What the AMA's brief actually argues
On October 29, 2026, the AMA's Center for Digital Health and AI published an issue brief titled "Liability for Clinical AI," MedTech Dive reported. Its premise: today's liability rules were built for software that behaves predictably and for physicians who have full visibility into what they're relying on. Clinical AI often fails both assumptions, so the AMA proposes dividing accountability among the three parties who actually have some control over an outcome — the physician, the developer, and the health system or practice that deployed the tool.
Jennifer Goldsack, who leads the AMA's Center for Digital Health and AI, framed the problem bluntly: per MedTech Dive, physicians are increasingly being asked to act as a safety net against clinical AI failures — tools whose failures physicians, in Goldsack's words, "in some cases, may not even be able to identify." The goal of the brief, she added, is to make sure "every different stakeholder gets very clear on their responsibility to proactively implement best practices."
The three-way split, plainly
The brief organizes responsibility around what it calls a control principle: liability should track whoever can actually foresee and prevent a given harm, not default to whoever is last in the chain. In practice, that breaks down into three lanes.
| Party | Responsible for |
|---|---|
| Physician | The clinical judgment they can meaningfully exercise — questioning or overriding an AI output when it looks wrong. Not liable for errors outside their control or visibility. |
| AI developer | Designing a tool that is safe and that lets users actually identify errors, and disclosing intended use, performance, and limitations — not just a legal disclaimer. |
| Health system / practice | Selecting the right tool for the clinical scenario, validating it in its own setting, training physicians on it, and monitoring its performance — with a process for responding when something goes wrong. |
The brief's own example makes the physician protection concrete: a triage model that silently deprioritizes a patient in a worklist gives the treating physician nothing to see but a reordered queue. There's no reasoning trail, no flag, no way to catch the error in the moment — so holding that physician solely liable punishes someone without the information needed to prevent the harm.
The brief also pushes back on one common contract pattern: vendor agreements that shift liability onto whichever party is least equipped to prevent harm, typically the physician or the health system, rather than the developer with the most visibility into how the model actually works.
Why the AMA is pushing this now
The brief leans on the AMA's 2026 Physician Survey on Augmented Intelligence. More than a third of physicians named clear liability rules as the thing most likely to increase their trust in and use of AI. Meanwhile, a quarter of physicians said they'd received no AI training at all, even as 92% said they wanted more. Put together: physicians are being asked to rely on tools they're undertrained on, inside a liability landscape nobody has clarified for them.
Health-system leaders echo the same concern from the deployment side. Barry Stein, MD, chief clinical innovation officer and CMIO at Hartford HealthCare, told healthsystemCIO that his organization doesn't want to "be reliant on other folks telling us how safe their products are without being able to independently verify it, 24/7, always on, not sporadically." Corewell Health's chief digital and transformation officer, Jason Joseph, made the scale problem explicit in the same piece: watch an AI tool make the same call ten times and a human notices patterns; a hundred times, attention fades; "a thousand times in a row, there is no way I'm going to find that needle in the haystack." Monitoring duty is easy to assign on paper and hard to execute at volume — which is exactly why the AMA ties it to a deploying organization's resources, not just its intentions.
The framework is harder on black-box tools than on reviewed drafts
Run the AMA's control principle against two different kinds of imaging AI and the gap shows up fast. A single-shot CAD or triage tool that outputs a binary flag or a risk score gives the radiologist almost nothing to check against — no visible reasoning, often little more than a liability disclaimer from the developer, and no second clinical check before the output reaches a human. Nearly all the practical liability exposure lands on whichever physician is staring at the flag, which is precisely the imbalance the AMA brief is trying to correct.
A workflow where AI produces a full report draft, a second clinician reviews it, and a signing radiologist checks a structured, already-reviewed document distributes that same control principle across three real checkpoints instead of concentrating it in one.
| AMA control-principle question | Black-box CAD / triage flag | AI-drafted, radiologist-reviewed report |
|---|---|---|
| Can the physician meaningfully exercise judgment? | Often not — a binary flag or score with no visible reasoning trail | Yes — a structured draft with findings a radiologist can check line by line |
| Does the developer help the user catch errors? | Varies; frequently limited to a liability disclaimer | A second clinical reviewer checks every draft before it reaches the signing radiologist |
| Who validates and monitors performance? | Left to the deploying health system alone, often discovered after the fact | Checked case by case by an independent reviewer, not just sampled in periodic audits |
Where xAID fits
The AMA brief isn't written about any one vendor, but its logic maps cleanly onto a draft-then-sign CT reporting workflow: xAID is accountable for the model that produces the structured draft, xAID's in-house radiologist reviews every preliminary as an independent clinical check before anyone downstream sees it, and the report reaches the client's reading radiologist ready-to-sign — a reviewed document to check, not a single opaque score to trust or reject blind. That's a materially different liability posture than a black-box tool with one physician as the only checkpoint, and it's also why the health system's own duty to select, validate, and monitor its AI vendor — the third leg of the AMA's framework — doesn't disappear just because the workflow already has two clinical reviews built in.
Frequently asked questions
What does the AMA's new issue brief say about AI medical liability?
On October 29, 2026, the AMA's Center for Digital Health and AI published an issue brief called "Liability for Clinical AI" arguing that liability for harm linked to clinical AI should be shared among the physician using the tool, the company that built it, and the health system or practice that deployed it — not placed solely on the physician who reviews the output.
Who is responsible for what under the AMA's liability framework?
The brief assigns physicians responsibility for the clinical judgment they can meaningfully exercise, such as questioning or overriding an AI's output. AI developers are responsible for designing tools that are safe and that let users identify errors, and for disclosing a tool's intended use, performance, and limitations beyond a legal disclaimer. Health systems and practices that deploy AI are responsible for selecting the right tool, validating it locally, training physicians, and monitoring its performance over time.
Why does the AMA say physicians shouldn't bear AI liability alone?
Because physicians often can't see or control everything an AI tool does. The brief's example is a triage algorithm that quietly reorders a patient queue: the physician sees only a reordered list, with no visibility into which patient it deprioritized or why. The AMA also cites a 2026 physician survey in which a quarter of respondents had received no AI training at all, 92% wanted more, and more than a third said clear liability rules would do the most to increase their trust in AI.
How does a draft-then-sign AI reporting workflow fit the AMA's liability framework?
A workflow where AI drafts a report, an in-house radiologist reviews every preliminary, and the client's reading radiologist signs the final gives each party the visibility the AMA brief calls for: the developer is accountable for the model, an independent clinical reviewer catches errors before the signing radiologist ever sees the draft, and the signing radiologist reviews a structured, already-checked report rather than a single opaque flag.
Source: MedTech Dive ("AI vendors, health systems join physicians in liability for clinical AI: AMA," Nov. 3, 2026); the AMA's "Liability for Clinical AI" issue brief (Center for Digital Health and AI, Oct. 29, 2026); and healthsystemCIO. Figures are rounded as reported.