The adoption of artificial intelligence (AI) across healthcare is accelerating rapidly. Predictive models flag patients at risk of deterioration 1 ; ambient AI scribes draft clinical notes 2 ; computer vision models triage scans 3 .
AI’s potential to improve the efficiency of healthcare and save lives is real. This has driven major investments, bold vendor promises of efficiency gains, and hopes for solutions to workforce shortages and variability in care.
Beneath this momentum, however, lingers a simple, uncomfortable question: is AI actually improving care outcomes? In many cases, we do not know.
The problem is not a lack of models, but a lack of evaluation focused on AI attribution that is aligned with clinical impact. Two main issues come to the fore when evaluating AI in clinical settings: first, an emphasis on accuracy and other metrics of model performance that may not translate into meaningful clinical improvement 2 , and second, a focus on clinical outcome metrics that may reflect heightened clinician vigilance and a change in workflows rather than directly implicating AI in improved performance 4 .
A.G.
Nature Medicine published a clinical update in Research Highlights on 21 Apr 2026.
The item focuses on Is AI actually improving healthcare?.
Review the original article for the full source wording and details.