The Diagnosticians — Summary

Priya Venkatesh keeps a thermos of chai on the console beside her keyboard. Three years ago, her reading room held twelve radiologists. Today it holds five. Not because seven were let go, but because the work reorganized itself around what remained.

The common narrative about AI and medicine is a replacement story: machines that read faster, more accurately, without fatigue, arriving to displace the people who used to do what machines now do. What the diagnosticians actually describe is something harder to diagram. A profession discovering what it was always for.

Priya used to read two hundred scans a day. She reads forty now. Every image on her screen is a genuine puzzle: the shadow that could be artifact or tumor, the asymmetry that might indicate early pathology or normal variation. She is, by every measure, a more skilled diagnostician than she was before the AI arrived. She is also more tired at the end of the shift. Easy cases provided rhythm. Difficult cases provide meaning, but not rest.

Adaeze Okafor is a pathologist in Lagos. Her AI system analyzes tissue samples with pattern recognition that matches or exceeds human accuracy on well-characterized pathologies. She reviews the reports, confirms most, adjusts perhaps fifteen percent, overrides perhaps two percent entirely — because something about the tissue, the patient’s history, the clinical context tells her the pattern-match has missed the point. That two percent has become the center of gravity of her profession.

A pathology report is not usually a recommendation. It is closer to a verdict. And when the system says “high probability of malignant neoplasm, confidence 94%,” someone must decide whether that is cancer. Someone must be accountable for that decision. That someone cannot be the AI. Not because the AI is wrong more often. It may be wrong less often. But accountability requires a moral agent — a person who can be questioned by the oncologist, challenged by the patient, overruled by a second opinion, who bears the weight of the call. The pathologist’s signature on the report is not a formality. It is the moment where computational probability becomes medical authority.

What this reveals is that the analysis was always separable from the accountability. We bundled them because we had no choice. Now we have a choice, and we are only beginning to understand what that means.

Less visible but equally important: Adaeze is reviewing AI output whose design choices she did not make and cannot fully see. Whether the system explains its confidence, names its competing interpretations, flags when tissue sits near the edge of its training data — these choices were made by developers she has never met. If the liability for a wrong finding lands on Adaeze, and the choices that shaped what the AI showed her were made by people she cannot hold to account, something about that distribution is worth examining.

Dr. Chen, Margaret’s endocrinologist in Ohio, offers the third variation. The AI tracks Margaret’s glucose patterns in real time — the dawn effect, the Thursday bridge dinners, the way her levels destabilize when she is anxious. When Margaret walks in, Dr. Chen already knows the last six months completely. The old visit was diagnostic: what has happened since we last met? The new visit is something different. “What do these patterns mean for how you want to live?” is roughly how Dr. Chen described it. The AI freed her from the computational labor of chronic disease management. What it freed her into was the human work of knowing a patient.

The replacement narrative frames AI as a threat: fewer jobs, displaced specialists. But there are not enough diagnosticians on Earth. Sub-Saharan Africa has approximately one pathologist per million people. India has roughly half the radiologists it needs. The profession does not shrink. The definition of who it serves expands. And Priya now reviews flagged cases from rural clinics that have never had a radiologist.

The equity risk is real: AI-assisted diagnosis is better than nothing but not the same as having a specialist who knows the local context, the equipment’s quirks, the clinical presentation patterns of a specific population. The technology creates a possibility. Whether that possibility becomes equity or a new layer of extraction depends on choices being made right now.

The deepest question is about training. Priya learned to read scans by reading a great many of them — not the interesting ones, the ordinary ones. The volume of routine work built the pattern recognition that now allows her to see significance in ambiguity. The AI has removed the easy cases from the queue. Trainees coming into radiology today are working with cases already filtered for difficulty. Whether auditing without the foundation of writing, reading, building, produces reliable judgment is the question the profession cannot yet answer, because the experiment has only just begun.

AI unbundled two things that were always bundled in diagnostic medicine: pattern reading and judgment. It took the reading and left the judging. In doing so, it revealed that the judging was always the harder, rarer, more valuable part. The part that required not just training but experience. Not just knowledge but something that accumulates over years and cannot yet be transferred by any means we have.

The diagnosticians are not disappearing. What is emerging is genuinely different from what existed before, and different things carry different losses even when the net is better. The accountability for the shape of that difference does not belong entirely to the profession adapting to it. Some of it belongs to the developers who decided how much of the AI’s reasoning to make visible. Some of it belongs to the institutions that deployed these tools without building the new developmental infrastructure to replace what the tools dissolved.

These are not rhetorical observations. They are questions about where responsibility lives when a technology reshapes a profession’s capacity to do its work and train its successors.