Who Is Responsible When Clinical AI Gets It Wrong?
When an AI tool contributes to patient harm, who answers for it? Physicians have been asking, and on October 29 the AMA's Center for Digital Health and AI offered its first detailed answer in a new issue brief, Liability for Clinical AI.
The starting point is simple. Responsibility should follow control. Physicians are responsible for the clinical judgment they can meaningfully exercise. Developers are responsible for what they design and build. Health systems and practices that put AI to work are responsible for where and how it's used. When control is shared, responsibility is shared too, but no one's obligations disappear because someone else also bears some.
Liability is one of the biggest barriers to adoption. In the AMA's 2026 physician survey on augmented intelligence, more than one-third of respondents said clear liability frameworks were the regulatory step most likely to increase their trust in AI.
What it means for physicians
The brief argues that physicians shouldn't be held responsible for errors they can't reasonably detect. Its example is a triage model that quietly deprioritizes a patient. The physician sees only a reordered list and has no way to know the tool caused the delay. Physicians who rely on well-validated tools used as intended should have strong liability protections, and they should be free to decline unproven AI without penalty.
What it means for practices
Practices that bring AI into patient care count as deployers in the AMA's framework. That carries real duties, from choosing and validating a tool to monitoring how it performs and deciding when to pull it. The brief also warns that vendor contracts and disclaimers can shift responsibility onto the people least able to prevent harm. Before signing, read the indemnification and liability terms closely, and make sure your team knows when it can override the tool.
The AMA calls this an initial contribution rather than new formal policy, and it's asking for physician feedback as the work continues.
