The Doctors Not AI Act of 2026 would set a clear boundary for artificial intelligence in U.S. health insurance claim reviews: AI could assist, but it could not issue or dictate an adverse benefit determination that involves clinical judgment. A qualified licensed health professional would have to make the decision after independently evaluating the person's circumstances.
Representatives Greg Landsman, Buddy Carter, Kim Schrier, and Tom Barrett introduced H.R. 10210 on 1 September 2026. The bill has been referred to three House committees. It has not passed either chamber and creates no current duty. Its operational design is still worth examining because it connects human judgment, disclosure, and evidence retention in one proposed control chain.
The bill draws a boundary around clinical judgment
The bill covers algorithms, predictive models, machine learning systems, and automated decision software used to materially influence coverage decisions. For an adverse determination involving medical necessity, appropriateness, level of care, effectiveness, clinical guidelines, or another question requiring evaluation of a patient's needs, the proposed rule would require a licensed professional with relevant training and experience.
That professional could not treat the system's output as presumptively valid. The text requires an independent evaluation using generally accepted standards of care. AI may still support claim processing, but the final clinical reasoning cannot be reduced to a formal human sign-off on an automated recommendation.
This distinction matters for governance teams. A human checkpoint is useful only when the reviewer has the authority, context, competence, and time to disagree. A workflow that presents one recommendation without the underlying evidence can satisfy the appearance of review while preserving automation bias.
Disclosure and administrative records are part of the control
If AI was used, the denial notice would have to state that fact, describe the system's role, and identify the licensed professional who made the decision. The administrative record would also need to describe the system, its role, its outputs, scores, recommendations, or determinations, and evidence of the professional's independent judgment. Those materials would be available to the covered person on request.
The proposal therefore treats explanation as a record-keeping problem, not merely a sentence in a notice. A useful record must connect the exact claim, the model or system version, source information, generated output, reviewer, decision basis, and final communication.
Maetra's AI audit evidence guide outlines the kinds of records teams can preserve around consequential automated decisions. The AI approval workflow guide explains how to define a real decision point instead of adding a cosmetic reviewer step.
A practical control map for insurers and plan administrators
The bill is only proposed, but teams can use its structure to test current claim-review systems.
| Control question | Evidence to retain |
|---|---|
| Which AI system influenced the review? | System name, version, owner, purpose, data sources, and change date |
| What did the system produce? | Output, score, recommendation, confidence information, and known limitations |
| Who made the clinical decision? | Reviewer identity, licence, relevant expertise, assignment, and timestamp |
| Was the review independent? | Patient-specific facts considered, departures from the model output, and decision rationale |
| What was disclosed? | Final notice, description of AI's role, named decision-maker, and delivery record |
| Can the record be reproduced? | Input provenance, system configuration, policy version, retained output, and access log |
An inventory is the starting point. Teams need to know every model, rules engine, vendor service, and automation that can influence a claim, even if it does not issue the final denial. The bill's definition focuses on material influence, which reaches farther than a narrow list of systems labelled as AI.
Mental health parity receives separate attention
The proposal would treat the use of an AI system in utilization review as a treatment limitation for federal mental health parity analysis. It would also require enough information about the system's function and effects for regulators to compare how it operates for mental health and substance use disorder benefits against medical and surgical benefits.
This does not prove that a particular system discriminates or violates parity law. It would create a route for examining whether automated review changes access differently across benefit categories. Teams would need evidence about actual operation, not just design documents.
What remains uncertain
H.R. 10210 is an introduced bill. Committee review, amendments, votes, enactment, and implementation may never occur. The effective date in the text would apply to plan years beginning after a future enactment threshold, not immediately. The bill also does not specify a technical audit standard, retention period, model-validation method, or minimum review time.
Independent reporting from The Washington Post confirms the bipartisan introduction and the main human-review and disclosure provisions. It also notes a material limitation: a doctor in the loop is not automatically an effective safeguard if the reviewer lacks time, information, or freedom to challenge the automated recommendation.
Maetra analysis
The most useful lesson is the difference between a named human and an accountable decision. Organizations can test that difference today. Select one adverse-decision workflow and map every system that influences it. Record which facts reach the reviewer, what the system recommends, which policy applies, whether the reviewer can override it, and what evidence survives an appeal.
Then test the workflow with a case where the model output should be rejected. If the reviewer cannot see the relevant context, explain a departure, or preserve the decision basis, the human checkpoint is not yet a reliable control.
The proposed act does not ban AI from health insurance operations. It defines a narrower principle: when clinical judgment determines access to care, assistance may be automated, but accountability, disclosure, and the decision record should remain human and inspectable.