The chest X-ray arrives at 3:47 a.m., one of several hundred that will pass through the department before sunrise. A decade ago, a bleary-eyed resident would have squinted at it alone, caffeine their only assistant. Today, before any human eyes see the image, an algorithm has already flagged a suspicious nodule in the upper left lobe, drawn a neat bounding box around it, and assigned a probability score. The radiologist still makes the call. But the nature of that call has fundamentally changed.

Radiology was always destined to be AI's beachhead in medicine. The specialty deals in pattern recognition across standardized images—precisely the terrain where machine learning excels. The FDA has now cleared well over five hundred AI-enabled devices for medical imaging, covering everything from mammography to retinal scans to bone-age assessment. Major academic medical centers have integrated these tools into their workflows. The technology is no longer experimental; it is operational.

The productivity bargain

The value proposition seems straightforward: AI handles the triage, flagging urgent findings so radiologists can prioritize their attention. Studies suggest these systems can reduce reading time for certain examinations and catch abnormalities that human eyes, fatigued or rushed, might miss. In a field facing chronic workforce shortages—demand for imaging has grown faster than the supply of trained radiologists for years—this assistance is not merely convenient but necessary.

Yet the productivity gains come with subtle costs. Radiologists report a phenomenon that researchers call "automation bias": the tendency to trust the machine's assessment, particularly when tired or overwhelmed. When the AI says an image is normal, the temptation to move quickly to the next case intensifies. The tool designed to catch what humans miss may, paradoxically, make humans more likely to miss what the tool misses.

The liability vacuum

Here lies the profession's unresolved dilemma. When an algorithm fails to flag a tumor that a radiologist, relying partly on that algorithm, also fails to catch, who bears responsibility? The physician who signed the report? The hospital that purchased the software? The company that trained the model on data that may not have represented all patient populations equally?

Medical malpractice law evolved around human judgment and human error. It has no settled framework for errors that emerge from the collaboration between human and machine. Some legal scholars argue that AI tools should be treated like any other medical device, placing liability on manufacturers. Others contend that the physician remains the learned intermediary, responsible for the final interpretation regardless of what assistance they received. Courts have barely begun to address these questions, and the answers will shape not just radiology but every medical specialty that AI touches.

The training data problem

There is also the matter of whose bodies these algorithms learned from. Many early imaging datasets skewed heavily toward patients from academic medical centers in wealthy countries—populations that do not reflect the diversity of people who will encounter these tools. An algorithm trained predominantly on lighter skin may perform differently on darker skin. A model that learned from one manufacturer's CT scanner may behave unpredictably when fed images from another. Radiologists are increasingly asked to trust systems whose training data they cannot fully audit and whose failure modes they cannot fully anticipate.

Our take

AI will not replace radiologists, but it is already replacing the radiologist's solitude. The profession is becoming a human-machine partnership whether its practitioners chose that arrangement or not. The technology offers genuine benefits—faster reads, fewer missed findings, stretched capacity in an understaffed field. But medicine has always been built on clear lines of accountability, and those lines are now blurred in ways that neither law nor professional norms have caught up with. The algorithm that never sleeps also never testifies, never apologizes to a patient's family, never loses its license. Until we decide who answers for its mistakes, we have not finished the work of integrating it into care.