This analysis synthesizes 8 sources published the week ending Nov 3, 2026. Editorial analysis by the PhysEmp Editorial Team.
The American Medical Association has proposed a rule of thumb for clinical AI harm: liability should follow control. The physician answers for judgment they could meaningfully exercise. The health system, hospital or practice that bought and switched on the tool answers for how it was chosen, validated and monitored. The problem is that physicians, by their own account, have almost no say in that choosing. That gap, between who the AMA says should carry the risk and who actually picks the software, is where the next round of employment negotiations in AI in Physician Employment & Clinical Practice is headed.
The framework released Oct. 29 assigns duties to four parties. Developers must show their products are safe and effective and keep them that way. Deployers are accountable for selection, implementation, governance and monitoring. Physicians are responsible for the clinical judgment within their reach. Patients get a path to a remedy. The accompanying issue brief adds two lines that read like contract language: physicians should not be liable for AI errors they cannot reasonably detect, and they should be able to decline unproven AI without penalty.
A rule written for errors nobody sees
The AMA’s own example is a triage model that quietly deprioritizes a case while the physician sees only a reordered list. Jennifer Goldsack, who heads the AMA’s digital health and AI center, gave Healthcare Dive a second one: a decision-support tool that omits critical data without the physician knowing, and a treatment plan built on top of it. Physicians, she said, are being asked to serve as the safety net for AI failures they may not be able to identify. She also said disclaimers are not an adequate safeguard.
That category of invisible error is not hypothetical. An emergency physician writing on KevinMD walked through an Emory study of two FDA-cleared pulmonary embolism algorithms across 17 facilities and more than 30,000 CT angiograms. The PE model caught 86.8% of emboli, against a clearance figure of 94.86%. It caught 72.9% of subsegmental clots. In patients who later had dedicated PE imaging, the tool missed 19 of 61 clots on the first scan. The researchers also could not confirm that the deployed model matched the cleared version.
A flag moves a study up the worklist. No flag lets it sit. When a radiologist at 1 a.m. reads in the order the software sets, the miss happens in the queue, not in the read, and the AMA’s principle says the physician should not own it.
The Emory figures turn an abstract principle into a measurable one. If a deployer can be shown to have run a tool with a real-world sensitivity well below its clearance number, and never checked, the AMA’s framework points the liability at the institution rather than the reader.
Physicians carry the risk; administrators pick the tools
Here is the divergence. In a national survey of 1,000 physicians cited by Northeastern’s Timothy Hoff in Medical Economics, just under three-quarters said they have little influence over which AI tools their organizations implement, and two-thirds want more. Separately, a 2026 AMA survey found a quarter of physicians had received no AI training, and 92% wanted more, according to Healthcare Dive.
Current legal thinking has not caught up to the AMA’s allocation. A CNS Quarterly review summarized by Becker’s Spine Review notes that legal scholarship still places responsibility for medical decisions on the treating physician. The same review flags automation bias running the other way: physicians may feel pressure to justify disagreeing with a model even when it is poorly validated or used outside its intended context. So the physician is exposed for following the tool and, socially at least, for overriding it.
The AMA framework is advocacy, not statute. Until courts or legislatures adopt something like it, the only place a physician can make “liability follows control” binding is in an employment agreement.
The stakes per claim keep climbing
The timing matters because malpractice is getting rarer and costlier at once. Becker’s Physician Leadership, drawing on AMA and industry data, reports that 1.8% of physicians were sued in 2024, down from 2.3% in 2016, yet premiums rose for the seventh straight year in 2025. Payments of $500,000 or more made up a record 36.5% of total malpractice payments in 2024, and the industry combined ratio rose to 107.4% in 2025. Writing for KevinMD, MagMutual’s executive chair William Kanich, MD, JD, described the same pattern: claim frequency down for years, severity up.
Fewer, larger claims mean a single AI-linked case can define a physician’s insurance history. A new category of claim, with no settled rule on who pays, arriving in a market where insurers are already losing money on the line, is the kind of risk that shows up in premiums and tail coverage before it shows up in case law.
Recruiters should expect the question to shift from “do you use ambient AI?” to “who validated it, who monitors it, and does your malpractice policy name it?” Organizations without a written answer will look riskier to experienced candidates than their pay grid suggests.
What deployer accountability looks like on paper
Some of the deployer duties are already taking concrete form. A BakerHostetler analysis on JD Supra argues that state recording statutes may require notice or consent for each encounter with ambient AI, not a one-time disclosure at intake. It recommends that organizations inventory every AI tool in use, including unapproved “shadow AI,” and require licensed clinicians to review and approve AI-generated notes before they enter the record. Each of those steps is a deployer obligation in the AMA’s terms, and each one also generates physician work: the review, the per-visit disclosure, the documentation that consent happened.
That is the employment consequence hiding in the framework. If deployers are accountable for governance, governance becomes a line in the physician agreement: a right to decline tools without local validation, indemnification when a vendor defect causes harm, an AMA-style bar on vendor contracts shifting developer defects onto physicians, and protected time for the review work these tools create. Hoff’s survey suggests physicians want that seat at the table. The AMA has now handed them a vocabulary for asking.
The AMA calls the brief a first step and lists liability for increasingly autonomous AI and automation bias among its next priorities. The number to watch is not a verdict. It is how many employed-physician contracts signed in 2027 include an AI clause at all, and whether the institutions writing them accept the deployer role the AMA has drawn for them or try to write it back onto the physician.
Sources
AMA lays out who should be liable when clinical AI harms patients – Becker’s Hospital Review
AI vendors, health systems join physicians in liability for clinical AI – Healthcare Dive
AI in radiology: A missing flag is not a negative read – KevinMD
Physicians and health systems must align on AI to avoid repeating EHR mistakes – Medical Economics
AI in neurosurgery raises privacy, liability concerns – Becker’s Spine Review
Why are malpractice premiums rising as lawsuits fall? – Becker’s Physician Leadership
Fewer malpractice claims, bigger consequences: what physicians should know – KevinMD
DSIR 2026 Deeper Dive: Ambient AI in the Exam Room – JD Supra (BakerHostetler)