Malpractice to Not Use AI? The AMA Draws a Hard Line Against Tech Giants

Maya: Welcome to AI in Medicine - Smart Summaries. Before we start, a quick reminder that our voices are AI-generated, and this is a summary of a public document.

Sam: And today we are digging into a truly fiery one. The American Medical Association is drawing a hard line in the sand regarding autonomous AI in clinical care.

Maya: Exactly. The single most interesting finding here is that the AMA is demanding a strict legal framework that protects physicians who choose not to use AI, pushing back aggressively against tech leaders who claim practicing without AI is malpractice. Today's document is titled Minimizing Patient Harm from Autonomous Artificial Intelligence. It is a late resolution published by the American Medical Association on October 22, 2026.

Sam: This document, specifically Late Resolution 1, introduced by Raymond Lorenzoni for the Interim 2026 YPS Assembly, is basically an emergency brake. The Young Physicians Section is looking around at the rhetoric coming from government officials and Silicon Valley and saying, wait a minute, we need to assert some boundaries right now. Maya, set the scene for us. Why is this dropping as a late resolution in late 2026?

Maya: It comes down to some highly publicized comments made just weeks prior. The resolution specifically cites statements from September 22, 2026. The HHS Secretary, Robert F. Kennedy Jr., stated that artificial intelligence can give you a second opinion that is much better informed than any doctor in the country.

Sam: That is quite a statement from the head of Health and Human Services.

Maya: It is. But the document notes he went further, relaying a claim by OpenAI founder Sam Altman. According to the text, Altman claimed that it would be malpractice if physicians diagnosed or prescribed without consulting AI.

Sam: Wow. Calling it malpractice not to use their product. I can see why physician organizations felt the need to respond.

Maya: They did respond initially. The AMA and five other national physician organizations issued a joint statement about physician leadership in AI development. But the authors of this late resolution argue that the initial response lacked specific demands. They feel the AMA has failed to protect patients and physicians from interests trying to equate AI to human doctors.

Sam: And the resolution points to specific marketing hype that is blurring these lines, right? They mention phrases like The first AI system legally authorized to practice medicine and The first ever end-to-end AI doctor. The document links these claims to things like the Utah AI physician experiment and a company called Nolla Health.

Maya: Yes, and the AMA Young Physicians Section argues these are misleading claims based on limited or simulated clinical tasks. They state these claims have dangerous implications for patient safety, public trust, and the physician-patient relationship. In medicine, individualized assessment, shared decision-making, and professional accountability are non-negotiable. Substituting an autonomous AI negatively alters that dynamic.

Sam: So they are looking for a legislative shield. They actually point to a law that already passed as a model, Delaware House Bill 191.

Maya: That is right. Delaware House Bill 191 was enacted in 2026 as 85 Del. Laws, chapter 250. It prohibits non-human entities, including AI agents, from being certified to practice medicine or assigned titles reserved for physicians.

Sam: Which sounds like a great start, but the resolution notes that the Delaware law is limited in scope. So, what exactly is the AMA resolving to do about this at a national level?

Maya: First, they want the AMA to vocally oppose granting or equating AI systems with physician licensure or certification to practice medicine. But the real meat is in the second resolved clause, where they demand the AMA write and disseminate model legislation with some very strict prohibitions.

Sam: Let us walk through those demands, because if you are a tech company building healthcare AI, this model legislation is basically a blueprint for the regulatory hurdles you are about to face in every state.

Maya: Exactly. The model legislation must, at a minimum, prohibit misrepresenting non-human entities with titles reserved for physicians. It must define the distinction between authorizing a medical technology and granting professional licensure.

Sam: Makes sense. An FDA clearance is not a medical license. What else?

Maya: This is a big one. It prohibits non-human entities from performing clinical tasks traditionally associated with practicing medicine without real-world evidence of superior safety and efficacy.

Sam: Wait, really? Superior safety and efficacy? Not just non-inferiority? That sets the bar incredibly high for autonomous AI developers.

Maya: It absolutely does. Then, tackling data, the model legislation must prohibit AI organizations from prospectively collecting clinical patient data for proprietary or research purposes without explicit authorization by the patient.

Sam: So companies cannot just quietly scrape real-time hospital workflows to train their next models without the patient explicitly saying yes.

Maya: Right. And then we get to the liability piece, which directly answers that Sam Altman quote. The AMA wants a liability framework protecting physicians independent decision-making from liability arising from non-use of AI recommendations or non-use of AI tools in general.

Sam: That is fascinating. They are legally demanding the right to ignore the machine without getting sued. And they also want a robust supervision framework for physicians overseeing AI, including adequate reimbursement for that supervision.

Maya: Exactly. The overarching theme is that AI should enhance human intelligence, not replace it. And if doctors are expected to babysit these models, they need to be paid for that time and protected from liability if they disagree with the model.

Sam: To really understand how the AMA is thinking about this, the document appends several massive sections of existing AMA policy. Let us unpack those, because they reveal exactly how the AMA wants to shift the financial and legal risks onto the tech companies.

Maya: Yes, let us look at policy H-480.939 on Augmented Intelligence. It states clearly that physicians should not be penalized if they do not use AI systems while regulatory oversight and standards of care are still in flux. It opposes any mandates by payers, hospitals, or government entities that force the use of AI as a condition of payment or licensure.

Sam: So if a hospital administrator buys a fancy new AI diagnostic tool, they cannot force the clinicians to use it just to justify the cost. In fact, the policy says that if a mandate prevents a doctor from mitigating risk, the entity issuing the mandate assumes all applicable liability.

Maya: Precisely. But where it gets really tough for tech companies is the liability for autonomous systems. The policy says developers of autonomous AI systems with clinical applications—meaning screening, diagnosis, or treatment—are in the best position to manage liability from system failure or misdiagnosis.

Sam: And how do they expect the developers to manage that?

Maya: The AMA states developers must accept this liability with measures such as maintaining appropriate medical liability insurance and writing it into their agreements with users.

Sam: Tech companies buying medical malpractice insurance? That completely upends the software-as-a-service business model. Usually, software vendors put a clause in their contract saying 'we are not providing medical advice, the doctor holds all the risk.' The AMA is saying, if your tool is autonomous, you hold the bag.

Maya: They are. The policy also strictly targets gag clauses. It says AI systems subject to non-disclosure agreements concerning flaws, malfunctions, or patient harm must not be covered or paid for. If a party enforces a gag clause, they assume liability for any harm.

Sam: That is huge for transparency. You cannot hide your AI's mistakes behind a corporate NDA. Which brings us to the transparency requirements detailed in policy H-480.391. If I am selling AI to a health system, what exactly do I have to disclose before they purchase it?

Maya: Quick note before we carry on. This spot is open for a sponsor. If your company builds or sells AI for healthcare and wants to reach the clinicians, health-system leaders and industry teams who listen to this show, the link to our sponsorship page is in the show notes.

Sam: And now, back to the document.

Maya: The list is extensive. Developers must disclose their regulatory approval status, intended population, and any limitations or possible disparate impacts. But the data requirements are the most intense. You must provide detailed information on data provenance, size, completeness, timeframes, diversity, and data labeling accuracy.

Sam: So no more 'we trained it on a proprietary black-box dataset.' You have to prove where the data came from and that it is actually diverse.

Maya: Exactly. You also have to show evidence of external validation processes for ongoing evaluation of model performance. The AMA specifically notes you have to account for AI model drift and degradation.

Sam: Because an algorithm trained in 2024 might perform terribly on new patient populations in 2027. It is not set-and-forget software. What about generative AI specifically? Large language models are notorious for making things up.

Maya: The policy requires evidence that the tool is constrained to evidence-based outcomes and mitigates hallucination or confabulation. It also demands clear disclosures on how the algorithm is maintained, including whether active patient data is being used for ongoing training.

Sam: That goes back to the patient privacy aspect. The policy states individuals should have the right to opt-out, update, or request deletion of their data from generative AI tools. And that right encompasses the training data itself.

Maya: Right, the document says tools should not reverse engineer, reconstruct, or reidentify an individual's originally identifiable data. Preventive measures must include secure enclaves, federated learning, and differential privacy.

Sam: Let us shift to how this impacts the daily clinical workflow. The AMA insists there must be a qualified human in the loop for medical decision making. How do they define a qualified human?

Maya: A qualified human is defined strictly as a licensed physician with the necessary qualifications and training to independently provide the same medical service without the aid of AI. And the policy states that as the potential for patient harm increases, the physician must intervene earlier in the care plan to interpret or act on an AI recommendation.

Sam: So you cannot just have an administrator or a junior staffer clicking 'approve' on an AI output. It has to be a doctor who could do the job themselves.

Maya: Exactly. And they require that any use of AI that impacts access to care or medical decision making must be documented in the medical record and disclosed to patients in a culturally and linguistically appropriate manner.

Sam: If I am a patient interacting with a hospital chatbot, do they have to tell me it is an AI?

Maya: Yes. The use of AI must be clearly disclosed at the beginning of the encounter. And where patient-facing content is generated by AI, its use must be noted within that content.

Sam: This document does not just target tech startups and hospitals. It takes direct aim at health insurance companies. We know payors have been using automated systems to process claims, sometimes resulting in mass denials. What does the AMA say about that?

Maya: The language here is incredibly forceful. The AMA states payors should only use automated decision-making to enhance efficiencies or reduce workflow burdens. They explicitly say these systems should never create or exacerbate overall or disparate access barriers by increasing denials or limiting benefit offerings.

Sam: But payors are constantly using algorithms to evaluate if a procedure is medically necessary. What guardrails is the AMA demanding?

Maya: First, the criteria must be publicly reported and based on easily accessible evidence-based clinical guidelines, not proprietary payor criteria. Second, decisions cannot be based on algorithms developed on similar or like patients. It must allow for the individualized assessment of a patient's specific circumstances.

Sam: And if the algorithm does spit out a denial, what happens?

Maya: This is a massive operational hurdle for payors. The policy demands that any automated recommendation indicating a denial of care, at both initial and appeal levels, must be automatically referred for review to a physician. But not just any physician. It must be a physician with a current, valid, non-restricted license in the state, and they must be of the same specialty as the physician who typically manages that medical condition.

Sam: Wow. Same specialty match required for every AI-driven denial. And does the treating doctor get a say?

Maya: Yes. Prior to issuing an adverse determination, the treating physician must have the opportunity to discuss the medical necessity directly with the reviewing physician.

Sam: That essentially neutralizes the speed and cost-saving advantages an insurance company gets from using an AI denial engine. Furthermore, the AMA wants payors to make their statistics regarding AI approval, denial, and appeal rates available on their website, broken down by patient demographics to monitor equity.

Maya: Correct. They even suggest insurance regulators should require reporting of payor AI use to monitor for negative impacts on access to care.

Sam: Let us talk about misinformation. We all know generative AI can confidently output completely wrong medical facts. What is the responsibility of the developers here?

Maya: The policy states algorithms should be developed to detect and flag potentially false and misleading content before it is widely disseminated. Developers must have mechanisms for users to report mis- and disinformation. And again, the policy ties this back to liability. Developers must accept liability for system failure or misdiagnosis arising from their tools.

Sam: They also briefly touch on medical publishing. Academic journals need clear guidelines requiring the disclosure of AI used in research methods, and they explicitly demand the exclusion of AI systems as authors.

Maya: Yes, and the authors of the paper remain entirely responsible for validating the veracity of any referenced content generated by AI. There is no passing the buck to the algorithm.

Sam: Cybersecurity is another massive risk. Hackers getting into clinical decision systems. What does the policy require?

Maya: AI systems must have strong protections against input manipulation and malicious attacks. Entities deploying these tools have to regularly monitor for anomalies, comparing AI outputs against known and normal behavior. And if there is a data breach, the entity is expected to act diligently in notifying the impacted individuals directly.

Sam: So looking at this entire resolution—the late pushback from YPS, the model legislation demands, the appended policies—what is the big takeaway? The resolution asks for these issues to be studied further with recommendations reported to the 2027 Annual Meeting of the House of Delegates. But what does this mean in practice right now?

Maya: For clinicians on the ward, it is a rallying cry. It validates the anxiety that tech companies are trying to force AI adoption by threatening the malpractice label. The AMA is clearly stating that your clinical judgment and your right to not use an AI tool must be protected. The human is the ultimate safeguard.

Sam: And for the boardroom? For the health-system leaders buying these tools or the medtech companies building them?

Maya: It is a severe reality check. Marketing your product as an autonomous AI doctor is going to attract intense legislative pushback. If you want to sell autonomous clinical AI, you need to be prepared to take on the medical liability insurance, provide massive transparency into your training data, and prove superior safety and efficacy in the real world.

Sam: The era of move fast and break things definitely does not apply when the AMA is actively drafting legislation to stop you. To recap, the AMA's late resolution aggressively pushes back against government and industry claims of AI superiority. They are demanding model legislation to ban AI from practicing medicine, a liability shield for doctors who choose not to use AI, and forcing autonomous AI developers to hold the legal risk.

Maya: It is a pivotal document for the future of clinical AI. As always, you can find a link to the full resolution in the show notes. Remember, this podcast is for informational purposes and is not medical advice. Thanks for listening.