Why don't radiologists benefit from AI predictions?

Agarwal, Moehring, Rajpurkar and Salz run an information experiment with professional radiologists and report a split result. Providing AI predictions "does not improve performance on average," whereas providing contextual information, meaning information the radiologists hold that the AI does not, does. The sharper claim concerns the shortfall. Radiologists "do not realize the gains from AI assistance," and the authors attribute this to "errors in belief updating." The excerpt treats the gains as real but unrealized.

The mechanism is an updating account, not a missing-information account. The excerpt names two errors: radiologists "underweight AI predictions," and they "treat their own information and AI predictions as statistically independent." The second error is the one that bears on combination. Two signals that share information cannot be weighed as if they were independent sources, so a clinician who assumes independence will merge them wrongly. The loss, on this account, sits at the join between the human's knowledge and the AI output, not in either source alone. The humans are not described as lacking information; the AI is described as lacking the contextual information the humans already have.

Set against the nearest notes, this excerpt locates a gap they describe at a different point. Does theory of mind predict who thrives in AI collaboration? finds that collaborative performance separates from solo performance and is predicted by perspective-taking. This excerpt points at the step after the user reads the AI output, which is how that output is weighed against the user's own view. The sharpest tension is with Can AI guidance reduce anchoring bias better than AI decisions?. That note prefers guidance to deferral. The excerpt's design conclusion cuts the other way: "the optimal human-AI collaboration design delegates cases either to humans or to AI, but rarely to AI assisted humans," unless the belief-updating mistakes can be corrected. Guidance only informs a decision if the human updates on it, and this excerpt says radiologists do not do so reliably. Can self-ratings replace objective performance scores for AI competence? is a parallel case in which what users believe about their AI-assisted performance does not track measured outcomes, though that note measures self-report, not belief updating.

The excerpt does not establish how large the gap is. It is an abstract, with no effect sizes, sample size, radiology task, case mix, or description of how belief updating was tested, so the direction of the independence error is not shown. It also does not show that AI assistance makes radiologists worse; it shows that AI predictions do not improve average performance in this experiment and attributes the shortfall to updating. The implication, at the strength the abstract supports, is narrow. Before assuming an AI output adds value in a human review workflow, one should check whether the human updates on it, and the paper's own conclusion conditions AI-assisted rollout on correcting those errors. Whether that correction is feasible is the open question the excerpt leaves.