Health apps are now part of everyday life. People use them to track periods, monitor fitness, assess mental health symptoms and interpret changes in their bodies. Many of these tools increasingly rely on artificial intelligence, creating the impression that their recommendations are objective.

Yet researchers argue that bias can enter a digital health product long before an algorithm is trained. The deeper problem may begin with the assumptions developers make about what health looks like, which problems deserve attention and whose experiences should shape the final product.

Bias Can Start With The Data

Digital health systems depend heavily on the information used to build and train them. If those datasets represent only a narrow section of the population, the technology may perform less reliably for everyone else.

One familiar example involves tools designed to recognize skin conditions. Systems trained predominantly on images of lighter skin may be less accurate when evaluating darker skin tones.

The same problem can appear in less obvious ways. A widely discussed health care algorithm once used medical spending as a proxy for how sick patients were. Because some groups historically spent less on health care despite having significant medical needs, the system underestimated their level of illness.

The result was not simply a technical error. A seemingly neutral variable reflected existing inequalities in access, trust and health care use.

The Problem Begins Before Coding

Researchers Mahima Kalla and Noushin Nazarian argue that developers should question their assumptions before creating a digital health tool at all.

They suggest asking whose definition of health is being used, which forms of knowledge are considered legitimate and who has meaningful influence over the design process.

This matters because health is not understood in exactly the same way everywhere. Some communities may place greater emphasis on relationships between physical health, mental well-being, family, culture and environment.

If a technology is built around only one model of health, it may overlook information that is highly relevant to the people it is supposed to support.

Consultation Often Comes Too Late

Another concern is that communities affected by a technology may be consulted only after major decisions have already been made.

By that point, developers may have chosen the problem, the data, the measures of success and even the type of technology being used.

Researchers argue that meaningful participation should begin much earlier. Communities should have the opportunity to question whether an app is needed in the first place and to influence how it collects information and delivers support.

Assumptions can also become barriers. Developers may decide in advance that certain groups have poor digital skills or are unlikely to use new technology, effectively excluding them from design discussions.

Health Technology Reflects Values

Every digital tool is built around priorities.

A developer may value speed, automation and individual independence. Users, however, may care more about family involvement, cultural relevance or stronger communication with health professionals.

Research involving Indigenous communities in several countries has shown that some users were more likely to accept assistive technologies when those tools strengthened relationships with families and health care providers rather than replacing them.

This demonstrates why technically sophisticated design is not necessarily the same as useful design.

Data Ownership Also Matters

Health data are often treated as information that can simply be collected and analyzed. For some communities, however, data may also be connected with identity, collective rights and cultural knowledge.

Researchers therefore emphasize the principle of data sovereignty. Communities should be able to influence how their information is collected, interpreted, linked, stored and shared.

This becomes increasingly important as AI systems require ever-larger amounts of health information.

Fairer Apps Need A Different Start

Reducing bias cannot be achieved simply by improving an algorithm after development is complete.

A more inclusive approach begins much earlier, with questions about knowledge, representation, power and values. Developers need diverse datasets, but they also need diverse voices involved in deciding what the technology should do.

As digital tools become more deeply embedded in health care, these choices will affect more than convenience. They may influence who receives accurate information, whose symptoms are recognized and who benefits from future medical technology.

Building fairer health apps therefore begins not with better code, but with better decisions about the people that code is meant to serve.