Whose Data, Whose Story? Reclaiming Nuance in AI-Driven Personalised Medicine

Personalised medicine is often framed as the future of healthcare: precise, data-driven and tailored to the individual. But as AI becomes increasingly central to how genomic information is analysed and explained, a more uncomfortable question begins to surface:

Whose version of “personal” is actually being represented?

In genomic research, AI-supported tools are already indispensable, helping to prioritise variants, synthesise literature and manage the sheer scale of biological data. Within expert environments, these tools are powerful precisely because they operate within a scaffold of training, context and critical interpretation. But that scaffold does not always travel with the output.

As AI moves beyond specialist use and into public-facing spaces where patients, families and individuals turn to it for answers it begins to take on a different role. It is no longer just supporting expertise. It becomes a proxy for it. This shift brings a set of tensions that are worth sitting with.

AI does not just analyse data it can shape meaning
In expert hands, genomic interpretation is rarely definitive. It involves weighing uncertainty, recognising gaps in evidence and understanding how context (take ancestry as an example) affects interpretation. But, AI-generated explanations, particularly those designed for general audiences, tend to smooth over this context. They produce outputs that are coherent, confident, and accessible yes, but often without signalling where knowledge is incomplete or uneven. The result is not necessarily incorrect information, but something more subtle: a shift in meaning. Probability can be interpreted as certainty. Population-level trends can feel like individual diagnoses. Speculation can sound like explanation. The risk, then, is not just technical inaccuracy, it is the quiet transformation of complex science into simplified narratives that carry unintended weight.
Data imbalance is not just a scientific issue it is a narrative one
Nowhere is this more evident than in African genomics. Despite representing the greatest human genetic diversity globally, African populations remain underrepresented in many genomic reference datasets and has well-documented implications for variant classification and disease association studies. In research settings, this limitation is visible and actively managed. Scientists account for it, qualify their interpretations and work within its constraints. But when AI systems (trained on these same imbalanced datasets) generate explanations for broader audiences, that limitation often disappears from view. What remains are statements that could sound complete: • That ancestry explains disease risk • That a single variant provides a clear answer • That population patterns apply neatly to individuals These are not inherently false claims… But without context, these claims can become misleading. They risk reinforcing reductive ideas about biology, identity and difference, particularly in settings where historical and structural inequalities already shape health outcomes.
The real fault line is not accuracy it is interpretation
Much of the current discourse around AI in healthcare focuses on improving models: better data, better algorithms, better predictions. But this focus risks overlooking a critical gap: what happens after the output is generated. AI does not understand how people interpret information when it becomes personal. It does not anticipate how genetic explanations intersect with identity, history, or lived experience. It cannot recognise when its outputs may be taken up in ways that extend beyond their scientific intent. This is where meaning slips. Not because the system fails computationally, but because it operates without awareness of the social and interpretive contexts into which its outputs are released.
Science communication is not an add-on
If AI excels at scale and synthesis, then science communication provides something equally essential: judgement, framing and responsibility. This is not about simplifying science, but about restoring context and reintroducing uncertainty where it has been smoothed away, naming the limits of datasets and distinguishing between what is known, suspected and unknown. In African genomics, this work carries additional weight. Genetic data from African populations has historically been extracted, analysed elsewhere and returned as conclusions rather than conversations. Without careful communication, AI risks accelerating this pattern; producing explanations that travel widely but remain disconnected from the contexts they describe. Science communication offers a way to interrupt that cycle. It shapes how people approach AI-generated information in the first place. It informs the language and narratives that ultimately feed back into AI systems themselves. And it helps ensure that interpretation does not drift too far from the realities it is meant to represent.

Opening the conversation

These tensions don’t resolve neatly, and perhaps they shouldn’t. Instead, they point us toward a set of questions that feel increasingly urgent as AI becomes embedded in how genomic knowledge is shared and understood. What does it mean to offer “answers” in contexts where the data itself is incomplete or uneven? How might AI systems be designed not just to inform, but to signal their own limits? And perhaps most importantly, who is involved in shaping the narratives that emerge from genomic data and who is left out? These questions move us beyond technical performance alone. They ask us to think more carefully about interpretation, responsibility and the kinds of knowledge we are putting into circulation. AI is already embedded in the future of personalised medicine across laboratories, clinics and increasingly, everyday life. As such, the question is no longer whether it belongs there. The question is whether the stories it tells will reflect the full complexity of the people behind the data, or whether, in the absence of context, they will flatten that complexity into something easier to compute, but harder to trust.