In our February analysis of the AI medical scribe problem, we outlined the technology, the risk landscape, and the emerging standard of care for physicians using AI-generated clinical documentation. Since that article was published, two things have happened. First, three new malpractice cases involving AI scribe errors have been filed in state courts in California, Texas, and Massachusetts. Second, I have been retained as an expert witness in one of those cases. I cannot discuss the specifics of the matter, but the general litigation dynamics are instructive and worth examining in detail.
This article is the playbook. It covers the litigation strategy, evidence preservation requirements, expert witness methodology, and practical considerations for both sides of AI medical scribe malpractice cases.
The Anatomy of an AI Scribe Malpractice Case
The typical AI scribe malpractice case follows a predictable pattern. A patient visits a physician. The physician uses an AI medical scribe (Nuance DAX, Abridge, Nabla, or a similar product) to generate clinical documentation from the encounter. The AI scribe produces a clinical note that contains an error: an omitted symptom, a mischaracterized finding, an incorrect medication dosage, or a fabricated element that was never discussed. The physician signs the note without catching the error. A subsequent treating physician relies on the note. The patient is harmed.
The legal question is straightforward: who is liable? The answer is complicated.
The physician signed the clinical note, making it a legal medical record. Under the prevailing standard of care, a physician is responsible for reviewing and verifying the accuracy of clinical documentation before signing. If the physician failed to catch an error that a reasonable physician exercising due care would have caught, the physician breached the standard of care.
The health system selected and deployed the AI scribe system. If the system was known to have accuracy limitations that the health system failed to disclose to physicians, or if the health system failed to implement adequate quality assurance processes for AI-generated documentation, the health system may be vicariously liable for the physician's reliance on defective output and independently liable for negligent deployment.
The AI vendor developed and sold the scribe system. If the system had a defect that caused the error, the vendor may be liable under product liability theories: design defect (the system's architecture made it prone to certain types of errors), manufacturing defect (a specific instance of the system malfunctioned), or failure to warn (the vendor did not adequately disclose the system's known limitations and error rates).
Evidence Preservation: The Critical First Step
AI scribe malpractice cases have a unique evidence preservation problem. The AI-generated note is only one piece of the evidence. The complete evidentiary picture includes:
The audio recording. Most AI scribe systems work by recording the patient encounter and transcribing it using speech-to-text technology, then summarizing the transcription into a structured clinical note. The original audio recording is the ground truth. It shows exactly what was said during the encounter, and comparing it to the AI-generated note reveals precisely where the AI introduced errors. Some systems retain audio recordings; others delete them after generating the note. If your case involves an AI scribe, send a preservation demand for the audio recording immediately. Every day of delay increases the risk that the recording has been deleted.
The intermediate transcript. Between the raw audio and the final clinical note, many AI scribe systems produce an intermediate verbatim transcript. This transcript is then processed by a summarization model that generates the structured note. Errors can be introduced at either stage: the speech-to-text model may misrecognize words, or the summarization model may omit, recharacterize, or fabricate content. Preserving the intermediate transcript allows the expert to determine where in the pipeline the error occurred.
The model version. AI scribe systems are continuously updated. The version of the model that generated the note at issue may differ from the current version. If the vendor has updated the model to fix the type of error that caused the harm, that update is powerful evidence that the vendor was aware of the defect. Preservation demands should specifically request the model version identifier, release notes for that version, and any known issue reports.
The edit history. Some AI scribe systems track edits made to the generated note before the physician signs it. If the physician made edits, those edits show what the physician reviewed and changed, and by implication, what the physician did not review or change. The edit history is relevant to the physician's standard of care: did the physician meaningfully review the note, or simply sign it?
System-wide error data. The vendor's data on the system's error rates, error types, and known failure modes is critical for both product liability claims and for establishing the standard of care. If the vendor knows that its system mischaracterizes medication dosages 2% of the time, that is a known defect. If the health system was not informed of that error rate, that supports a failure to warn claim.
The audio recording is the single most important piece of evidence in an AI scribe malpractice case. If it is deleted before you can preserve it, you have lost your best evidence. Send the preservation demand on day one.
The Expert Witness Methodology
As an AI technical expert in these cases, my analysis follows a structured methodology.
Step 1: Reconstruct the pipeline. I examine the AI scribe system's architecture to understand how it processes audio input into clinical documentation. This includes the speech recognition model, the natural language processing pipeline, the summarization model, and any post-processing rules or templates. Understanding the pipeline is essential to identifying where the error was introduced.
Step 2: Compare outputs to ground truth. I compare the AI-generated clinical note to the audio recording (if available) or to other contemporaneous records of the encounter. This comparison identifies every discrepancy between what actually occurred during the encounter and what the AI documented. Discrepancies are categorized by type (omission, mischaracterization, fabrication, dosage error) and severity (clinically insignificant, clinically relevant, clinically dangerous).
Step 3: Determine the error mechanism. For each clinically significant discrepancy, I analyze the technical cause. Was it a speech recognition error (the model misheard a word)? A summarization error (the model omitted or recharacterized information during summarization)? A hallucination (the model generated content that was not present in the source audio at all)? The error mechanism determines which defendant bears responsibility: speech recognition errors may be a vendor defect, while summarization errors may reflect a design choice that the vendor made and the health system accepted.
Step 4: Assess detectability. I evaluate whether the error in the AI-generated note would have been detectable by a physician reviewing the note with reasonable care. This is the crux of the physician's standard of care defense. Some AI errors are obvious on their face (a note that says "patient denies chest pain" when the encounter was clearly about chest pain). Others are subtle (a note that lists 500mg of a medication when the physician said 50mg). The detectability of the error determines the allocation of liability between the physician and the AI vendor.
Step 5: Evaluate industry standards. I assess the AI vendor's system against the current standard of care for clinical documentation AI, including published accuracy benchmarks, FDA guidance (where applicable), and industry best practices for deployment and quality assurance. This assessment is relevant to product liability claims and to the health system's duty of care in selecting and deploying the system.
Strategic Considerations for Plaintiffs
Name all defendants. The physician, the health system, and the AI vendor should all be named as defendants. The allocation of liability among them will be determined at trial, but naming all parties ensures that the full evidentiary picture is available through discovery from each defendant.
Focus discovery on the vendor. The vendor has the most information about the system's known limitations, error rates, and failure modes. Discovery should target the vendor's internal testing data, customer-reported errors, model version history, and any communications about known issues. This evidence is essential for product liability claims and for establishing that the physician's reliance on the system was reasonable (or unreasonable, depending on what was disclosed).
Retain the AI expert early. The technical analysis in these cases is complex and time-consuming. Retaining an AI expert early in the case allows the expert to advise on evidence preservation, guide discovery strategy, and begin the technical analysis while the evidence is still available.
Strategic Considerations for Defense
Physician defendants should emphasize the standard of care. If the physician reviewed the note and the error was not reasonably detectable, the physician did not breach the standard of care. Expert testimony on the detectability of the specific error type is critical to this defense.
Vendor defendants should emphasize the learned intermediary doctrine. The physician is the learned intermediary between the AI system and the patient. If the vendor adequately disclosed the system's limitations and the physician assumed responsibility for verifying the output, the vendor may argue that the physician's failure to catch the error breaks the chain of causation.
Health system defendants should emphasize compliance with deployment guidelines. If the health system followed the vendor's recommended deployment guidelines, implemented quality assurance processes, and provided physician training on AI scribe limitations, the health system may have a strong defense against negligent deployment claims.
The Standard of Care Is Forming Now
These early cases will establish the standard of care for AI-assisted clinical documentation. The outcomes will determine how much review physicians are expected to perform, what disclosures vendors are required to make, and what quality assurance processes health systems must implement. Every case that is filed, every expert report that is submitted, and every judicial decision that is issued contributes to the emerging standard.
For practitioners on both sides of these cases, the expert witness is the differentiator. The technical analysis that determines where in the AI pipeline the error occurred, whether the error was detectable, and whether the system met industry standards will determine the outcome. Retain the right expert, preserve the right evidence, and build the case on solid technical foundations. That is the playbook.
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 info@thecriterionai.com or call (617) 798-9715.