A single human genome contains roughly 3 billion base pairs. A clinical geneticist reviewing a patient's genomic data might need to interpret thousands of variants, each with different levels of evidence, different clinical significance, and different implications depending on the patient's ethnicity, family history, and existing conditions. We're asking humans to do work that's outpacing human capacity. And that gap is only growing.
What You Need to Know
- Genomic data is doubling every seven months, far outpacing our ability to manually interpret it
- AI can dramatically accelerate variant classification, but the clinical decision still requires a human who understands the patient
- The biggest risk isn't AI getting it wrong. It's deploying AI without clinician oversight and losing the contextual judgement that makes genomic medicine safe
- Organisations building genomic AI need to design for clinician trust from day one, not bolt it on later
The Scale Problem
When I was leading Edison, we built a precision health platform that included genomic reporting. The experience taught me something fundamental about the future of genetics: we're generating data far faster than we can make sense of it.
40 exabytes
estimated annual genomic data generation by 2025, doubling roughly every seven months
Source: National Human Genome Research Institute, 2022
Clinical genetics used to be a manageable discipline. A geneticist would review a handful of variants in known disease genes, check the literature, make a classification. That workflow worked when we were looking at single genes for rare diseases.
It doesn't work anymore. Whole genome sequencing identifies millions of variants per patient. Multi-gene panels for common conditions like cardiovascular risk or cancer predisposition flag hundreds of variants that need evaluation. And the evidence base for variant interpretation is itself growing exponentially, with thousands of new studies published annually.
A clinical geneticist I worked with put it bluntly: "I can spend four hours interpreting one complex case. I have 40 cases in my queue. The maths doesn't work."
Where AI Fits
This is where AI becomes not just useful but necessary. Machine learning models can process variant data at a scale humans can't. They can cross-reference a variant against thousands of databases, published studies, and population datasets in seconds. They can flag the variants that need human attention and deprioritise the ones that don't.
In practice, this looks like a triage system. AI handles the high-confidence classifications, variants that are clearly benign or clearly pathogenic based on strong evidence. The uncertain cases, variants of unknown significance, novel variants, cases where the evidence is conflicting, go to a human geneticist for review.
This isn't about replacing clinicians. It's about giving them a manageable workload. Instead of reviewing 500 variants manually, a clinician might review 50, the ones where their expertise and judgement actually matter.
The time savings are significant. What took eight hours of manual work for a complex genomic report could potentially be done in under an hour with AI-assisted triage. That's not an efficiency gain. That's the difference between a service that scales and one that doesn't.
Why AI Needs Clinicians
Here's where it gets complicated. And where I've seen health tech companies get it dangerously wrong.
Genomic interpretation isn't pattern matching. It's clinical reasoning. A variant that's benign in one population might be pathogenic in another. A classification that's correct in isolation might be wrong in the context of a specific patient's family history, symptoms, and co-existing conditions.
AI models trained on population-level data can miss these nuances. Worse, they can be confidently wrong, presenting a classification with high certainty that a clinician would immediately question if they saw the full picture.
25%
of variants classified as 'pathogenic' in databases were reclassified after expert review in one study
Source: Manrai et al., New England Journal of Medicine, 2016
The reclassification problem is real. Variants that were once considered disease-causing get reclassified as benign as evidence accumulates. AI models that don't account for this, that treat a database entry as ground truth, will propagate outdated classifications at scale.
This is why the "clinician in the loop" isn't a nice-to-have. It's a safety requirement.
The Trust Architecture
When we built Edison's genomic reporting tools, we learned that the technical accuracy of the AI wasn't the hardest problem. The hardest problem was getting clinicians to trust the output enough to use it, without trusting it so much that they stopped questioning it.
That balance requires specific design choices.
Show the reasoning, not just the answer. When the AI classifies a variant, the clinician needs to see why. Which databases were consulted? What evidence supports the classification? Where is the evidence weak? A black-box classification, even a correct one, doesn't build trust.
Make disagreement easy. Clinicians need to override AI classifications without friction. If the system makes it hard to disagree, clinicians will either stop using it or stop questioning it. Both outcomes are dangerous.
Track and learn from overrides. When a clinician overrides an AI classification, that's training data. It's also a quality signal. If clinicians are frequently overriding a specific type of classification, the model needs retraining. This feedback loop is what makes the system get better over time.
Be explicit about uncertainty. The most dangerous AI output is a confident wrong answer. Systems need to clearly communicate their confidence level. A variant classified with 95% confidence should look different from one classified with 60% confidence. Clinicians can handle uncertainty. What they can't handle is false certainty.
Building for the Right Future
The intersection of genetics and AI is one of the most promising areas in healthcare. I genuinely believe that within a decade, AI-assisted genomic interpretation will be standard clinical practice, reducing diagnostic delays, improving treatment selection, and catching risks years before they become conditions.
But we'll only get there if we build these systems the right way. That means starting with clinician workflows, not AI capabilities. It means designing for trust, transparency, and human oversight. And it means treating AI as a tool that amplifies clinical expertise, not one that replaces it.
The question isn't whether AI can interpret a genome faster than a human. It can. The question is whether we trust a system that doesn't understand the patient sitting across the desk.
Jay Harrison
Health Technology Advisory
The organisations getting this right are the ones that hire clinicians and engineers in equal measure, that test with real clinical workflows, and that measure success in patient outcomes rather than processing speed.
That's the standard we should hold every genomic AI product to.
- Can AI fully replace clinical geneticists?
- No. AI can handle high-confidence variant classification at scale, but clinical genetics requires contextual reasoning about the individual patient, their family history, ethnicity, and symptoms. AI is a triage and acceleration tool, not a replacement for clinical judgement.
- What's the biggest risk of AI in genomics?
- Overconfident misclassification at scale. An AI model that confidently classifies a variant as benign when it's actually pathogenic could affect thousands of patients before the error is caught. That's why clinician oversight and transparent reasoning are essential.
- How should organisations evaluate genomic AI platforms?
- Ask three questions: Does the system show its reasoning? Can clinicians easily override classifications? Does it track and learn from those overrides? If the answer to any of these is no, the system isn't ready for clinical use.