For a recent university paper I dug into one of the most hyped frontiers in medicine: using artificial intelligence to make diagnoses. The promise is easy to state — faster, more consistent reads of scans, lab values, and patient data, with scarce expertise spread further. The reality, it turns out, is more interesting than the headlines.
Where AI already shines
The strongest results come from narrow, well-defined imaging tasks. As far back as 2017, a neural network trained on roughly 130,000 clinical images classified skin lesions at a level comparable to board-certified dermatologists. Today you find similar systems across radiology (spotting lung nodules, measuring volumes), pathology (sorting tissue samples), and dermatology. Where there's a mountain of labelled images and a clear question, the models do well.
There's also a quieter benefit that gets less attention than raw accuracy: efficiency. A lot of the real value isn't replacing the doctor — it's automating the repetitive scaffolding around the diagnosis, like pre-sorting, prioritising, and measuring, so clinicians have more time for the hard cases.
Why it's not that simple
The hype tends to skip the asterisks, and there are a few big ones:
- The headline numbers don't generalise. A 2025 meta-analysis of 80-plus studies found that generative AI models, on average, weren't clearly distinguishable from physicians — but they performed worse than genuine specialists. Peak scores from ideal conditions rarely survive contact with the messy clinical everyday.
- Bias is baked in. A model mirrors its training data. Skin-cancer models trained mostly on lighter skin perform worse on darker skin tones — and the failure often goes unnoticed when the test set isn't diverse enough to reveal it.
- The reasoning is opaque. Many high-performing models can't really explain why they reached a recommendation. That's a problem when a patient has a legitimate interest in understanding a decision that affects them — and when someone has to be accountable if it's wrong.
The takeaway
The evidence points toward a complementary role rather than a replacement one. AI is most valuable where expertise is scarce, but it still trails experienced specialists. That argues for a hybrid setup: AI as a second reader or triage tool that backs up the clinical decision instead of taking it over — deployed under human oversight, inside transparent processes, and continuously validated.
Less "the algorithm will see you now," more "the algorithm flagged this, and your doctor took a closer look." That's the version worth building toward.