Toward Responsible AI in Health Insurance Decision-Making is a Stanford HAI policy brief, released February 2026 in the HAI Policy & Society series and authored by Michelle Mello, Artem Trotsyuk, Abdoul Jalil Djiberou Mahamadou, and Danton Char. It examines AI in health insurance utilization review, primarily prior authorization and claims adjudication, and argues that responsible deployment in this domain is structurally difficult. It documents that 84% of large health insurers in 16 states were using AI for some operational purposes by 2024, and frames the gap between that adoption and the regulatory frame governing it as the source of the wrongful-coverage-denial controversies and the resulting policy attention.
- Authors: Michelle Mello, Artem Trotsyuk, Abdoul Jalil Djiberou Mahamadou, Danton Char
- Series: HAI Policy & Society
- Released: February 2026
- Source: hai.stanford.edu
Summary of argument
The brief documents adoption rates, the decision-points where AI is being deployed, and the structural reasons the authors give for why responsible deployment in this domain is hard. Per the survey the brief cites, 84% of large health insurers in 16 states were using AI for some operational purposes by 2024. AI is deployed primarily on prior authorization, which the brief describes as the most resource-intensive utilization-review process, and on claims adjudication. The authors attribute the uptake to the hope that AI will streamline resource-intensive tasks and automate clearly approvable requests, which they describe as having legitimate efficiency rationales.
The brief connects the same deployments to wrongful-coverage-denial controversies at scale, citing UnitedHealth's nH Predict, Humana's algorithmic determinations, and Cigna's PXDX system, which it identifies as drivers of the current wave of policy attention.
The responsible-deployment gap
The central argument is that responsible AI deployment in health insurance is structurally hard for reasons distinct from AI deployment in other settings. The brief identifies five reasons:
- Information asymmetry. Patients cannot effectively contest denials they do not know are AI-generated, or whose criteria they cannot inspect.
- Velocity asymmetry. AI denials are issued at machine speed, while appeals are processed at human speed.
- Misaligned incentives. Insurer-deployed AI is optimized to insurer cost-control objectives; patient welfare and physician judgment are not the optimization target.
- Fragmented oversight. State insurance commissioners, federal Medicare/Medicaid CMS rules, ERISA preemption, and tort doctrine all apply unevenly across plan types.
- Definitional ambiguity. What constitutes "AI" in a coverage-determination context is contested; many insurer tools sit on a spectrum from rule-based decision trees to large language models, and the regulatory regime treats them inconsistently.
Policy recommendations
In its Key Takeaways framing, the brief argues that responsible deployment requires transparency about AI use in coverage determinations; human review for denials, particularly in clinically consequential cases; audit trails and outcome monitoring for AI tools used in utilization review; patient-facing disclosure when AI has contributed to a coverage determination; and cross-jurisdictional harmonization to address the fragmented-oversight problem. The authors note that these align with the regulatory direction many state legislators are pursuing, including provisions in Colorado, California, and Texas AI legislation that touch healthcare deployments.
Provenance and reception
The brief is the first primary-record source in this collection on AI in health insurance utilization review, addressing the gap between widespread industry adoption (84% of large insurers) and the still-thin regulatory frame for responsible deployment. It is a position document and is treated as such: the documented adoption rates and the structural-gap framing are well-supported by the brief's underlying citations, while the prescriptive policy recommendations reflect the authors' judgment about what responsible deployment requires, on which observers may differ in emphasis (for example, transparency-first versus human-review-first).
The brief serves as the canonical primary reference for AI prior-authorization adoption rates and the responsible-deployment gap, and is the relevant citation for healthcare-AI adoption figures across related pages. Healthcare — AI Deployment documents the sector-level deployment; the empirical adoption-rate baseline here supports future updates to that page and to Colorado AI Act (SB 24-205) and SB 25B-004 (Date Amendment) and California AB 3030 (Healthcare AI Disclosure), both of which touch healthcare AI. The brief tracks insurer-side AI use, which complements the clinician-side terrain covered in AMA Physician AI Sentiment Report (2026) (physician AI sentiment) and the clinical-AI deskilling concern in Lancet Endoscopist Deskilling Study (2025); the coverage-denial concern and the deskilling concern are responsible-deployment questions with different structural mechanisms. Mobley v. Workday, Inc. addresses the AI-vendor-as-employer-or-agent question under Title VII, the ADA, and the ADEA in employment contexts; the brief identifies a structurally similar but separately litigated question for healthcare, namely the AI vendor's liability for wrongful coverage denials. On the regulatory frame, the brief notes that the FDA — Food and Drug Administration (AI Deployer) regulates AI as a medical device and National Institute of Standards and Technology (NIST) sets cross-sector standards, with health insurance AI sitting in the gap between the two, where neither is the primary regulator. Stanford HAI is the publisher of the policy-brief series, and the brief is a domain-specific anchor for algorithmic-decision-making policy (Algorithmic Decisionmaking, planned).
Confidence note
High for the documented adoption rates and the structural-gap framing, which are well-supported by the brief's underlying citations. Medium for the prescriptive policy recommendations, which reflect the authors' judgment about what responsible deployment requires and on which reasonable observers may differ in emphasis.
Relationships
- supports: Healthcare — AI Deployment (canonical adoption-rate anchor), Algorithmic Decisionmaking (planned domain-specific anchor), Colorado AI Act (SB 24-205) and SB 25B-004 (Date Amendment) (healthcare AI provision context), California AB 3030 (Healthcare AI Disclosure) (healthcare AI provision context)
- related: AMA Physician AI Sentiment Report (2026) (physician-side companion), Lancet Endoscopist Deskilling Study (2025) (different responsible-deployment failure mode), Mobley v. Workday, Inc. (parallel vendor-liability question in employment), Stanford HAI, FDA — Food and Drug Administration (AI Deployer) (regulatory-gap context), National Institute of Standards and Technology (NIST)
- instance-of: Responsible AI Deployment