University of Alabama professor Rebecca K. Ivic is calling for a fundamental shift in how artificial intelligence is designed and evaluated for health: an AI system should not be considered effective only because it provides accurate information.
Ivic is a commissioner and the University of Alabama’s co-secretariat lead for the Nature Medicine Commission on Quality Health Information for All. UA is one of three co-secretariat institutions supporting the Commission, which brings together international experts to develop approaches for strengthening the quality of health information worldwide.
In a new article published in Nature Human Behaviour, one of the Nature Portfolio’s most selective and influential multidisciplinary journals, Ivic and colleagues introduce a model for ensuring that AI-generated health information does more than provide technically accurate answers. It must also help people understand what information means, judge its significance, make informed decisions, and determine what to do next.
“An AI response can be technically accurate and still fail the person using it,” said Ivic. “Someone can receive medically accurate information and still be left uncertain about how worried they should be, what the information means for them, what uncertainty remains or whether they need to seek human expertise. Accuracy is essential, but it is not sufficient.”
The article introduces a model of health-literate AI organized around four dimensions: comprehension, agency, accountability and proportionality.
Together, the dimensions provide a framework for reconsidering how health-related AI systems are designed, tested, deployed and governed.
The model responds to a rapidly changing information environment in which millions of people increasingly encounter AI-generated information through search engines, conversational systems, health applications and other digital technologies.
Much of the development and evaluation of AI has emphasized measures such as accuracy, predictive performance, efficiency and technical capability. Ivic argues that health-related AI requires another layer of evaluation: whether information is communicated in a way that allows people to understand and act on it appropriately.
“The question cannot only be, ‘Did the AI provide the correct answer?’” Ivic said. “We also have to ask whether it communicated that answer in a way that supports sound judgment and appropriate action. Does the person understand what matters? Is uncertainty communicated appropriately? Does the response convey when something is urgent and when it is not? Does the user know when AI is no longer enough and a health professional is needed?”
That distinction has implications far beyond individual interactions with chatbots.
AI is increasingly being incorporated into health care, consumer technologies, workplaces, government agencies and public information systems. Decisions made now about how these systems communicate health information could shape how people understand symptoms, treatments, risk, prevention and health recommendations at enormous scale.
The framework therefore positions health literacy not simply as a characteristic of an individual user, but as a design and governance responsibility for AI systems and the institutions that develop and deploy them.
“For decades, health literacy research has asked how we create information and systems that people can navigate and use,” Ivic said. “Artificial intelligence changes the scale of that challenge dramatically. We now have systems capable of generating individualized health communication almost instantaneously. We need to make sure that capability is being developed around human understanding and decision-making.”
Ivic’s work with Nature is intended to provide a common conceptual foundation for researchers, technology developers, health organizations, policymakers and regulators. The model can inform the development of new measures and benchmarks for evaluating AI systems alongside traditional measures of technical performance.
It also places health communication at the center of the broader international discussion surrounding responsible AI.
“This is not simply a communication problem after an AI system has already been built,” Ivic said. “Health literacy needs to become part of the architecture of responsible AI- part of how systems are designed, evaluated, governed and regulated from the beginning.”
The publication places the work before audiences spanning the behavioral and social sciences, artificial intelligence, medicine, public health, policy and technology, giving the model relevance well beyond health communication.
For Ivic, the two efforts address different dimensions of the same challenge: ensuring that rapidly changing information systems actually serve the people who depend on them.
“AI is becoming part of the infrastructure through which people encounter health information,” Ivic said. “We have an opportunity now to establish what good looks like before these systems become even more deeply embedded in everyday life. Health-literate AI provides a framework for beginning to do that.”
“Building health-literate artificial intelligence” is authored by Rebecca K. Ivic, Scott C. Ratzan and Ruth Parker and was published in Nature Human Behaviour on Sept. 15.