Radiology was supposed to be the first medical specialty to fall to artificial intelligence. Geoffrey Hinton, the computer scientist often called the godfather of deep learning, predicted in 2016 that training radiologists was essentially pointless because machines would soon outperform them. A decade later, radiologists remain employed, busier than ever, and increasingly reliant on AI tools that have quietly transformed how they work without replacing them.
The reality that emerged is more interesting than the extinction narrative. AI has become a second set of eyes—tireless, consistent, and utterly incapable of understanding what it sees.
The shift in the reading room
Modern radiology AI typically operates as a triage system. When a chest X-ray arrives, an algorithm scans it in seconds, flagging potential pneumothorax, nodules, or cardiomegaly. Urgent findings jump the queue; routine images wait their turn. The radiologist still reads every scan, but the order has changed, and so has the cognitive experience of the work.
This sounds like pure efficiency gain, and often it is. But radiologists describe a subtler shift in how they approach their craft. When software pre-highlights a suspicious region, the human reader's attention is primed. Studies suggest this can accelerate detection of genuine pathology while also creating anchoring bias—the tendency to fixate on what the machine noticed and miss what it did not. The algorithm becomes a collaborator whose suggestions are difficult to ignore.
What machines actually see
Deep learning models trained on medical images do not understand anatomy. They recognize statistical patterns in pixel arrangements that correlate with diagnostic labels. A model might correctly identify a tumor not because it grasps what a tumor is, but because it learned that certain texture gradients predict the label "malignant" in training data.
This distinction matters because it explains both the technology's power and its brittleness. AI excels at standardized tasks with abundant labeled examples—mammography screening, diabetic retinopathy detection, bone age assessment. It struggles when images deviate from training distributions: unusual patient positioning, rare pathologies, artifacts from unfamiliar scanner models. The radiologist's job increasingly involves knowing when to trust the algorithm and when to override it, a skill no one was trained for in medical school.
The liability question no one has answered
If an AI flags a lung nodule and the radiologist dismisses it as artifact, who bears responsibility when that nodule turns out to be cancer? If the AI misses a finding entirely and the radiologist, lulled by the software's silence, overlooks it too, does the algorithm's manufacturer share liability?
These questions remain largely untested in court. Regulatory frameworks treat AI as a tool, placing responsibility squarely on the physician. But the psychological reality of human-AI collaboration complicates this legal fiction. Radiologists report that after months of working with a reliable AI system, they begin to trust its negative findings almost unconsciously. The tool shapes the reader's attention in ways that are difficult to document and harder to litigate.
Our take
Radiology's AI experiment offers a preview of how intelligent automation will infiltrate other knowledge professions. The pattern is not replacement but entanglement—humans and algorithms developing a mutual dependency that neither fully controls. Radiologists have not been rendered obsolete; they have been given a collaborator that works at inhuman speed, learns from inhuman quantities of data, and possesses no understanding of the stakes. The profession's future depends less on whether AI can read scans than on whether humans can maintain the judgment to know when the machine is wrong. That is a harder problem than anyone predicted in 2016.
