In early October 2026, Utah authorized something unprecedented: an acne patient pointed their smartphone front camera at their face, scanned it, checked off a few medical history items, and walked away with a prescription for topical medication. The prescription bore the name and ten-digit professional licensing identifier of a licensed Utah physician, but the entire severity assessment and the formulation selection from among eight approved topical medications were completed autonomously by background algorithms. Under the agreement, during the initial phase of the pilot, each prescription still requires case-by-case pre-review by two physicians; what Utah truly handed over for the first time was the authority of “autonomous algorithmic prescribing” itself.
The tech community often treats such news as another leap in AI physician capabilities. But looking at the longer arc, having algorithms make autonomous medical judgments is hardly new. As early as 2018, U.S. federal regulators approved an AI retinal disease screening system that operates with autonomous algorithmic interpretation, requiring zero ophthalmologist involvement. Given that autonomous machine diagnosis was permitted eight years ago, why is issuing a prescription for a patient still treated by the entire medical establishment as an impassable red line today?
A physician the patient has never met has their license printed on the prescription. What defines the boundary between machines and medicine is not how smart the algorithms are, but to whom real-world authority to act is granted. No matter how superior a model’s scores or how reliable its diagnoses, it is merely providing information, rendering assessments, and estimating probabilities. Issuing a prescription, however, invokes legally recognized authority to place a physical drug into a patient’s hands. Rather than arguing in court over whether AI can practice medicine autonomously, Utah seized upon a jurisdictional division in U.S. healthcare regulation: medical licensure and prescribing authority are governed by individual states, while medical device clearance is governed by the federal government—two halves of a divided mandate. Bypassing device clearance, Utah used an administrative contract promising temporary non-enforcement to transfer prescribing authority to algorithms from the state-regulated side of medical practice.
In 2018, medical technology company Digital Diagnostics developed an autonomous screening system named IDx-DR (later rebranded as LumineticsCore) targeting diabetic retinopathy, a common complication among diabetic patients. While the traditional process required patients to schedule appointments with specialists for dilated eye exams, IDx-DR moved the workflow to primary care clinics: nurses capture retinal photographs, and a deep learning model analyzes microaneurysms and exudates in real time, delivering a verdict within minutes on whether the patient has more than mild retinopathy.
The system received FDA De Novo classification, review number DEN180001. De Novo is a specialized regulatory pathway established by federal regulators for novel medical devices that lack legally marketed predicate devices but present manageable risks. The milestone significance of this clearance lies in the fact that it was the first medical system permitted to produce an autonomous screening diagnosis without an image read by a specialist physician. If the algorithm yields a negative result, the patient is recommended for a rescreen in 12 months; if positive, the system directly outputs a referral recommendation. Legally speaking, algorithms obtained the legitimate standing to render autonomous diagnostic conclusions eight years ago.
However, this “delegation” at the diagnostic level did not ignite a revolution in autonomous medical practice. Eight years later, the medical community has instead placed a firm ceiling on algorithmic autonomy. In 2026, Qiu et al. published a study on a seven-level framework for medical AI autonomy in Nature. Across three dimensions—autonomy, automation degree, and task breadth—they constructed a seven-level ladder spanning Level 0 through Level 6, and conducted a systematic audit of more than 2,800 medical AI devices approved across major global regulatory jurisdictions.
The audit revealed that not a single one among thousands of approved medical AI products worldwide exceeds Level 3. By definition, Level 3 already represents conditional automation, where algorithms make decisions within an extremely narrow scope and physicians intervene only when anomalies arise; an autonomous clinician capable of independently conducting consultations, ordering diagnostic tests, and prescribing medication resides at Level 5. Prior imaging interpretation and screening models, even those achieving autonomous diagnosis, remain capped within L3.
Regulators have been willing to let algorithms participate in diagnosis because diagnosis produces information. Detecting a lesion merely changes the understanding of the patient’s condition for both clinician and patient. As for what medication to administer or what surgery to perform next, the executing party remains a human physician, who signs the medical record and bears the legal consequences of treatment and misdiagnosis. As long as algorithms do not overstep, the closed loop of medical liability remains intact, with physicians retaining ultimate control. The institutional wall stands firm precisely at the threshold where information attempts to cross over into physical action.
Diagnosis depicts a physiological state; a prescription is a legal directive mobilizing societal material resources. If a patient receives a diagnostic report indicating dermatitis or hypertension and locks it in a drawer, nothing in the physical world changes. But once a prescription is generated, it immediately steps outside the clinic and into the societal supply chain: pharmacists dispense and deliver prescription drugs based on it, and these molecules enter the human body, initiating physical interventions on living tissue in the real world. (The prescription drugs referred to here are standard prescription drugs, excluding controlled substances such as opioids, which Utah’s pilots do not touch.)
What follows is a complete transfer of legal liability. What is signed on a prescription is not just a name, but a professional credential and joint legal liability. In the United States, every practitioner legally qualified to prescribe holds a unique ten-digit identifier: the National Provider Identifier (NPI). Once a prescription carries this identifier, pharmacies can identify the prescriber, and insurance companies can adjudicate and reimburse claims. If a patient suffers acute organ failure or a severe allergic reaction due to drug-drug interactions, the physician named on the prescription is the point of liability recourse. A prescription is the institutional switch in modern healthcare that converts abstract judgment into physical intervention and pins potential malpractice liability onto a specific individual.
Utah’s drive to let algorithms prescribe autonomously met immediate, direct resistance from medical professional associations. On April 20, 2026, 11 of the 14 members of the Utah Physicians Licensing Board signed a joint letter to the Utah Department of Commerce in protest, demanding an immediate halt to the ongoing automated prescription renewal pilot. Board members wrote that renewal is a statutory duty reserved for licensed practitioners and cannot be delegated; when facing patient renewal requests, practitioners must reassess clinical stability, monitor potential side effects, screen for newly emergent contraindications, and verify drug interactions, lest patients remain on outdated or suboptimal regimens for months or even years.
The next day, the Division of Professional Licensing (DOPL) and the Office of Artificial Intelligence Policy (OAIP) issued a joint response denying the licensing board’s request to pause, stating that the program retained case-by-case physician review. The Federation of State Medical Boards (FSMB) stated in August 2026 that AI is not ready for independent licensure; the American Medical Association (AMA) reiterated in June 2026 that AI should augment rather than replace physicians, with its CEO John Whyte noting that removing physicians from the prescribing loop creates systemic risks to patient safety.
From an institutional perspective, prescribing authority has long been delegable across professional chains. State laws already provide well-established conduits: physicians can delegate partial prescribing authority to advanced practice registered nurses or physician assistants, sign collaborative practice agreements with clinical pharmacists to titrate dosages, or authorize nurses via standing orders to administer medications according to protocols. Prescribing authority is itself an institutionally chainable conduit. Utah’s innovation did not conjure an algorithmic physician out of thin air; it merely followed existing delegation logic, transferring prescribing authority previously delegated to allied health personnel to a set of computer protocols. It was precisely by recognizing the delegable nature of prescribing authority that state administrative agencies pried open a breakthrough.
Utah was able to pry open this seam through a regulatory sandbox mechanism. Rather than debating in courts or ethics committees whether algorithms surpass humans, the state government crafted an administrative tool. On May 1, 2024, Utah formally enacted Senate Bill 149 (SB 149), creating the Office of Artificial Intelligence Policy (OAIP) within the Department of Commerce and granting it the authority to enter into Regulatory Mitigation Agreements (RMAs) with selected businesses.
This RMA is essentially an administrative contract promising forbearance from enforcement, not a statute passed by a legislature. The state government promises the business: as long as it operates within the agreed scope, regulatory agencies will not pursue administrative penalties or civil prosecution for specific unlawful practice provisions. Entering the sandbox does not confer official endorsement, and the state explicitly disclaims liability for personal injury. One example is Nolla Health and its operating entity Magic Health, launched on October 5, 2026; its executed Nolla RMA Agreement Schedule A waives unlawful and unprofessional conduct provisions § 58-1-501(1)(2), unlawful conduct provisions under pharmacy and medical practice acts §§ 58-17b-501, 58-67-501, 58-68-501, unprofessional conduct provisions §§ 58-17b-502, 58-67-502, 58-68-502, and restrictions on substantive provider-patient communication under telemedicine statutes § 26B-4-704(6). Section 19(D) of the agreement also explicitly extends safe harbor protections to collaborating supervising physicians. The agreement was jointly executed by the director of the OAIP and the director of the Division of Professional Licensing (DOPL). In 2025, Utah extended the repeal date of the entire Artificial Intelligence Policy Act to July 1, 2027; in 2026, the legislature amended the law to add auditing requirements.
Why this proved viable stems from the division of labor in U.S. healthcare regulation—a judgment based on institutional division: the FDA statutorily regulates “devices,” exercising review over software that qualifies as a medical device; medical licensure and prescribing authority are powers reserved to the states, outside the FDA’s device review remit. If a company sought to push autonomous AI prescribing through the FDA device pathway, it would face formidable clinical trial hurdles. Rather than pursuing device clearance, Utah categorized algorithmic prescribing under intra-state medical practice, using non-enforcement forbearance to shield against state-level practice liability, thereby handing over prescribing authority first. This does not waive any federal obligations: if the FDA later determines that such AI must be regulated as a medical device, state authorization would still have to yield. At the federal level, the only relevant bill is the Health Technology Act of 2025, H.R. 238, which aims to recognize state-authorized and FDA-approved AI as prescribing entities, but it has languished in committee since its introduction in January 2025.
In practice, Nolla Health became the first pilot approved to issue initial prescriptions. A patient uploads facial photographs; a vision Transformer model extracts lesion features and outputs a severity score (Investigator’s Global Assessment, IGA 0 to 4), then combines medical history to recommend a regimen from eight approved topical medications. The pilot is strictly restricted to topical treatments, excluding oral medications, steroids, isotretinoin, and severe or nodulocystic acne. When the prescription is routed to a pharmacy, the signature block is populated with the collaborating supervising physician’s NPI. The official state position is: prescriptions are signed directly by licensed physicians or executed on their behalf via AI protocols, with records mandatorily flagged as AI-initiated. Through this, the algorithm secured the credential needed to mobilize real-world pharmaceuticals.
Following Nolla’s approval, numerous media outlets reported that AI had begun prescribing autonomously. But from the perspective of an administrative contract, institutional approval is not equivalent to AI operating fully without human supervision. Nolla’s operations are divided into three phases: Phase 1 runs for at least 4 weeks and covers at least 100 patients, with each prescription requiring case-by-case pre-approval by two independent licensed physicians; Phase 2 runs for at least 8 weeks and after a cumulative 500 patients, where AI transitions to direct prescribing while physicians review all records weekly; Phase 3 begins after reaching 750 cumulative patients, transitioning physicians to routine spot-checks of at least 10% monthly. Advancement through each phase requires prior written approval from the OAIP, rather than automatically triggering once patient quotas are met. Currently, Nolla remains in Phase 1 with two-physician case-by-case pre-review.
Doctronic, the startup that first entered the sandbox, reflects the same pace of progression. Having signed its agreement on October 24, 2025, Doctronic specializes in renewals for 191 chronic maintenance medications, strictly excluding controlled substances, dosage adjustments, or new prescriptions. Nearly a year later, as of October 2026, it remains in Phase 1 with case-by-case physician review. On May 19, 2026, Utah published five-month performance outcome data from Doctronic’s operations: AI recommended renewal in 72% of cases, with reviewing physicians agreeing in 91%; AI recommended escalation/referral in 28% of cases, with physician agreement at 69%. Official notes state that this dataset is entirely vendor-physician self-reported, and third-party independent verification has been initiated.
In contrast to this lies an unresolved legal liability vacuum. Even though the sandbox mandates the purchase of AI liability insurance, legal scholar Michelle Mello noted in JAMA Health Forum that Doctronic’s terms of service disclaim liability for system accuracy and personal injury, while its concession contract waives the obligation to provide patients with full medical records. Across the United States, judicial precedent regarding autonomous prescribing litigation currently stands at zero. Adjacent litigation, such as Pennsylvania v. Character.AI in May 2026, alleged unauthorized practice impersonating a licensed psychiatrist rather than a prescribing mishap. Stanford HAI examined 803 relevant cases (spanning health and non-health domains) and found that judgments awarding plaintiffs personal injury damages due to software defects were exceedingly rare. Should an incident occur with software disclaiming liability, the risk of damages and license revocation falls upon the supervising physician supplying their NPI code—a liability attribution mechanism that has never yet been tested by a real-world incident.
Empirical evidence from security evaluations has further amplified concerns. Doctronic previously cited vendor self-reported data in a non-peer-reviewed preprint claiming a 99.2% concordance rate with physicians across 500 acute cases. Red-teaming by security firm Mindgard in January and March 2026 demonstrated that simple jailbreaks could prompt the system to recommend triple-strength doses of OxyContin (a potent opioid), classify methamphetamine as an unrestricted therapeutic, and generate false vaccine information. The system has never undergone prospective clinical trials for its prescribing functions, and its acute-care sample was noticeably disconnected from the sandbox’s chronic maintenance population.
Looking globally, Utah’s administrative exemption appears remarkably isolated. Across examined jurisdictions—including China, the European Union, the United Kingdom, Canada, Japan, South Korea, Israel, Saudi Arabia, India, and Kenya—not a single one permits AI to prescribe independently. China’s regulatory framework imposes explicit bans on algorithmic prescribing; for instance, Tianjin’s 2026 Implementation Measures for the Supervision of Internet Diagnosis and Treatment (Trial) explicitly prohibits AI from automatically generating prescriptions, and the National Medical Products Administration (NMPA) prohibited AI review of prescriptions in its Compliance Guidelines for Online Retail of Prescription Drugs issued on May 25, 2026. In Europe, Article 14 of the EU AI Act mandates that high-risk medical AI must be subject to effective human oversight, allowing individuals to override system outputs in specific circumstances. Member state laws likewise reserve prescribing authority to licensed natural persons. The Digital Omnibus postponed the compliance date for certain embedded medical devices from August 2, 2027, to August 2, 2028, but the regulatory principles remain unchanged. In the UK, the MHRA characterizes algorithms as devices, reserving the legal right to prescribe exclusively to licensed prescribers. The World Health Organization (WHO), in its 2021 Guidance on Ethics and Governance of AI for Health and 2024 guidance on large multimodal models, emphasized that safeguarding human autonomy is paramount and human physicians must invariably remain at the core of decision-making.
Across other U.S. states, attitudes toward medical AI are equally cautious. General sandboxes in Arizona and Wyoming have not accepted applications for medical prescribing (Wyoming remains in discussions), public activity from Texas’s AI sandbox (TRAIGA) stands at zero, and automated licensure legislation promoted by the Cicero Institute in Idaho and Iowa died upon legislative adjournment. More states are instead legislating restrictions: Delaware barred non-human entities from using physician titles, while California, Illinois, and Colorado have progressively passed laws restricting independent clinical decision-making by AI.
Why Utah, and why now? It stems from an analytical inference woven from four converging forces: first, advances in multimodal generative models, which made end-to-end lesion image processing and medical consultations viable; second, Utah’s SB 149, which empowered administrative officials with enforcement forbearance authority; third, a federal regulatory vacuum—FDA workforce reductions in 2025 and guidance issued on January 6, 2026, loosening oversight on clinical decision support, leaving the agency silent on Utah’s experiment; and fourth, an acute scarcity of primary specialist care in Utah, where the average dermatology wait time reaches 61 days, 11 counties have no resident dermatologist, and prescription refills account for nearly 80% of prescribing activity. It is precisely because these pain points concentrate in scarce specialties and repetitive tasks that the sandbox chose to anchor itself in chronic refills and topical acne treatment, designing a phased transition mechanism.
When observing the evolution of AI in heavily regulated sectors, first look at action authorization. A model scoring high marks on an exam remains within the realm of information production. The criterion for determining whether a system has entered the era of autonomy is whether it has secured legal authority to mobilize real-world material resources. So long as it remains at the advisory level, the system remains inside the authorization wall. The true watershed occurs at the instant it is granted the authority to trigger real-world consequences, initiate physical interventions, and bind legal liability.
Second, distinguish institutional admission from real-world unsupervised operation. Even securing a sandbox pass with enforcement forbearance does not mean the system is operating independently in reality. Between signing an agreement and genuinely removing human physicians lies a protracted and potentially reversible gauntlet of validation phases and spot-check thresholds. Many pilot programs linger indefinitely in early stages of human-machine collaboration.
Third, innovation born within administrative seams is inherently fragile. Enforcement forbearance relies on administrative discretion rather than durable statutory guarantees provided by legislation: it hinges on the inclinations of local officials and federal forbearance, and remains vulnerable to lawsuits brought by medical professional associations at any time. In the present landscape of mutual liability disclaimers and absent legal precedent, the occurrence of a single major adverse event could cause this structure of forbearance to collapse rapidly.
Health law scholar Michelle Mello once invoked a classic legal adage: once the camel gets its nose under the tent, the whole body follows. Utah’s sandbox mechanism moved from permitting chronic renewals in October 2025 to authorizing de novo prescriptions for new patients in October 2026—in less than a year. A single tube of topical acne cream has already set an institutional precedent.
The boundaries of AI in high-stakes domains will not be established organically through parameter scaling, nor will they be arbitrated by benchmark leaderboards. They are being forged by contractual clauses signed inside administrative conference rooms. When confronting a prescription that touches human physical health, whether authority remains held by natural persons who can be subpoenaed in court or is entrusted to code is an institutional struggle that has only just begun.