For decades, American products-liability law has treated physicians as “learned intermediaries” between manufacturers and patients.
The doctrine is most familiar in prescription-drug and medical-device cases. A manufacturer generally satisfies its duty to warn by giving adequate warnings to the prescribing physician, not directly to the patient. The law assumes that the physician, because of training, clinical judgment, and knowledge of the patient’s condition, is best positioned to evaluate risks and benefits and then advise the patient accordingly.
That framework fits comfortably when the product is a drug, implant, or device. The manufacturer provides information. The physician interprets it. The patient relies on the physician’s judgment. But artificial intelligence in healthcare complicates this traditional model. When an AI system recommends a diagnosis, flags an abnormal image, predicts deterioration, or suggests a treatment pathway, it does not merely sit passively like a drug label. It participates, at least functionally, in the clinical reasoning process.
This raises a central question: if the AI system is capable of generating clinically meaningful conclusions, does the physician remain the sole learned intermediary, or does the AI become something closer to a second decision-maker?
The initial answer, almost certainly, is that courts will continue to place responsibility on the physician. Existing malpractice law is flexible enough to absorb new technologies. Courts have long evaluated physicians’ use of diagnostic tests, imaging studies, monitoring devices, and clinical guidelines by asking whether the physician acted as a reasonably prudent practitioner under similar circumstances. AI will likely be treated the same way at first: not as a new legal category, but as another clinical tool.
If AI is seen as a clinical tool, physicians may face liability for over-reliance on it. If an AI system produces a recommendation that conflicts with the patient’s presentation, accepted medical practice, or common clinical sense, a physician who follows it blindly may be accused of abandoning independent judgment. The “AI told me to do it” defense is unlikely to succeed if the physician could not articulate why the recommendation made clinical sense, much as a driver who follows GPS directions off a bridge cannot blame the navigation system.
But the more difficult problem lies in the opposite direction. As AI systems become more accurate, more widely adopted, and more deeply embedded in clinical workflows, physicians may also face liability for under-reliance. If a validated AI tool flags a malignancy, predicts sepsis, or recommends further testing, a physician who ignores the alert may need to explain that choice. Courts have already begun to grapple with analogous questions: physicians who override automated ECG interpretations or dismiss CT alert flags have faced scrutiny over whether their independent judgment was reasonable given the technology available to them. In that setting, AI begins to resemble not merely a tool, but a source of clinically relevant information that cannot safely be disregarded without documented justification.
This is where the learned-intermediary doctrine may begin to bend. The doctrine assumes that the physician is the superior interpreter of medical risk. But in narrow, well-validated domains such as radiology image classification, sepsis prediction, and diabetic retinopathy screening, AI systems are already demonstrating accuracy that rivals or exceeds the average clinician. The physician may remain the legal intermediary, but no longer the sole epistemic authority. Courts may increasingly ask not only whether the physician exercised judgment, but whether that judgment reasonably accounted for the AI output.
For practicing physicians, the practical implication is documentation. The safest future chart may not simply record the final diagnosis or treatment plan. It may also record what the AI suggested, whether the physician agreed or disagreed, and why. In this sense, AI may become legally analogous to a specialist consultation or diagnostic test result: not binding, but not casually dismissible.
A related question, one that is largely unresolved, is whether patients have a right to know that AI played a role in their care. If the learned-intermediary doctrine is premised on the physician mediating between manufacturer and patient, transparency may be part of that mediation. Several states have begun exploring disclosure requirements for AI-assisted clinical decisions. Whether informed-consent doctrine will eventually require disclosure of AI involvement, and under what circumstances, is a thread the courts and legislatures have yet to fully pull.
Hospitals and health systems will also face liability exposure under this framework. As the entities that select AI tools, negotiate vendor contracts, configure systems, train clinicians, and decide whether alerts are mandatory, advisory, or suppressed, institutions are well-positioned defendants for claims of negligent implementation, inadequate validation, poor post-deployment monitoring, or failure to disclose system limitations to treating physicians. These claims will look less like traditional malpractice and more like product selection and institutional governance failures.
The law is unlikely to announce a dramatic new rule for AI malpractice. More likely, it will adapt old doctrines incrementally: applying existing negligence standards to new facts, extending familiar frameworks for specialist consultations and diagnostic tools, and calibrating expectations of physician conduct as AI becomes a routine part of clinical practice. Through that process, the physician will almost certainly remain the legal learned intermediary, the person the law holds responsible for the final clinical judgment.
But that framing may increasingly obscure what is actually happening at the bedside. The question is whether, in the age of clinical AI, the physician is mediating between patient and manufacturer, or between patient and machine judgment.
