An AI tool helped 21 physicians diagnose a rare kidney cancer more accurately, but doctors who were more receptive to its advice also accepted incorrect recommendations more often. The study, reported by Korea Biomedical Review, puts a useful point beyond raw accuracy: working well with AI may be a skill of its own.
Watch Desk analysis
What happened
Researchers developed a deep-learning tool to identify fumarate hydratase-deficient renal cell carcinoma, a rare and difficult-to-spot kidney cancer. In tests involving 21 physicians and 30 challenging cases, diagnostic accuracy rose from 60 per cent without AI assistance to 73.3 per cent with it. Precision also improved, and pathologists agreed more often, without a significant increase in reading time.
The results were not all in the AI’s favour. Physicians classed as more receptive to its recommendations accepted incorrect ones 80 per cent of the time, compared with 60 per cent among a more cautious group. The researchers said these differences were not associated with clinical experience or subspecialty. Korea Biomedical Review reports that the study was published in npj Digital Medicine.
Why it matters
A tool can improve average diagnostic performance and still create a sharp trap for someone who follows it too readily. That matters especially here: the cancer is rare, can resemble other kidney cancers, and may be missed if clinicians do not suspect it and order further tests.
The study suggests that deploying an AI tool is not the whole job. Clinicians may also need training that helps them judge when to accept its recommendation and when to challenge it. The results come from a small study, not evidence that AI-assisted diagnosis will improve outcomes for every patient or clinical team.
Our read
This is a promising result with its most useful warning attached. The tool helped in difficult cases, but “use the AI” is not a complete clinical strategy. Better training in how to work with recommendations deserves attention alongside the model’s diagnostic performance.
What to watch
- Whether larger studies reproduce the accuracy gains.
- Whether tailored training helps clinicians spot incorrect recommendations.
- Whether the tool can help identify patients who need confirmatory testing.
Discussion spark: Should hospitals prioritise improving AI models, or training clinicians to recognise when an AI recommendation is wrong?
Sources and evidence
- Source update (13 November 2026, 04:57 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.