What Should AI Actually Do in Oncology?

There is no shortage of claims about artificial intelligence in oncology. Much of the discussion concerns what these systems can generate. Rather less concerns what oncology actually needs.

Those are different questions, and starting from the second one produces a more useful list.

Where the real bottlenecks are

Volume of evidence. The oncology literature expands faster than any individual can track. Even within a subspecialty, staying current is difficult. The bottleneck is not that the information is unavailable — it is that finding the relevant part at the moment of decision is hard.

Unstructured information. Much of the clinically important record is prose: pathology narrative, clinic letters, operation notes, imaging reports. Structured fields capture a fraction. Anything that needs the whole record needs someone to read it.

Connecting across domains. A molecular finding matters in relation to the disease, the treatment history, the evidence base and the trial landscape. Each lives in a different place. Assembling them is manual, repetitive work.

Trial identification. Registries are searchable but not well aligned to the question "is there a trial that fits this specific patient". Retrieval is easy; relevance assessment is not.

None of these bottlenecks is a shortage of generated text.

What follows from that

Retrieval before generation. The valuable operation is usually finding the right evidence and showing it, not composing a fluent summary. A system that surfaces three relevant papers with their findings is more useful than one that produces a paragraph the clinician cannot check.

Extraction that preserves qualification. Clinical text is full of negation, hedging and temporal structure. "No evidence of metastatic disease", "cannot exclude", "previously treated" — an extraction pipeline that discards these produces confident nonsense. Preserving them is harder and matters more than raw extraction accuracy.

Calibration over accuracy. A model that is right 90% of the time and knows which 10% it is unsure about is more clinically useful than one that is right 93% of the time with uniform confidence. Knowing when to distrust an output is what makes it safe to use.

Showing the reasoning. A recommendation a clinician cannot interrogate is one they must either accept on faith or ignore. Neither is a good outcome. Systems that expose which evidence drove a conclusion can be challenged, and a system that can be challenged can be trusted more, not less.

Evaluation is where most programmes go wrong

The recurring failure is optimising a metric that no clinical decision depends on. Accuracy on a curated benchmark is not the same as usefulness in a workflow. A model can perform well on retrospective data and change nothing prospectively, because the decision it informs was not actually difficult, or because its output arrives after the decision point.

Better evaluation asks: which decision does this touch? Does the output arrive in time to affect it? When the model is wrong, is the error detectable? What is the cost of each error type — and are those costs symmetric? They rarely are.

What should not be automated

Some things in oncology are not information-processing problems. Deciding how much treatment burden is worth accepting is a value judgement belonging to the patient. Communicating a poor prognosis is a human act. Weighing an uncertain benefit against a certain toxicity involves preferences a model does not have access to.

Confusing these with prediction problems produces systems that answer the wrong question well.

A modest position

The most defensible role for machine assistance in oncology is as an instrument that extends what a clinician can attend to: more literature, more of the record, more of the trial landscape, in the time available.

That is a smaller claim than much of the surrounding discourse. It is also achievable, checkable, and would represent a real improvement.