Who is liable when an AI scribe makes an error?

Short answer

The clinician who signs the note, according to every medical board, malpractice insurer and scribe vendor statement found. No published court decision on an AI-written note was found as of Sept. 28, 2026. Published vendor terms for individual clinicians cap the vendor's liability at small sums, and Abridge's make the clinician responsible for patients' consent to recording, which is where the lawsuits are.

Boards say the signature decides

The Federation of State Medical Boards, in policy adopted in April 2024: "the physician is ultimately responsible for the use of AI and should be held accountable for any harms that occur." On scribes, it warns that AI-generated records "without proper oversight, may lead to inaccurate documentation and subsequent patient harm for which the physician will likely be accountable" (FSMB). The North Carolina Medical Board's records statement, amended in September 2026, says it is "incumbent upon the licensee to ensure that the transcription of notes is accurate," particularly when records are generated "with the assistance of dictation software or artificial intelligence" (NCMB). The Oregon Medical Board says it "will continue to hold licensees responsible for the care they provide" (Oregon Medical Board). In Texas, a practitioner who uses AI for diagnostic purposes must review all records created with it and disclose the use to patients, under a law in effect since Sept. 1, 2025 (Texas SB 1188).

Malpractice insurers say the same. The Texas Medical Liability Trust: "physicians remain fully responsible for the content of all medical documentation, regardless of how it was generated," and "Automatic signatures on AI-generated content are discouraged" (TMLT).

The AMA wants the risk shared

The American Medical Association's principles argue that liability and incentives should be aligned so that those "best positioned to know the AI system risks and best positioned to avert or mitigate harm" do so, and that physicians who had no reason to know of a tool's problems "should not be held liable for the performance of the technology in question." They also say AI tools "cannot augment, create, or otherwise generate records" on a physician's behalf "without that physician's consent and final review" (AMA).

No published ruling yet

No public malpractice cases are tied to AI scribes, Healthcare Dive reported in September, but "that doesn't mean there are zero," said Bill Satterwhite, a physician and lawyer at Huron Consulting Group (Healthcare Dive). A 2025 journal article in JCO Oncology Practice on the liability risks of ambient documentation found no case law on the subject, according to a summary by the European CLASSICA project (CLASSICA). Until a court rules, boards and insurers treat an AI-drafted note like any other documentation the physician signs.

How often AI notes get it wrong

In an experimental study of notes generated by GPT-4 from consultation transcripts, 1.47% of sentences contained a hallucination and 3.45% of transcript sentences were omitted; 44% of the hallucinations and 16.7% of the omissions were judged major (npj Digital Medicine). In 14 simulated visits run through four commercial scribes and one free one, the mean error rate in the notes was 26.3%, with an average of three errors per case that could cause moderate to severe harm (Mayo Clinic Proceedings: Digital Health). In a randomized trial at UCLA, physicians reported that notes "occasionally" contained clinically significant inaccuracies, most often omissions or pronoun errors (UCLA Health). Almost all general practitioners in a Dutch study said the tool did not summarize the physical exam correctly (npj Digital Medicine).

What the contracts say

Published terms for individual clinicians cap what the vendor owes. Freed limits direct damages to the greater of $100 or three months of fees (Freed). Abridge caps its liability at the greater of $1,000 or 12 months of fees and says the clinician, "and not Abridge," is "solely responsible" for obtaining patients' consent (Abridge). Microsoft's terms for Dragon Copilot require the customer to ensure that the provider of record reviews all output and takes clinical responsibility "by finalizing and signing Encounter Documentation in their own name" (Microsoft). Health systems negotiate their own contracts, which are not public.

The lawsuits are about recording

California requires the consent of everyone in a confidential conversation before it is recorded, and a person recorded without consent can sue for $5,000 per violation or three times actual damages, whichever is greater (Penal Code 632; 637.2). Patients have sued Sharp HealthCare in San Diego County (Becker's) and Sutter Health in federal court in Northern California (complaint), both over Abridge's scribe. About 11 states generally require the consent of everyone recorded, and Missouri and Oregon require it for in-person conversations, according to the Reporters Committee for Freedom of the Press (RCFP).

What a physician can do

  1. Read every note before signing, and never sign automatically.
  2. Check the parts studies flag: omissions, the physical exam and who said what.
  3. Ask for consent before recording, document it, and do not let the software fill in consent language by itself.
  4. Read the vendor terms for liability caps, indemnity and who is responsible for consent.
  5. Confirm a BAA, what the vendor may do with recordings, and how long audio is kept.
  6. Ask the malpractice carrier whether it has scribe-specific guidance.

What would change this answer

The first court decision on an AI-written note; rulings in the Sharp or Sutter cases; new state laws requiring disclosure of AI scribes; new policy from the FSMB, whose AI workgroup was announced in May 2026 (FSMB), or from the AMA; and changes to vendors' published terms.

General information, not legal or medical advice. Every fact links to its source, and the page shows the date it was last reviewed.