A new WHO discussion paper on artificial intelligence and health policy has put formal language around a problem that’s usually described in vaguer terms: the evidence used to shape health decisions systematically underrepresents most of the world’s population, and that gap is structural, not incidental.

A familiar problem, now written into policy guidance

The paper names what it calls “epistemic injustice and knowledge exclusion.” AI systems trained predominantly in data-rich settings, it argues, tend to inadequately capture the realities of data-scarce environments, which quietly narrows what counts as legitimate evidence in health policy in the first place. It also flags “power concentration”: a small number of commercial actors dominate AI development, leaving countries outside that circle with less adaptable tools and widening global inequities.

These aren’t abstract concerns. The paper points to real cases where biased algorithms produced concrete harm at scale, from wrongful accusations against families based on flawed risk-scoring to health forecasting tools that underperform the moment a population differs from whoever the model was trained on. In the earliest stage of the policy cycle, simply understanding a health problem, the paper warns that AI can distort priorities by leaning on whatever data is available and quietly overlooking the populations that aren’t well represented in it.

Why the gap compounds, and what closing it looks like

The concern isn’t just that individual AI tools perform worse for underrepresented groups. It’s that the gap compounds through the entire policy cycle: what data exists shapes what gets studied, what gets studied shapes what evidence policymakers can draw on, and what evidence is available shapes which populations end up considered at all. The paper’s own risk mapping traces this thread from problem definition through solution design to monitoring and evaluation, and at every stage the same populations risk being left out again.

Closing a gap like this takes more than better intentions. It takes access to data from the places that are usually missing in the first place. PaiX Navigator was built around exactly that: a Data Request Agent and Data Insights Agent that give researchers and health organizations direct access to data and insights from across 60+ countries, including many of the populations conventional health AI training sets leave out. The goal is to make underrepresented populations visible in the data itself, not just in the conversation about it.

See how PaiX Navigator surfaces insights across 60+ countries

References

• World Health Organization. Artificial intelligence and evidence-informed policy – emerging challenges and opportunities: discussion paper. Geneva: World Health Organization; 2026. Available from: https://doi.org/10.2471/B09667