Claims that AI will diagnose better than doctors have been made for a decade. In specific, narrow contexts they have been substantiated. In general they have not, and the pattern is instructive.
Where it works
Image interpretation with a clear ground truth. Detecting diabetic eye disease in retinal photographs. Flagging suspicious regions in mammograms. Identifying skin lesions likely to be malignant.
Each of these is a task where a large labelled dataset exists, the answer is verifiable, and the value is in not missing something. Systems doing this are deployed and improve outcomes, usually by catching cases a tired human would miss rather than by replacing the human.
Where it does not
Diagnosis from history and examination – the general work of medicine. The information is unstructured, the ground truth is often uncertain, and the datasets to train on do not exist in usable form.
Rare conditions, where there is too little data. Patients who differ from the training population, where performance degrades silently.
The deployment problem
A system validated in one hospital frequently performs worse in another, because patients, equipment and practice differ. Validating in each new setting is expensive and often skipped.
The responsibility question
When a system flags something and the clinician disagrees, who is right and who is liable. That has not been settled and it shapes how these tools are used in practice – as a second opinion rather than a decision.
The reasonable expectation
Continued expansion in imaging and other structured tasks, slow progress elsewhere, and a persistent gap between demonstrated performance and real-world deployment.