Will any US state legally authorize autonomous AI diagnosis and prescription for primary care (no human sign-off per encounter) before 2031?

TL;DR

I forecast a 41.5% chance of YES. Utah already has a binding AI prescription-renewal pilot, while Iowa and Idaho introduced bills that could authorize fully autonomous diagnosis and prescribing, but none has crossed the required line. The most likely YES is a narrow pilot for one condition; physician opposition, federal ambiguity, and the lack of real-world autonomous clinical evidence still make NO more likely.

Context

No state currently qualifies. Utah’s Doctronic pilot remains limited to renewals of medications previously ordered by a licensed physician; it forbids new prescriptions or treatment changes, remains in a phase requiring practitioner authorization for every request, and keeps a physician attached to each prescription (Utah Office of AI Policy).

The policy pipeline is real. Iowa HSB 766 received a favorable subcommittee recommendation on March 24, 2026, while Idaho H945 was introduced on March 25 but did not advance beyond committee referral (Iowa bill history, Idaho bill history). Montana’s Economic Affairs Interim Committee reviewed the same model legislation on May 19, 2026, but did not endorse or introduce a bill (Montana Legislature).

Evidence

The historical backbone argues for caution. State scope-of-practice changes normally diffuse slowly even when the new prescribers are trained humans: a statutory dataset covering 1998–2022 shows staggered expansion of nurse-practitioner independence rather than rapid national change (NBER). A January 2025 Vermont review listed only seven states with psychologist prescribing laws, although the first such law dated to New Mexico in 2002; Vermont then enacted its own law on April 27, 2026 (Vermont sunrise review, Vermont Act 84). Autonomous AI is a larger trust and liability jump. The counterweight is that this question needs only one state and allows one narrow, temporary pilot.

Iowa supplies the strongest positive evidence. Its bill defines fully autonomous AI as operating without human supervision or intervention for each case, creates an L3 independent-operation license, and explicitly allows L2 and L3 providers to issue non-controlled prescription-drug orders (Iowa bill text). The bill also says that a medical director’s role does not constitute practicing medicine in individual AI encounters. If enacted substantially intact and made operational for even one ordinary primary-care condition, it would satisfy the resolution. But neither Iowa nor Idaho’s first attempt received a chamber vote, so the current evidence shows an organized campaign, not a demonstrated legislative majority (Iowa history, Idaho history).

Utah is the strongest administrative route. Its law permits temporary agreements that modify specific regulatory requirements, although the statute is scheduled to repeal on July 1, 2027 unless extended or replaced (Utah Code §13-72-401, Utah repeal provision). The state has already continued its pilot despite a medical-board request for immediate suspension, showing that the board is influential but not an absolute veto player (board letter). Expansion from renewals to initial diagnosis and treatment would still be a large step: Utah’s early public report omitted the encounter count and said its sample was limited, so it provides little clinical evidence for removing oversight (May 2026 report).

Technology is advancing faster than real-world validation. MIRA completed end-to-end diagnostic and treatment workflows on 574 retrospective cases inside a simulated electronic-health-record environment, while AMIE was tested against 21 primary-care physicians on 100 simulated multivisit scenarios (Nature: MIRA, AMIE study). Nature’s own summary said neither system was ready for real clinical use (Nature research highlight). ARPA-H is nonetheless funding a program intended to produce the first FDA-authorized autonomous agentic cardiovascular-care system, which could create a regulatory bridge before 2031 (ARPA-H ADVOCATE).

Federal law is friction rather than a clear prohibition. Federal prescription law requires a practitioner licensed by law, and H.R. 238 would expressly clarify that qualifying state-authorized, FDA-authorized AI can be such a practitioner; the bill remains only introduced (21 U.S.C. §353, Congress.gov). FDA guidance also keeps patient-facing diagnostic and treatment software within potential device oversight (FDA guidance). Meanwhile, the AMA has adopted policy opposing autonomous clinical decision-making, and the FSMB has begun developing guidance for systems operating with limited or no physician supervision (AMA, FSMB).

My pathway model assigns 31% to an in-effect autonomous-provider statute, 13% to a qualifying sandbox conditional on no such statute, and 2.5% to another route such as a specially enacted pilot or board rule. The combined estimate is 41.47%, which I report as 41.5%. These are conditional pathways rather than independent state-by-state bets, because the same safety event, FDA decision, or lobbying campaign would move many states together.

What's non-obvious

The number of usable sandboxes is smaller than headline inventories suggest. Kansas expressly bars its sandbox from waiving licensing requirements, Delaware merely ordered the development of a framework and then enacted a law barring nonhuman entities from receiving medical licenses, and Wyoming’s statute is framed around medical digital assessments rather than prescribing (Kansas statute, Delaware HJR 7, Delaware HB 191, Wyoming statute). Utah and Texas are the two clear administrative pathways because Texas can allow approved participants to test AI systems without ordinary licensing or registration (Texas Business and Commerce Code, Chapter 553). This correction is why my estimate is below the near-50% forecasts.

The opposite correction keeps the estimate well above a remote probability. Iowa’s proposal is not a generic AI study bill. Its operative language nearly mirrors the resolution, and the first qualifying authorization could cover one low-acuity condition, one formulary, and consenting patients rather than a general-purpose AI family doctor (Iowa bill text).

Limitations

The resolution leaves some ambiguity about whether an effective statute creating a usable L3 licensing pathway is enough, or whether a specific vendor must receive a license before the state has truly permitted real patient care. I put the estimate near 47% under the first reading and near 35% under the second, then weight toward the stricter operational reading.

The evidence base is thin. There are only two directly qualifying introduced bills, Utah’s public pilot report gives no encounter count, and pending sandbox applications are not comprehensively public. Absence of another visible proposal is therefore not evidence that none exists. A single safe autonomous trial, FDA authorization, serious patient injury, or federal enforcement action could move the forecast by more than ten percentage points.