Introduction: The illusion of progress
Artificial Intelligence (AI) has rapidly become one of the most promising tools in healthcare capable of powering diagnostic imaging, predicting disease outbreaks, supporting clinical decision-making, and enabling mobile health interventions. In countries with overburdened health systems, AI’s potential to do more with less has been particularly exciting. But as researchers and product teams rush to build and deploy these AI tools, we are increasingly facing an uncomfortable truth: the real barriers to adoption aren’t technical but ethical. Over the past few years, I have worked across the intersection of AI, public health, and human-centered research. I have participated in projects evaluating mobile health technologies, explored user behavior in digital app adoption, and collaborated with both developers and frontline health professionals. One pattern repeats itself across almost every initiative: no matter how smart or efficient the technology, if it doesn’t build trust, it doesn’t get used. We tend to obsess over algorithmic precision, model tuning, and pipeline optimization but neglect to ask: Is this system fair? Understandable? Inclusive? Ethical? AI in healthcare won’t fail because of bad code. It will fail because we didn’t design for trust.
Bias is built in
Bias in AI isn’t a bug. It is a reflection of human systems, amplified at scale. Nowhere is this more dangerous than in healthcare, where decisions directly affect diagnosis, treatment, and outcomes. A significant challenge lies in the lack of diverse datasets used to train health AI systems. If a model is trained predominantly on data from Western populations, how will it perform in Nigeria, India, or rural South America? The answer is: often poorly. In one of our recent studies evaluating mobile diabetes apps, we analyzed user adoption behavior and decision-making patterns across various populations. What we observed was telling: apps designed with assumptions about literacy levels, smartphone behavior, and even health beliefs often failed to resonate with real users. The technology wasn’t flawed rather the perspective was. It didn’t reflect the reality of its intended users. When bias is baked into the data, the model can reinforce existing disparities which include misdiagnosing darker skin tones in dermatology, underestimating risks in marginalized populations, or missing patterns entirely due to underrepresentation. In some cases, this leads to direct harm. In others, it simply erodes trust, especially in communities already skeptical of institutional healthcare. The fix isn’t just “more data.” It’s better data, inclusive practices, and deliberate decisions about whose health gets prioritized during model development.
Explainability: The black box problem
Trust in medicine is not optional. When lives are on the line, patients and their doctors deserve to know why a recommendation was made. Unfortunately, many AI systems in healthcare operate as “black boxes,” offering high-confidence predictions without human-understandable reasoning. For a data scientist, this might be a tolerable tradeoff between performance and transparency. For a physician or patient, it’s unacceptable. Imagine being told you are at a high risk for a rare disease based on a model’s prediction but not being told why. Or being denied a medical procedure because an AI system flagged you as low-priority, based on variables you can’t see or understand. In health tech, explainability isn’t just a technical feature, it is a moral obligation. When designing AI systems, we must prioritize not only predictive power but interpretability. Tools like SHAP values, LIME, and transparent decision trees are steps in the right direction. More importantly, we must design interfaces and communication flows that help real people make sense of the output. In public health contexts, especially in low-resource settings, this means accounting for literacy levels, language, and local health knowledge.
A path forward: Building responsible AI in health
So how do we move beyond lip service and start building AI systems that are genuinely responsible?
First, we need diverse teams designing these systems, not just racially and culturally diverse but inclusive of patients, nurses, clinicians, public health experts, and ethicists. If everyone at the table is a machine learning engineer, we’ve already lost perspective. Second, we must build explainability into the model lifecycle. Tools that clarify how decisions are made should be tested as rigorously as accuracy metrics. Third, we should treat ethical reviews with the same seriousness as technical audits. Independent ethics panels, transparency reports, and community advisory boards can help keep models accountable to the people they serve. Finally, in global health contexts, we must co-create AI systems with the communities they are intended to help. This means engaging local researchers, using participatory design methods, and adapting models to reflect local values and constraints.
AI in public health will not succeed on technical merit alone. It will succeed or fail based on whether people believe in it, understand it, and see it as a force for good in their lives. That’s not just a design challenge. It’s a moral one. We must move beyond accuracy as our only metric. We must ask: Is this model fair? Is it respectful? Is it transparent? And if the answer is no, we must have the courage to pause and rethink. Because the future of AI in health isn’t about smarter machines. It’s about dignity by design.
.Onyekwelu is an AI researcher and Digital Health innovator focused on creating ethical, human-centered solutions at the intersection of technology and public health, and passionate about advancing equitable innovation in Global HealthTech.
