Why it matters
Two things are happening at the same time. Clinical data is being pooled and turned into models. And the licensure rules that stand between those models and patients are being rewritten. Here is the record, with links to the primary sources so you can check every line.
Data flows one direction and rules flow the other. Clinicians supply the documentation, the reasoning, and the outcomes that make clinical AI possible. Vendors and their trade groups then work to remove the requirement that a clinician be involved when that AI is used.
None of the individual steps look dramatic. A de-identification clause. A pilot program. A model bill about "access." Stacked together, they move the authority to make clinical decisions away from the people who carry the license and the liability.
The load-bearing legal fact: under HIPAA's Safe Harbor method, once the 18 listed identifiers are removed, the information is no longer protected health information. HIPAA's use and disclosure limits stop applying to it — which is why "we only train on de-identified data" is a description of what the law permits, not a promise to you.
Chronological. Every item links to a primary source below.
May 1, 2024 Utah · S.B. 149
The Artificial Intelligence Policy Act creates an Office of Artificial Intelligence Policy inside the Department of Commerce, along with an AI Learning Laboratory and the power to sign "regulatory mitigation agreements" — tailored, time-limited relief from state rules for companies testing AI. This is the mechanism everything below runs on.
Jan 7, 2025 U.S. House · H.R. 238
The Healthy Technology Act of 2025, introduced by Rep. David Schweikert, would amend the Federal Food, Drug, and Cosmetic Act so that artificial intelligence and machine learning technology can qualify as a practitioner licensed by law to prescribe drugs — provided the state authorizes it and the FDA has cleared the software. It sits in the House Energy and Commerce Committee; a near-identical bill was introduced in 2023. The American College of Physicians wrote to object in March 2025, arguing AI should complement rather than supplant physician decision-making.
2025 American Medical Association
The AMA adopts policy holding that AI in health care must remain assistive and under physician oversight, not an autonomous decision-maker — and continues to call it "augmented intelligence" for exactly that reason. It also pushes for transparency where AI drives prior authorization and coverage denials.
Oct 2025 – Oct 2026 Utah · Doctronic agreement
Utah's Office of AI Policy signs a regulatory mitigation agreement with Doctronic, announced publicly on January 6, 2026. Roughly 20–200 non-controlled maintenance medications for chronic and mental-health conditions are in scope; controlled substances, injectables, and new prescriptions are not. In phase one a licensed physician still authorizes each renewal. Phase two — conditional on hitting safety benchmarks — lets the AI send renewals directly to pharmacists, who retain the ability to escalate. State officials describe the standard as "doctor, not device." However it is framed, a state has now waived the rule that only a licensed human participates in prescribing, and the FDA has signaled it does not plan to intervene.
Dec 11, 2025 Executive Order
"Ensuring a National Policy Framework for Artificial Intelligence" directs the Attorney General to stand up an AI Litigation Task Force within 30 days to challenge state AI laws as unconstitutional or preempted, and directs Commerce to publish, within 90 days, a list of state AI laws considered onerous and suitable for referral. The state-level guardrails on clinical AI are squarely inside the blast radius.
Jan – Feb 2026 Cicero Institute · model bill
The AI Medical Services Act — a model bill written for states to adopt — would create an "AI Augmented & Autonomous Service Provider" (AAASP) license with four autonomy tiers. L0 is advisory. L3 is fully autonomous: the AI is authorized to independently diagnose, treat, or prescribe. Its own summary is blunt about the premise: "If an AI acts like a doctor, it is licensed, insured, and regulated like a doctor." Supporting provisions include a two-year provisional license, malpractice insurance and bonding, and value-based payment. Framed as a fix for provider shortages and rural access, it establishes a licensure path that runs parallel to — rather than through — the physician licensure system your medical board administers.
2025 – 2026 Statehouses
States are also legislating the other way. Washington's S.B. 5395 says only a licensed physician or health professional may deny a prior authorization request. Colorado H.B. 1139, Alabama S.B. 63, Georgia S.B. 444, Iowa H.F. 2635, and Utah S.B. 319 impose versions of the same requirement on AI-driven coverage decisions. Colorado, Maine, Rhode Island, Tennessee, and Vermont have restricted AI from independently providing therapy. Iowa H.B. 475 requires disclosure before AI records a visit for transcription. This is the layer the December executive order targets.
The capability being licensed had to be built out of something. It was built out of clinical work.
A competitor's dataset holds de-identified records covering more than 300 million patients and 16.3 billion encounters from 310 health systems. Its foundation models were pretrained on 118 million patients and about 115 billion medical events, to predict a patient's future course. Your documentation is in there — contributed under an agreement your health system signed.
Enterprise scribes are deployed across hundreds of health systems, and training on stored, de-identified encounter data is standard practice across the category. The most intimate record in medicine — the unedited conversation between a doctor and a patient — has become a training corpus.
Free clinical-reference AI is generally free because the clinician is the product: usage is tracked and monetized through advertising targeted by specialty and interest. At least one health system has told its physicians not to enter PHI into one such tool. Your questions reveal your uncertainty — that is valuable data about the practice of medicine.
AI has no license to practice medicine, and the case for giving it one rests on a false comparison — autonomous AI against an overstretched status quo. The right comparison is autonomous AI against a clinician equipped with strong AI. That pairing is safer, faster, and more accountable than either alone, and it does not require anyone to give up the license or the liability that make medicine trustworthy.
The way to get there is to build tools good enough that clinicians want them, on terms where the value that clinical work creates accrues to clinicians rather than to whoever is selling it back to them. That is the whole product thesis.
Read them directly. We would rather you check our summary than take it.
A note on fairness. Several of the organizations named here make genuinely useful products, and de-identified research has produced real clinical value. The objection is not that anyone broke the law — mostly they didn't. It is that the current arrangement lets the party holding the data decide, unilaterally, what medicine becomes. Terms differ by vendor and by contract; check yours.
That's what we're building. Come use it early and tell us where it falls short.
Request early access