Autonomous AI Prescribing: The Clinician's Veto and Safety

The Clinician's Veto, an arXiv paper submitted on 23 Jun 2026, argues that autonomous AI systems that prescribe medications need three minimum architectural features before they can safely perform prescriptions without continuous clinician control. The authors tested their claims with a survey of 136 U.S. prescribing clinicians and use those responses to ground regulatory and technical recommendations.

What did the paper find?

The paper reports that clinicians surveyed would not permit autonomous prescribing without a calibrated confidence-based escalation mechanism; they preferred a competing-options summary when uncertainty was aleatoric but shifted to abstention when uncertainty was epistemic; and they were only willing to accept additional liability when inferential transparency enabled a substantive judgment under acknowledged uncertainty. The authors present these findings as evidence that clinician trust and liability preferences depend on per-decision confidence, the kind of uncertainty present, and the ability to inspect the model's reasoning at decision time.

The survey results undergird the paper's regulatory argument: aggregate model metrics are insufficient because they do not provide per-prediction calibrated confidence, they do not separate epistemic (model ignorance) from aleatoric (genuine clinical ambiguity) uncertainty, and they do not provide inferential transparency required to assign liability at the moment of decision.

What architectural requirements do the authors propose?

The authors propose three specific minimum architectural requirements: calibrated per-prediction confidence for action-gated thresholds; differentiated communication of epistemic versus aleatoric uncertainty; and inferential transparency at the moment of decision that permits liability allocation. Each requirement maps to a distinct safety or governance need identified in their survey-backed argument.

Calibrated per-prediction confidence would allow the system to escalate or withhold action based on a measured likelihood tied to a decision threshold. Separating epistemic from aleatoric uncertainty changes the communication pattern: clinicians in the study wanted competing-option summaries when the problem was genuinely ambiguous (aleatoric) but preferred the system to abstain when uncertainty came from model ignorance (epistemic). Inferential transparency means the system must expose enough of its reasoning at decision time for a clinician or institution to make a substantive judgment and accept or reject liability accordingly.

How would a compliant system behave?

A system meeting the authors' requirements would behave less as an autonomous agent and more as a heavily supervised decision-support tool, collapsing much of what "autonomy" conventionally means. In practice that means the system would compute a calibrated confidence score for each prescription, tag the uncertainty source as epistemic or aleatoric, expose the inferential basis for the recommendation, and trigger clinician escalation whenever confidence or uncertainty type crosses pre-set thresholds.

The paper frames this architecture as enabling a clinician veto and aligning liability with institutional actors who control design and deployment, rather than leaving accountability opaque. The authors link these architectural features directly to higher clinician adoption in their survey results.

Why it matters

Recent policy moves make this debate urgent: the paper notes that United States bill H.R. 238 and Utah's prescription-renewal pilot both authorize AI to prescribe medications in an agentic capacity. If regulators clear systems based only on aggregate performance metrics, the paper argues, deployments will lack the per-decision signals clinicians and institutions need to safely accept autonomy and liability. Requiring the three architectural features would constrain how much operational autonomy regulators grant and make accountability traceable to system design and deployment choices.

What to watch

Watch implementation outcomes from Utah's prescription-renewal pilot and any legislative progress on H.R. 238 for how regulators allocate authority to AI in prescribing. Also watch whether regulators or clearance pathways begin to require per-prediction calibration, uncertainty decomposition, or inferential transparency as part of their approval criteria.