Artificial intelligence could help make health care more accessible, but poorly designed implementation may deepen the same disparities it aims to solve, according to a commentary published in Learning Health Systems.

The commentary was authored by Sandeep Reddy, MBBS, DPH, PhD, Professor of Healthcare Management and Medical Informatics in the School of Public Health and Social Work, Queensland University of Technology, Brisbane, Australia.

It proposes a structured implementation framework built around 4 dimensions: accessibility, affordability, usability, and ethical regulation. Rather than cataloging AI’s capabilities optimistically, Dr. Reddy anchors the framework around the growing equity problem. The World Health Organization (WHO) has projected a global shortfall of 11 million health workers by 2030, with the steepest gaps in specialized fields and underserved regions. Geographic and financial barriers already exclude millions from basic care. AI, he argues, could help close those gaps or deepen them, depending on how it is deployed.

“The core contribution is a structured, actionable framework for policymakers rather than another optimistic survey of AI capabilities,” Dr. Reddy told AACE Endocrine AI. “The inclusion of a Learning Health System lens is critical because it shifts the framing from deployment to ongoing adaptation.”

That design moves the framework from one-time implementation guidance toward a more dynamic model. Under a Learning Health System approach, AI tools would be expected to generate ongoing signals about real-world performance, equity gaps, and community needs, with those signals feeding back into health system improvement rather than treating initial deployment as a finished product.

The commentary draws on real-world examples to illustrate both the potential and the complexity of AI-enabled access. These include cited examples of AI-assisted tuberculosis detection in India using cough sound analysis, locally trained language models in Kinyarwanda to support frontline health workers in Rwanda, and AI triage systems in the United Kingdom’s National Health Service that directed patients to appropriate care settings.

Failure Modes Identified

But Dr. Reddy's piece cautions that these examples do not tell the full story of AI implementation risk. It identifies several failure modes that governance frameworks may fail to anticipate: generative AI hallucinations that produce authoritative-sounding but incorrect clinical guidance; automation bias, in which clinicians defer to algorithmic recommendations even when their own judgment diverges; clinician deskilling as complex cognitive tasks are gradually automated; and silent performance drift, in which systems trained on one population fail when applied to another without triggering any visible alarm.

“Legal, regulatory, and governance frameworks are essential,” Dr. Reddy wrote, “but they are inherently reactive; they codify risks that have already been identified and typically lag behind the rapid pace of technological change.”

On bias specifically, the framework offers what Dr. Reddy described as a tiered safeguard model. Upstream, it calls for demographic diversity in training datasets drawn from multiple clinical centers. During development, it calls for algorithmic impact assessments before AI is deployed in vulnerable communities. After deployment, it calls for continuous bias monitoring rather than one-time premarket review. Throughout the process, the framework emphasizes community representation embedded in development rather than added after key design decisions have already been made.

“The most actionable bias-related contributions are mandatory demographic diversity in training datasets, continuous post-deployment bias monitoring, algorithmic impact assessments before deployment in vulnerable communities, and community representation embedded structurally in AI development processes, not consulted after design decisions are fixed,” Dr. Reddy said.

The framework aligns with international governance standards, including the European Union’s AI Act, which classifies health care AI as high risk, and the WHO’s guidance on AI ethics. Dr. Reddy said future priorities should include longitudinal studies of clinician deskilling, standardized post-deployment surveillance methods designed for resource-limited settings, offline-capable AI systems that can function without reliable connectivity, and greater harmonization of AI regulation across jurisdictions.

The central risk, according to the commentary, is also one of the most subtle: AI tools designed to democratize care could instead flow toward populations already advantaged by proximity, infrastructure, and resources, widening the gaps they were built to close.

Dr. Reddy reported no conflicts in the commentary.