AI and Liability in Radiology Workflows

My feed has been full of this story about the head of NYC Health + Hospitals saying he’s ready to replace radiologists with AI for first reads. This came weeks after Dario Amodei, CEO of Anthropic, said on a podcast that AI has already taken over the core technical work in radiology.

Lots of strong takes on both sides but here’s a study that speaks directly to what they’re proposing: Bernstein et al. (2026, Nature Health) ran 282 mock jurors through a radiology malpractice case where a radiologist missed a brain bleed that AI caught. Three medicolegal vignettes. No AI involved: 56.3% of jurors sided with the plaintiff. Radiologist reviewed alongside AI in one pass: 74.7%. Radiologist formed their own read first then cross-checked against AI: 52.9%.

The patient outcome was the same in every vignette but what changed was how AI fit into the workflow. That alone swung the jury over 20 points. The single-pass vignette, essentially what Amodei described and what those hospital CEOs are pushing for, performed the worst. The vignette where the physician read independently first and then checked AI actually dropped below the no-AI baseline.

Everyone’s arguing about whether AI should replace radiologists. Bernstein’s data suggests the more pressing question is how you design the workflow around it. Thats what moved the jury, and it’s not part of the conversation yet.

(Image generated with AI assistance)