Skip to content
Physician Case Review Consult

When an AI Scribe Gets the Note Wrong

Ambient AI scribes draft notes from recorded visits. When a draft adds, drops, or misattributes a finding, the evidence sits in records most productions miss.

Consider a hypothetical. A patient tells an emergency physician that she felt pressure in her chest the day before. An ambient AI scribe, which listens to the visit and drafts the note, records “denies chest pain.” The physician reviews the draft between patients, signs it, and discharges her with a diagnosis of reflux. She returns two days later with a heart attack. The signed note says she denied the symptom that matters most.

Ambient documentation is spreading quickly because it addresses a real problem: clinicians spend a large part of their day writing notes. The tools work well most of the time. But when one gets a detail wrong, the error lands in the legal record of care, signed by a clinician, and it can look exactly like a clinician’s own mistake. Telling the two apart requires evidence that a standard record production does not include.

Three ways a draft goes wrong. An AI scribe can add something that was never said, such as a finding, a medication, or a negative response. It can leave out something that was said, such as a symptom, a worry, or a plan the clinician spoke aloud. Or it can get a real fact attached to the wrong thing: the wrong side of the body, the wrong date, or what a family member said recorded as the patient’s own history. Omissions are the hardest to catch, because nothing on the page looks wrong. A reviewing clinician has to remember what is missing.

The signature still matters. In most organizations the clinician who signs the note is responsible for its content, and AI policies usually say so. A signed note with an error is not excused because a tool drafted it. But the signature is the beginning of the inquiry, not the end. The question for a case is how the tool, the workflow, and the clinician’s review together produced the record, and whether the error changed what happened to the patient.

What evidence may exist. These records are worth asking about specifically, because they are not part of the printed chart:

  • The recording and transcript. Retention varies by product and by contract. Some configurations keep the audio for a period of time; others discard it once the draft is created. If it exists, it is the closest thing to a record of what was actually said, and it may need prompt preservation.
  • The AI draft and the signed version. Note version history can show what the tool wrote, what the clinician changed, and what was left as drafted.
  • Review timing. The audit trail can show when the draft was opened, how long it was open, and when it was signed. A short review time is a fact to be explained, not proof that no review happened.
  • The tool’s version and configuration in use on the date of the visit, including any specialty templates.
  • Governance records. The organization’s AI policy, the approval of the tool, training given to clinicians on reviewing drafts, and any monitoring of draft accuracy.

What the guidance expects. The Joint Commission and the Coalition for Health AI (CHAI) issued voluntary guidance in 2025 on the responsible use of AI in healthcare. It calls for a governance structure and policies, training on the proper use and limits of each tool, and telling patients when AI directly affects their care, where appropriate. It also calls for monitoring that is scaled to risk, and it suggests that tools supporting documentation can be checked less often than tools that drive clinical decisions.

That last point deserves attention in a case like this one. A documentation tool sounds low risk, but the note it drafts becomes the information the next clinician relies on. CHAI’s risk categorization approach makes a similar point in a different way: risk depends on how directly a tool reaches the patient, how much a person reviews its output, and how serious an uncaught error could be. An ambient scribe whose drafts are routinely signed with light review sits closer to the patient than its label suggests.

What to expect from the defense. The strongest plaintiff evidence here is a recording that contradicts the note and an AI error that went uncorrected. Expect the defense to answer that the clinician’s actual assessment was sound and the note error did not change the treatment decision, or that the organization had reasonable review training and monitoring in place. A theory survives that answer only if the evidence connects the error to a decision. Neither the existence of an AI error nor the absence of governance documents establishes negligence on its own. Voluntary guidance is not a legal standard of care. What the evidence can do is show what was said, what was written, who changed what, and when.

For counsel, the practical step is early. Ask whether an ambient scribe was used for the encounters at issue. If it was, ask what the vendor and the hospital retained, and send a preservation request before routine deletion schedules run. The difference between a documentation error and a clinical error is often provable only from records that exist for a limited time.

Sources

  • Joint Commission and Coalition for Health AI. The Responsible Use of AI in Healthcare (RUAIH). 2025. jointcommission.org
  • Coalition for Health AI. Risk Categorization Tool, version 3. chai.org

The scenario above is hypothetical. Disclosure: Dr. Bansal is Chief Medical Officer of Diagnose Data, Inc., a healthcare technology company. This article does not describe or endorse any product.

This article is educational and does not constitute legal advice. Safety classifications and voluntary guidance describe reporting and improvement frameworks, not a legal standard of care or any outcome. Sources are listed at the end of the article.