This is the last of three parts. In Part 1, I used Chi-way Li’s count of FDA-authorized AI devices to show that clearance is a license to sell, not a payment: of 1,614 devices, Li could tie only 79 to a billing code of their own, and most of the rest are bought from provider budgets. In Part 2, I looked at the best-measured cases, AI scribes and inbox drafts. There the physician’s review is real work that I found no one who has totaled or priced, and perceived time savings outrun measured ones, though the well-being gains are real. This part asks what happens when that review fails, and who is left holding the cost. It is written to be read on its own.
My vantage point is that of a gastroenterologist with more than 30 years in practice, now working as a hospitalist, who is also founder and CEO of AiMOS Systems Corporation. Weigh my arguments with that in mind.
The AMA framework
On October 16, 2026, the American Medical Association released a one-page governance framework for what it calls agentic AI assuming clinical functions. Note the scope. It is written for agentic systems, meaning AI that takes on clinical functions, and not specifically for note drafting or inbox replies. I am applying it to the tools physicians use now because the accountability question it raises does not stop at the edge of that category.
The framework asks four questions: what function the AI performs, what evidence shows it can do so and what happens if it fails, what the AI is permitted to do and under what oversight, and who is responsible for risks and outcomes. For the third question it uses a ladder of increasing authority: inform, recommend, decide, act. It adds a governing principle that clinical authority should be earned through evidence and maintained through demonstrated performance.
On accountability, the framework is specific. It says accountability should be allocated according to authority, control, knowledge of risk, and the ability to prevent or mitigate harm. The parties it lists are the developer, deployer, health system, physician, payer and other actors. It does not place liability on the physician’s oversight step. Liability appears as one of three levers, alongside regulation and licensure, and the framework defines it as responsibility for decisions, actions and failures. The AMA’s own page describes the framework in the same terms.
I think that is a sound principle. It is also silent on one point. I found nothing in the framework about who pays for the oversight it presumes. It tells us who should answer for harm in proportion to control. It does not say who funds the time, staffing or tooling that make oversight possible. That is my reading of a one-page graphic, and a longer AMA document could say more.
What carriers say
Allocation on paper meets a different reality in a courtroom, so I looked at what malpractice carriers have said publicly. The record is thin.
An executive of The Doctors Company told Medical Economics in October 2025 that the company currently has no AI exclusion and would defend and potentially indemnify a physician in an AI-related claim. That is one carrier speaking through its own executive, and the wording is hedged.
Peter Kolbert of Healthcare Risk Advisors, part of TDC Group, told Business Insurance in 2023 that coverage applies even when AI is a big factor in a claim. The article noted that this position had not been tested in court. It is also three years old, in a field that has changed quickly.
I searched for AI position statements from MedPro, MagMutual and Coverys and did not find any. That is a limit of my searching, not proof that none exist.
So the carriers on record are telling physicians they will be defended. That is useful to know. It is also an answer to a different question from the one this series has been asking.
Claims so far
I found no reported malpractice case against a clinician over the use of AI. Absence in my search is not absence in the world, and claims may move through insurers and settlements before they become public.
What I did find are lawsuits of other kinds, and all of them are allegations, not findings. A privacy suit, reported in April 2026, alleges that an AI scribe unlawfully recorded patient and clinician conversations in a case involving Sutter Health (accounts differ on the defendants named). It does not concern a missed diagnosis. It concerns consent to recording.
Separately, a man has sued OpenAI alleging that medical advice from ChatGPT delayed care for a pulmonary embolism, and Pennsylvania has sued Character.AI alleging that a chatbot posed as a licensed psychiatrist. These are suits against AI companies by a patient and a state. I cite them only to show where the first legal tests are appearing, which is not yet in malpractice claims against physicians who relied on a tool.
The exclusion trend
There is one more signal, and it is a warning rather than a fact. In January 2026, according to a law firm’s summary, ISO introduced a generative AI exclusion for general liability policies, and coverage for AI risk is fragmenting. That is general liability, not medical malpractice, and the source is law-firm marketing that says nothing about malpractice.
An Australian company planning to sell standalone AI cover has forecast that AI exclusions may reach malpractice and professional indemnity policies. That company is not yet selling the product and has a financial interest in the prediction, and it is a forecast, not an event.
I do not read either as evidence that physicians will lose coverage. I read them as a reason not to treat today’s reassurance as permanent. If exclusions spread, the physician’s position would change from defended to exposed, while the expectation of review stayed exactly the same.
Does the check work?
Everything above assumes the physician’s review is a meaningful safeguard. The best test of that assumption I found is a medRxiv preprint by Barrit and colleagues. It is not peer reviewed, and its design limits what it can say.
Here is what the study did. It gave 225 physicians written neurocritical-care cases with advice from a conversational AI. The AI was deliberately steered toward harmful choices in some conditions, which is manipulation and not the routine error of a scribe or a drafting tool. It used one open-source model, and an LLM judged the outcomes. With those limits stated, physicians made a harmful decision in 49.4 percent of targeted decisions when the AI was steering them, compared with 2.0 percent when the AI was aligned. Experience gave no clear protection. Of the 190 physicians who made a harmful decision, 141 flagged nothing, and only 2 physicians detected the manipulation, as the preprint reports.
What does this show? It shows that in a written-case setting, with deliberately manipulated advice from one model, physicians were often led to harmful decisions and mostly did not notice. It does not show how often real tools mislead real clinicians in practice. It does not measure routine errors. It does not show that the review step is useless, and the 2.0 percent figure suggests aligned advice produced little harm in this setting.
My inference, which is not the authors’ finding, is that the phrase “a clinician will review it” describes an intention and not a tested safeguard. The preprint gives one reason to doubt it under adversarial conditions. Part 2 gave another: I found no study that measures how much review time the intention requires. Taken together, the check is both unpriced and unproven.
The strongest objection
The strongest objection to my argument is that none of this is new. Liability for a tool has always sat with the physician. We are responsible for the labs we order, the monitors we read and the references we consult, and malpractice carriers exist to pay when care fails. If the carriers are saying they will defend physicians who use AI, the system is working as designed.
I take that seriously. It is the reason I started with the carriers on record. My answer is an argument, not a finding. Carriers price failure, not the review minutes that prevent it. A policy that defends me after an error does not fund the time I spend looking for one. When a lab value is wrong, the review of that single value is a small part of an encounter that is already paid. When AI produces a note, a reply draft, an image read or an alert at volume, the review scales with the number of outputs, and no payment I know of scales with it. The liability framework and the payment system answer different questions, and the gap between them is where the cost of the check falls on the individual physician.
What a fix would need
A fix needs three things, and none is exotic.
First, a payer for the review. Someone has to be assigned the cost of the check. The AMA framework already names the candidates: developer, deployer, health system, physician, payer. Allocating accountability in proportion to control, as the framework proposes, implies allocating the cost of oversight in the same way. That step is my extension of the framework, not something it says. A developer who sells a tool that requires review, or a health system that deploys it to save staffing, are plausibly the parties with the most control over how much review the tool will need.
Second, a measure of review time. I found no study that totals review time across tools or puts a price on it, and I found none measuring it for imaging AI. Without a measure, the cost cannot be assigned, and the promised saving cannot be netted against what review takes. The first requirement depends on the second.
Third, checking tools that are independent of the AI doing the advising. If the same system produces the suggestion and supplies the reassurance, the check inherits the system’s blind spots. I offer this as my proposal, not the preprint authors’.
Closing the series
Across the three parts, the argument has one shape. Clearance lets a developer sell a tool, but it does not create a payer for it. The buyer is usually a provider’s budget, so the cost of making the tool safe lands on the provider. The physician’s review is the safeguard most deployments rely on, yet it is real work that I found no one totaling or pricing. And when the safeguard fails, the physician is the one defended, and potentially the one exposed.
None of that is an argument against medical AI. The well-being gains from scribes are real, and some AI reads and measurements will earn their payment. My argument is narrower. A system that depends on a safeguard should say who pays for the safeguard, measure how long it takes and test whether it works. In what I found, it does none of the three.
Li ended his first article with a question for hospital leaders: which budget paid for the last AI tool you rolled out, and did the promised saving show up? I would add one. In the contract for that tool, who is assigned the cost of checking it?
Brian K. Hudes, MD, is a practicing gastroenterologist and writes The Renaissance Physician. Disclosure: he is the founder and CEO of AiMOS Systems Corporation and has a commercial interest in health care technology and clinical artificial intelligence.
Selected sources
Chi-way Li, Healthcare AI Insights, Oct 14, 2026, updated Oct 16 (analysis of who pays for cleared AI): Most cleared AI is bought, not billed
American Medical Association, Oct 2026 (one-page governance framework, PDF): Governance framework for agentic AI
American Medical Association, Oct 2026 (accompanying page): AMA framework page
Littrell, Medical Economics, Oct 2025 (interview on malpractice insurers and AI risk): AI on trial
Wilkinson, Business Insurance, Oct 2, 2023 (insurer positions on medical AI): Insurers pause to take pulse of medical AI
Hopkins, Fenwick, Jun 15, 2026 (law-firm summary of AI exclusions): Fenwick publication
Mitchell, IT Brief Australia, Dec 16, 2025 (forecast by an insurer planning AI cover): IT Brief article
Alder, HIPAA Journal, Apr 14, 2026 (reported allegations in a scribe recording suit): HIPAA Journal article
Attridge, Courthouse News, Jul 22, 2026 (reported allegations against OpenAI): Courthouse News article
NPR, May 5, 2026 (state suit against Character.AI): NPR report
Barrit et al., medRxiv preprint, 2026, not peer reviewed (physician decisions with steered AI advice): Subversion of clinical judgment by conversational AI