A patient presents to the emergency department with chest pain radiating to the left arm. The attending physician conducts a thorough evaluation, considers and rules out a cardiac event, and diagnoses musculoskeletal strain with a clear plan for follow-up. The AI medical scribe, listening to the encounter through an ambient microphone, generates a clinical note. But the note contains a subtle and catastrophic error: it documents the chest pain but fails to record the cardiac workup that ruled out myocardial infarction. Two weeks later, the patient returns with an actual heart attack. The plaintiff's attorney pulls the medical record and finds no documentation of the prior cardiac evaluation. From the chart, it looks like the physician ignored obvious warning signs.
This scenario is not hypothetical. It reflects the type of clinical documentation failure that AI medical scribes are producing in hospitals across the country. I have reviewed technical documentation and error reports from multiple AI scribe deployments, and the pattern is consistent: these systems are good enough to be trusted, but not reliable enough to be trusted without verification. That gap is where malpractice liability lives.
The Scale of AI Scribe Deployment
The adoption of AI medical scribes has accelerated dramatically. Microsoft's Nuance DAX Copilot, the market leader, is now deployed in hundreds of healthcare systems. Abridge has secured partnerships with major academic medical centers including UCSF, Yale, and the University of Kansas Health System. Suki, Nabla, and DeepScribe are expanding rapidly in outpatient settings. By early 2026, it is estimated that AI scribes are documenting tens of millions of patient encounters annually.
The appeal is obvious. Physicians spend an estimated two hours on documentation for every one hour of patient care. Burnout driven by documentation burden is a leading cause of physician attrition. An AI scribe that listens to the patient encounter and automatically generates a structured clinical note promises to return hours of productive time to physicians every day. And for the most part, these systems work well. The clinical notes they produce are generally accurate, well-structured, and clinically appropriate.
The problem is that "generally accurate" is not the same as "reliable." And in medicine, the errors that matter most are precisely the ones that are hardest for a busy physician to catch during a quick review.
An AI scribe that is 97 percent accurate sounds impressive until you calculate what three percent error means across fifty patient encounters per day, five days per week, across thousands of physicians. The error volume is staggering.
How AI Scribes Fail
Having reviewed the technical architecture and error patterns of several leading AI scribe systems, I can identify the most common and most dangerous failure modes.
Omission of negative findings. This is the most clinically dangerous error type. When a physician says "no signs of peritoneal irritation" or "lungs clear bilaterally," the AI scribe sometimes fails to document the negative finding. In clinical practice, the absence of a documented negative finding can be interpreted as the physician failing to check. This is exactly the kind of error that transforms a defensible malpractice case into an indefensible one.
Misattribution of symptoms. AI scribes occasionally attribute symptoms to the wrong body system or misinterpret the clinical significance of a finding. A physician describing "tenderness in the right lower quadrant with rebound" might find the note documenting "abdominal tenderness" without the specific localization and associated findings that would indicate appendicitis. The clinical note reads as unremarkable when the actual examination was anything but.
Hallucinated content. Like all large language model-based systems, AI scribes can generate content that was never spoken. A scribe might insert a medication that the physician did not prescribe, document an allergy that was not discussed, or include a review of systems that did not occur. These hallucinations are particularly dangerous because they create false documentation that subsequent providers may rely upon.
Context collapse in complex encounters. When a patient presents with multiple complaints, or when the encounter involves interruptions, consultations, or changes in clinical direction, AI scribes struggle to maintain accurate context. The note may blend elements from different parts of the encounter, attribute statements to the wrong speaker, or lose track of the clinical reasoning that connected symptoms to diagnoses.
The Malpractice Framework
Medical malpractice requires four elements: duty, breach, causation, and damages. AI scribe errors intersect with this framework in ways that create novel liability questions.
Who breached the duty of care? The physician has a non-delegable duty to maintain accurate medical records. If the physician relies on an AI scribe and fails to catch an error in the generated note, the physician bears liability for the inaccurate record. This is true even though the physician did not write the note, because the physician signed it. From a legal perspective, signing an AI-generated note without adequate review is functionally identical to writing an inaccurate note.
What is the standard of care for AI scribe review? This is the question I expect to be litigating for years. If a physician reviews a ten-paragraph AI-generated note in thirty seconds before signing, is that adequate review? What about sixty seconds? What constitutes reasonable verification of an AI-generated clinical document? The standard of care has not crystallized, and expert testimony will be essential to establishing it.
Does the hospital bear institutional liability? When a hospital system mandates or encourages the use of AI scribes, it takes on institutional responsibility for the adequacy of the system. If the hospital deployed an AI scribe without adequate validation, without training physicians on its limitations, or without establishing review protocols, the hospital may face direct institutional liability under theories of corporate negligence.
The Vendor's Exposure
AI scribe vendors face potential products liability claims when their systems produce clinically dangerous errors. The key question is whether the error represents a defect in the product. If the vendor knew, or should have known, that the system had a tendency to omit negative findings, and failed to warn users or implement safeguards, that looks like a classic failure-to-warn or design defect claim.
Vendors typically attempt to disclaim liability through terms of service that characterize their product as a "documentation aid" that requires physician review. These disclaimers may limit contractual liability but are unlikely to insulate vendors from tort claims, particularly when the vendor's marketing materials emphasize accuracy and reliability.
What Attorneys Need to Know
Request the raw audio. Most AI scribe systems record the ambient audio of the patient encounter. This recording is the ground truth against which the AI-generated note can be compared. If the audio confirms that the physician conducted an appropriate examination but the AI scribe failed to document it, that fundamentally changes the malpractice analysis.
Examine the system's known error rates. AI scribe vendors conduct internal accuracy assessments. These documents, obtainable through discovery, will reveal the system's known failure modes, its error rates by clinical specialty, and any documented instances of the specific type of error at issue in the case.
Evaluate the hospital's deployment protocols. Did the hospital train physicians on the AI scribe's limitations? Did it establish minimum review times or verification procedures? Did it monitor error rates after deployment? The answers to these questions will determine the strength of institutional liability claims.
The AI medical scribe is here to stay. It solves a real problem, and physicians overwhelmingly prefer it to manual documentation. But the legal framework for allocating liability when these systems fail is still forming. The cases that establish the standard of care for AI-assisted clinical documentation are being filed now, and the technical evidence will determine their outcomes.
The Criterion AI provides expert witness services and litigation support for matters involving artificial intelligence, machine learning, and algorithmic decision-making. For a confidential consultation on an active or anticipated matter, contact us at criterion@thecriterionai.com or call (617) 798-9715.