The training data problem: bias and AI in genomic medicine
CHELSEA WAGNER, MS CGC | July 14th, 2026
We already know our databases have a representation problem. Non-European ancestry is underrepresented in nearly every major genomic resource we rely on. It's why a variant gets classified as a VUS in one patient and benign in another, and why "we don't have as much data" is a sentence too many of us have had to say out loud.
Now layer AI on top of that. It doesn't fix the problem. It can quietly make it worse, and harder to see.
Here's the mechanism: An AI model learns from the data it's trained on. If that data underrepresents a population, the model underperforms for that population. Same failure we already know from gnomAD and ClinVar, except now it's wrapped in fluent, confident output that gives no signal that it's on shaky ground.
That last part is what makes AI bias more dangerous than database bias, not less. When you query a genomic database, the gap is visible. You see the allele frequency is "not reported" for a given ancestry group. The absence is right there on the screen, and you counsel accordingly.
An AI tool doesn't show you the gap. Ask it about a variant in a population it has thin data on, and it answers in the same confident tone it uses for a well-studied variant in a well-studied population. The uncertainty gets smoothed over. You have to know to look for it, because the tool won't volunteer that it's guessing.
Say you use an AI tool to help draft patient education material about a condition, and you ask it to make the language culturally appropriate for a specific community. It'll produce something confident and clean. But the model's sense of that community came from whatever was in its training data, which skews toward whoever writes the most on the internet in English. The output can carry assumptions you didn't ask for and can't easily see. If you're not from that community, or don't know it well, you might not catch what's off.
This connects to something GCs already hold as a core value. We care about health equity. It's in the professional identity, not bolted on. Which means we're actually well positioned to be good at this, better than most fields, because we already ask "who's missing from this dataset?" as a reflex. The task is turning that same question toward AI: who was underrepresented in what trained this tool, and how would I know?
A few habits that help. Ask what a tool was trained on before you trust it for anything population-specific. Treat confident output about underrepresented groups as a flag to verify, not a reason to relax. And when you use AI to generate patient-facing material, run it past the same equity lens you'd apply to anything else, because the tool didn't.
I'm not saying don't use these tools. I use them. All the time. I'm saying the bias in AI isn't a separate new problem to learn from scratch. It's the representation problem we already know, moving into a system that hides it better. Our advantage is that we already see it coming.
The uncomfortable part, and I'll say it plainly: our field has its own diversity and representation gaps, in the workforce and in the data. Being rigorous about AI bias means holding both things at once, the technology's blind spots and our own. That's not a comfortable place to sit. But sitting with complexity is what we're trained to do.
