AI in medicine — liability debate

A medical error used to point at one person in one room. When an artificial intelligence tool is involved, a lawsuit can now point at the doctor who used it, the hospital that bought it, and the company that built it. (stanford.edu) Medical artificial intelligence usually works like a second set of eyes. In radiology, for example, software scans an image for patterns such as a lung nodule or a brain bleed and sends the clinician a suggestion, not a final legal decision. (fda.gov) That distinction matters because United States malpractice law still treats the clinician as the person responsible for the patient-facing call. Stanford’s 2024 policy brief says courts are likely to keep focusing on whether the doctor’s conduct met the standard of care even when software was part of the chain. (stanford.edu) Now the standard-of-care question is splitting in two directions. A doctor can be accused of relying too heavily on a flawed tool, or accused of ignoring a stronger Food and Drug Administration-cleared tool that was available and increasingly common. (medicaleconomics.com) (physicianaihandbook.com) That is why malpractice insurers are starting to pay attention to procurement and workflow, not just bedside mistakes. Medical Economics reported that carriers are considering policy riders and looking at how heavily a practice relies on artificial intelligence tools when they price or structure coverage. (medicaleconomics.com) The vendor side is changing too. If a company markets software as accurate, safe, or clinically useful, plaintiffs can argue product defect, bad warnings, weak testing, or misleading promotion when the tool fails in a real case. (stanford.edu) (censinet.com) The cleanest recent example is not in image reading but in the operating room. Reuters reporting, echoed by multiple outlets, found that after machine-learning software was added in 2021 to Acclarent’s TruDi Navigation System for sinus procedures, the Food and Drug Administration received at least 100 malfunction and adverse-event reports, with at least 10 patients reported injured by November 2025. (news18.com) (futurism.com) Those reported injuries were not minor software bugs on a screen. The cases described skull-base punctures, cerebrospinal-fluid leaks, carotid artery injuries, and strokes during sinus surgery, and at least two patients sued over the device’s role in their procedures. (medboundtimes.com) (futurism.com) The Food and Drug Administration’s own public list shows how fast this category has grown. The agency says its Artificial Intelligence-Enabled Medical Device List is meant to identify devices authorized for marketing in the United States, which gives hospitals and lawyers a concrete record of what was cleared, when, and for what use. (fda.gov) Clearance does not end the argument. A Food and Drug Administration-cleared tool can still be used on the wrong patient, in the wrong workflow, with the wrong training, or with performance that drifts outside the setting where it was first validated. (stanford.edu) (censinet.com) That is why professional groups are now treating artificial intelligence liability as a live operational problem instead of a future thought experiment. The Society for Imaging Informatics in Medicine has scheduled an April 27, 2026 webinar called “AI on Trial: The Legal Landscape of Human-AI Interaction in Radiology,” with sessions on malpractice trends and how legal outcomes can change depending on whether artificial intelligence was used or ignored. (siim.org) (my.siim.org) The practical fight is no longer over whether artificial intelligence belongs in medicine. The fight is over which miss becomes negligence: trusting the software too much, buying the wrong software, or failing to use the best software when the chart later shows it was sitting there all along. (medicaleconomics.com) (stanford.edu)