Building health-literate AI: a blueprint for machines that empower patients

Artificial intelligence is rapidly becoming an intermediary between patients and health information, from chatbots that answer questions about symptoms to algorithms that summarize clinical records and recommend screening decisions. Yet a growing body of research warns that these systems may be quietly shifting an enormous interpretive burden onto the very people they are meant to help. A new Perspective published in Nature Human Behaviour by Rebecca K. Ivic of the University of Alabama, Scott C. Ratzan of the CUNY Graduate School of Public Health and Health Policy, and Ruth M. Parker of Emory University School of Medicine argues that the field must go beyond building accurate or explainable machines. The authors introduce the concept of health-literate artificial intelligence: systems deliberately designed so that information, guidance and responsibility align with users’ abilities, contexts and needs, rather than assuming users can absorb whatever the machine produces.

The argument rests on decades of health literacy research. Since the Institute of Medicine’s landmark report Health Literacy: A Prescription to End Confusion, and through foundational work by Nutbeam, Parker, Ratzan and colleagues, scientists have documented that limited health literacy is widespread and is associated with poorer health outcomes, higher rates of hospitalization, less use of preventive services and greater difficulty navigating care. Classic instruments such as the Test of Functional Health Literacy in Adults and more recent multidimensional tools like the Health Literacy Questionnaire and the European Health Literacy Survey have shown that literacy is not merely a property of individuals. It emerges from the interaction between people’s skills and the complexity of the demands placed on them by health systems, communication materials and technologies. When those demands exceed capacity, understanding breaks down, no matter how intelligent the user may be.

Against that backdrop, Ivic and colleagues ask a deceptively simple question: can AI-mediated systems support health communication and health literacy instead of eroding them? Their answer is conditional. Generative models can translate technical language, personalize content and answer questions at any hour, which gives them genuine potential to lower communication barriers. But the same fluency can mask uncertainty, produce confident-sounding errors and create the illusion of comprehension where none exists. The authors therefore define health-literate AI through four interrelated components: comprehension, agency, accountability and proportionality. These components, they argue, help distinguish systems that are merely accurate or technically explainable from systems designed to help people understand what matters, what uncertainties remain and what to do next.

Comprehension is the first pillar. Drawing on research showing that plain language interventions, teach-back techniques and carefully framed risk presentations measurably improve understanding, the authors contend that AI systems should be engineered to match the health literacy demands of their output to the abilities and circumstances of their users. This is a technical challenge as much as a linguistic one. It means evaluating not just whether a model’s text is grammatically simple, but whether a diverse population of users can actually grasp the meaning, weigh the numbers and act on the advice. A chatbot that produces polished prose at a sixth-grade reading level but fails to convey the difference between absolute and relative risk has not met the comprehension standard, because the cognitive demand of the decision remains out of reach.

Agency is the second component, and it reflects one of the deepest traditions in health literacy theory: the idea, articulated by Nutbeam and others, that health communication should build people’s capacity for informed action rather than simply transmitting instructions. Health-literate AI, in this framing, should preserve and expand user agency. It should offer meaningful choices about how information is presented, support people in questioning outputs, and avoid paternalistic designs that funnel users toward a single algorithmic answer. Research on how patients develop health literacy as an agentic behavior, including studies of immigrants navigating online information during the COVID-19 pandemic, suggests that people actively construct understanding from the tools available to them. Systems that undermine that construction, by overwhelming users or by hiding the basis of their recommendations, fail a core test of health-literate design.

Accountability is the third pillar, and it may be the most provocative. The authors argue that when AI mediates health communication, responsibility for understanding cannot simply be transferred to the user. Transparency alone is insufficient: studies of algorithmic accountability have shown that disclosing the existence of a model does not guarantee that anyone can meaningfully scrutinize it, and research on explainable AI in medicine has questioned whether current explanation techniques genuinely help patients or even clinicians. Health-literate AI therefore requires institutions, developers and health systems to share responsibility for whether communication succeeds. If an AI-driven interface leads patients to misunderstand their treatment or misjudge a screening decision, that is a system-level failure, not merely an individual one. The Perspective aligns this view with broader work on AI governance in healthcare, which emphasizes shared responsibility, institutional guidelines and regulatory frameworks.

Proportionality, the fourth component, addresses how much information and how much technology is actually appropriate for a given health task. Not every health interaction warrants a sophisticated model, and not every output needs every caveat. Proportionate design calibrates the complexity, framing and intrusiveness of the system to the stakes of the decision at hand. A scheduling assistant and a tool that summarizes cancer treatment options carry very different risks of harm, and the authors suggest that evaluation standards should reflect that gradient. Proportionality also implies restraint in the deployment of conversational agents, where systematic reviews have found evidence of potential benefits but also significant gaps in demonstrated effectiveness, safety and equity across populations.

What emerges from these four components is a framework that reframes the debate about AI in health. Much of the current discourse centers on model accuracy, hallucination rates and explainability metrics. Those matters are important, but the authors argue they are insufficient. A model can be highly accurate by benchmark standards and still leave users confused about what the result means for their lives. It can be formally explainable while the explanation remains opaque to a layperson. Health-literate AI shifts the evaluation question from ‘is the machine right?’ to ‘did the human being come away understanding what matters, what remains uncertain and what to do next?’ That question demands new outcome measures, including assessments of user comprehension, decision quality and downstream health behavior, rather than relying solely on technical performance tests.

The implications extend across the AI ecosystem in health, from clinical decision support and patient portals to public health campaigns and consumer wellness apps. The authors call for health-literate principles to inform design, evaluation, research and governance alike. Designers would build systems starting from users’ real abilities and contexts. Evaluators would test whether communication goals are achieved, borrowing from established health literacy assessment traditions. Researchers would investigate where AI genuinely reduces literacy demands and where it merely relocates them, including studies of AI literacy among health professionals and the public. Governance frameworks, including those being developed by healthcare institutions and policymakers, would treat comprehensibility and shared responsibility as requirements rather than aspirations. The Perspective also connects this agenda to broader efforts, such as the Quality Health Information for All Commission, to reinvent health communication for a digital environment saturated with algorithmically generated content.

The timing of the argument is significant. Surveys of trust, digital health literacy and information quality show that public confidence in health information sources is strained, while the volume of AI-generated content continues to grow. Studies have found that people’s perceptions of medical advice shift depending on whether they believe AI is involved, and that gaps in AI literacy among clinicians and patients alike can compromise learning and safety. Ivic, Ratzan and Parker conclude that innovation and understanding are not opposing forces, provided responsibility is distributed deliberately. An AI ecosystem that supports comprehension, preserves agency, embeds accountability and calibrates proportionately could help close long-standing health literacy gaps rather than widen them. The alternative, the authors imply, is a future in which machines grow ever more eloquent while the humans they serve grow ever more responsible for decoding them, a division of labor that health systems, and the people they exist to serve, can ill afford.

Subject of Research: A framework for designing artificial intelligence systems that support health literacy and health communication

Article Title: Building health-literate artificial intelligence

Article References: Ivic, R. K., Ratzan, S. C., & Parker, R. M. (2026). Building health-literate artificial intelligence. Nature Human Behaviour. https://doi.org/10.1038/s41562-026-02595-1

Image Credits: AI Generated

DOI: 10.1038/s41562-026-02595-1

Keywords: health-literate AI, health literacy, health communication, artificial intelligence, Nature Human Behaviour, comprehension, agency, accountability, proportionality, AI governance, generative AI in healthcare, digital health literacy