across skin tones, ages and languages, along with clear escalation logic and audit trails. Expect your role to shift toward supervising, auditing and handling the complex cases AI escalates. That requires new skills in reviewing algorithmic decisions and new workflows that count oversight as real clinical work.

Wrapping Up

Utah's decision to let AI write a first-time prescription, however narrow, marks a turning point. For years, health AI has been framed as a helper that drafts, flags and suggests while a human decides. These pilots test something new: a carefully fenced space where software can act, with human oversight that steps back only as the evidence builds.

The early data from Utah's first experiment suggest caution rather than recklessness, and the addition of independent evaluators answers some of the sharpest criticism. But real questions remain about bias, accountability and where the line should be drawn. The rest of the world is watching. If Utah gets it right, millions of people stuck on waiting lists could get faster care for everyday problems. If it gets it wrong, it could set back public trust in health AI for years. Either way, the debate over whether a machine may prescribe is no longer theoretical. It is already happening.