Artificial Intelligence & Machine Learning: AI in Healthcare
The Algorithm That Sees Cancer Before the Doctor Does
In several clinical trials, an AI trained on medical images has outperformed specialist radiologists at detecting early-stage cancer — and nobody is entirely sure how it does it.
The Idea
Machine learning models used in medical imaging don't work the way most people imagine. They aren't given a checklist of what cancer looks like. Instead, they are shown hundreds of thousands of labelled scans — this one malignant, this one benign — and they learn to extract patterns that no human explicitly defined. The result is a model that can flag anomalies with remarkable accuracy, sometimes catching tumours at stages where they are still highly treatable. But here's the genuinely unsettling part: when researchers try to interrogate what the model is actually responding to, the answer is often murky. It may be picking up on subtle textural gradients or spatial relationships between tissues that radiologists have never been trained to consciously notice — because until now, no one knew those features mattered. This is the defining tension of AI in healthcare. The performance gains are real and measurable. But medicine has always demanded explainability — not just 'the test says yes', but 'here is why, here is the mechanism, here is what we do next'. An opaque model that is right 95% of the time is a genuinely novel kind of tool, one that our existing frameworks for clinical decision-making weren't designed to handle. The question isn't whether AI is useful in medicine. It clearly is. The harder question is how much we trust a system that cannot show its working.
In the World
In 2020, Google Health published results from a deep learning model trained on mammography scans from tens of thousands of patients across the US and UK. The model was tested against a panel of experienced radiologists reading the same images. It reduced false negatives — missed cancers — by 9.4% in the US dataset, and reduced false positives by 5.7%. In other words, it was more accurate in both directions: catching more real cancers and raising fewer unnecessary alarms. The study was published in Nature and attracted enormous attention. But it also drew pointed scrutiny. Critics noted that the model was trained and tested under controlled conditions that don't fully reflect the chaotic reality of clinical practice — variable scan quality, incomplete patient histories, the judgment calls that happen outside the image itself. There was also a structural issue: the AI was evaluated against individual radiologists, whereas standard clinical practice in many countries involves two radiologists reviewing each scan independently. When the model was compared against that two-reader standard, the gap narrowed. None of this negates the finding. But it illustrates something important: deploying AI in healthcare isn't a software update. It is a reorganisation of an entire system of human expertise, institutional trust, and liability — and the technology is moving faster than any of those structures are designed to adapt.
Why It Matters
It is easy to receive news like this in one of two modes: uncritical optimism ('AI will cure cancer') or reflexive suspicion ('they're replacing doctors with algorithms'). Neither is useful. What's worth carrying forward is a more precise kind of scepticism — one that asks not just 'does it work?' but 'work in what context, validated how, governed by whom, and with what recourse when it's wrong?'. If you ever find yourself in a healthcare system that uses AI-assisted diagnostics — and statistically, you will — you have a right to understand what role that system played in your care. More broadly, the pattern here repeats across almost every domain where AI is being deployed: impressive benchmark performance, genuine real-world utility, and a troubling gap between what the model can do and what we can explain about why it does it. Learning to sit with that gap, rather than collapsing it prematurely into either fear or faith, is one of the more important cognitive skills of this decade.
A Question to Ponder
If an AI is consistently more accurate than a human expert but cannot explain its reasoning, at what point does demanding an explanation become an obstacle to better outcomes — and who gets to decide that threshold?
Get a new one of these every morning.
Start learning with Thinkable