Who is legally responsible if an AI scribe puts something wrong in my note?

The signing physician. The AMA's position is that physicians retain full responsibility for anything they sign off on, including AI-generated documentation [10]. Vendor indemnification governs the commercial relationship between your organization and the vendor, not your professional liability. Have counsel read the actual language.

Do I have to tell patients I am using an AI scribe?

Practically, yes, and increasingly as a legal matter. Class actions filed in 2026 against Sutter Health, MemorialCare entities, and a scribe vendor allege recording without adequate consent under California's all-party consent statutes and the federal Wiretap Act [14]. Malpractice carriers now recommend adding AI scribe language to consent forms and documenting the discussion [13].

Do AI scribes actually reduce burnout, or only shift the work?

The burnout signal is more consistent than the time signal. A randomized crossover trial of 160 clinicians found both tested products reduced burnout scores while leaving pajama time largely unchanged [6], and a three-arm randomized trial found improvements in Mini-Z and work exhaustion scores even when time-in-note did not move [3]. Feeling present in the room appears to be worth something independent of minutes saved.

Are AI scribes regulated by the FDA?

Generally not. Most are positioned as administrative or HIPAA-eligible services rather than medical devices, which keeps them outside formal FDA review despite their direct impact on the medical record [9]. The Joint Commission and the Coalition for Health AI issued voluntary guidance in September 2025 recommending local validation and ongoing monitoring [12], but that is guidance, not clearance.

Which patients are highest risk for an inaccurate AI note?

atients with accents, limited English proficiency, aphasia, dysarthria, dementia, or significant dyspnea, plus any encounter with multiple speakers such as family members or interpreters [8][9]. Speech recognition error rates have been documented as higher for Black speakers specifically [9]. These are exactly the encounters that warrant a full read rather than a scan.

Back to blogs

Are AI Medical Scribes Safe? What Doctors Must Know

Audits found omissions in 18% of AI scribe notes and hallucinations in 11.5%. Here is what physicians should verify before signing.

Reviewed by O. Daniel Odutola, MD, MBA, MPHNovember 5, 202615 min read

Are AI medical scribes safe? Mostly yes, with a real asterisk. The best published audit of ambient AI notes found that about 95 percent were free of significant errors, but roughly 1 in 20 contained something a physician judged capable of causing serious harm if it went uncorrected [1]. That is a good tool with a bad tail, and the entire question of AI scribe safety comes down to whether you are the person catching that tail.

Most of us have had the 9 p.m. version of this conversation with ourselves. Twenty-two notes in the queue, dinner cold, and a draft note sitting there that reads beautifully. It has a tidy HPI. It has a plan. Signing it takes one click. Reading it takes four minutes you do not have.

That gap, between how finished the note looks and how verified it actually is, is where the safety problem lives.

The honest answer on whether AI medical scribes are safe

Ambient scribes are not a new category of danger. They are an old category of danger with better prose.

Physicians have been signing notes they skimmed for as long as there have been templates and copy-forward. What changed is that the draft now arrives fluent, confident, and clinically plausible, which makes skimming feel safer than it is. A garbled human transcription announces itself. A hallucinated physical exam does not.

The University of California, Davis ran the cleanest look at this. Over two months in 2024, 31 physicians generated 7,545 ambient AI notes and then formally evaluated a random sample of 356 against four error types [1]. What they found:

Error type in AI-generated notes	Notes affected (of 356)	Rate
Accidental omission (something said, not documented)	64	18.0%
Hallucination (documented, never said)	41	11.5%
Accidental inclusion (irrelevant content pulled in)	33	9.3%
Bias	4	1.1%
Errors rated serious or imminent risk (severity 4 to 5)	19	5.3%
Notes free of significant errors	337	94.7%

Source: Taylor et al., JMIR Medical Informatics, 2026 [1].

Read the first row again. Omission was almost twice as common as hallucination. Everyone worries about the AI inventing a lung exam. The more frequent failure is the AI quietly dropping the part where the patient mentioned the chest pressure while you were washing your hands.

The same study tracked how much physicians actually edited. The median physician changed 9 percent of the AI-generated words. Individual practice ranged from 1.9 percent to 69.3 percent, and 14.9 percent of notes were signed with no edits at all [1]. Those numbers describe the same technology being used two completely different ways.

What the evidence says about time saved, and what it does not say

The efficiency case is real but smaller than the marketing.

Study	Design and scale	Headline finding
Rotenstein et al., JAMA, 2026 [2]	8,581 ambulatory clinicians, 1,809 adopters, 5 academic health systems	13.4 fewer minutes of total EHR time and 16.0 fewer minutes of documentation per 8 scheduled patient hours; 0.49 additional visits per week
Lukac et al., NEJM AI, 2025 [3]	Randomized trial, 238 outpatient physicians, 14 specialties, 2 vendors vs usual care	One product cut time-in-note 9.5% (P=0.02); the other showed no significant change (P=0.66)
Tierney et al., NEJM Catalyst, 2025 [4]	2,576,627 encounters, 7,260 physicians, 63 weeks at The Permanente Medical Group	Roughly 15,700 documentation hours saved; pajama time fell 1.03 minutes per day (P=0.02)
Preiksaitis et al., JMIR AI, 2026 [5]	10,344 ED encounters, 100 attendings, 13-month rollout	72.6 fewer seconds of on-shift documentation per encounter; after-shift time rose 9.1 seconds
Chowdhury et al., JAMIA, 2026 [6]	Randomized crossover, 160 outpatient clinicians	Both tools reduced burnout scores; pajama time largely unaffected

Notice the pattern. Burnout and satisfaction move reliably. Wall-clock time moves modestly, and in the UCLA randomized trial one of the two commercial products did not beat usual care on time-in-note at all [3]. Anyone quoting you a single dramatic number is quoting one study.

The subjective benefit is not nothing. In the Permanente rollout, 84 percent of surveyed physicians said the scribe improved their visit interactions [4]. Looking at a patient instead of a keyboard changes how the visit feels, and that matters. It just is not the same claim as "safe."

See how your offer compares

Compare real pay data before you sign anything

Physicians' Copilot Clinical AI

The four ways an ambient scribe gets it wrong

Understanding ambient AI documentation risks means understanding that these tools fail in specific, predictable ways rather than at random.

Omission. The most common failure. Content that was discussed does not make it into the note. This is not unique to AI; a 2022 study of home health nursing found roughly half of verbally discussed patient problems never reached the chart [7]. But an AI omission is more dangerous because the note looks complete. No blank field reminds you something is missing.

Hallucination. Fabricated content. The classic example is a documented physical exam maneuver that never happened. Research on the underlying speech-to-text layer helps explain why: in one analysis of a widely used transcription model, about 1 percent of audio segments produced entire invented phrases, and 38 percent of those hallucinations carried explicit harms such as fabricated associations or false authority [8]. For a deeper treatment of why generative models invent content and how to catch it, see our piece on AI hallucinations in medicine (/clinical/ai-hallucinations-medicine).

Misattribution. The tool cannot reliably tell who is talking. A daughter's report of the fall becomes the patient's own account. A student's differential becomes the attending's assessment.

Bias and unequal accuracy. This one gets the least attention and deserves more. Speech recognition systems have documented higher error rates for Black speakers, and accuracy degrades with non-standard accents and limited English proficiency [9]. The same analysis noted that hallucinations cluster around longer pauses in speech, which means patients with aphasia, dysarthria, dementia, or simple breathlessness generate worse notes [8]. Your sickest and least fluent patients get your least accurate documentation. That is exactly backwards from what anyone wants.

Overall published error rates for modern ambient scribes sit in the range of 1 to 3 percent [9]. Multiply that by your annual encounter volume and decide whether the number feels small.

Physicians' Copilot keeps you in the loop. Every output is traceable to a source you can check in seconds, because a tool that cannot show its work is asking you to sign for it blind.

Who is liable when the AI scribe is wrong

You are. That is the entire answer, and it has not been seriously contested.

The AMA's guidance is blunt about it: physicians should read the note, should not assume the draft is right, and carry "100% of the responsibility" for what they sign [10]. Bill Satterwhite, a physician and attorney at Huron Consulting Group, put it the way I now use with residents: think of AI scribing as a copilot, not the pilot, and definitely not autopilot [11].

Three practical points get missed.

First, your signature is an attestation. Signing an AI draft is legally identical to signing a note you dictated yourself. There is no partial credit for "the software wrote it."

Second, the audio and the transcript may exist. Vendor retention policies vary, and a stored recording that contradicts your signed note is a plaintiff's exhibit, not a defense. Ask your vendor, in writing, how long audio and intermediate transcripts are kept and who can access them.

Third, accuracy drifts. Jennifer Geetter of McDermott Will & Schulte described AI performance degradation as "like a knife that gets dull" [11]. A tool validated at go-live is not permanently validated. Model updates, vendor changes, and new note templates all shift behavior, which is why the Joint Commission and the Coalition for Health AI published guidance in September 2025 calling for ongoing local validation and monitoring rather than one-time approval [12].

Malpractice carriers have started publishing their own expectations. The Texas Medical Liability Trust recommends amending documentation policies to require review, building dedicated time to read AI drafts rather than batch-signing, correcting post-signature errors by addendum with the reviewer's name and date, and confirming a business associate agreement with any third-party scribe vendor [13].

This article is educational and is not legal advice. Contract terms vary by state and employer. Have any agreement reviewed by a licensed attorney before signing.

In 2026, patients filed class actions against Sutter Health, MemorialCare entities, and the vendor Abridge AI, alleging that visits were recorded without adequate consent. The claims were brought under the California Invasion of Privacy Act, the federal Wiretap Act, and the California Confidentiality of Medical Information Act [14].

The core allegation is not that AI scribes are illegal. It is that patients did not get clear notice their conversation would be recorded, sent outside the exam room, and processed by a third party. In all-party consent states, a general privacy notice buried in registration paperwork may not be enough [14].

At the bedside this is simple. Tell the patient, in one sentence, before you start: "I use an AI assistant that listens and drafts my note, so I can pay attention to you instead of the screen. Is that okay?" Document that you asked. If they decline, turn it off and type.

Roughly 30 percent of physician practices were using AI scribes as of late 2025 [15]. Very few of those conversations are happening consistently.

A 90-second review protocol before you sign

You will not re-read every note against your memory of the visit. Nobody does. What works is a targeted scan of the fields where errors are both likely and consequential.

1. Check the medication list against what you actually decided. Dose changes and discontinuations are the highest-yield errors. Read the numbers, not the drug names.

2. Check allergies and adverse reactions. These carry forward into every downstream decision and get transcribed wrong more than you would expect.

3. Verify every physical exam finding you did not perform. If the note describes a maneuver you skipped, delete it. This is the single most common serious hallucination.

4. Confirm who said what. Scan for family or interpreter statements attributed to the patient, and for your own thinking attributed to a trainee or vice versa.

5. Hunt for what is missing, not only what is wrong. Ask yourself one question: what was the most important thing the patient said today, and is it in this note? Omission was the most frequent error in the UC Davis audit [1].

6. Read the assessment and plan word for word. Everything else is context. This part is the medicolegal spine of the note and the part a reviewer will read first.

7. Raise your scrutiny for high-risk encounters. Interpreter-mediated visits, patients with speech impairment, crowded rooms, telephone encounters, and any visit where you seriously considered admission deserve a full read, not a scan.

If you are new to ambient tools, run steps 1 through 6 on every note for your first two weeks. You will learn your product's failure pattern faster than any vendor training will teach you, and after that the scan takes about 90 seconds.

What to ask before your group signs a vendor contract

If you sit on a committee, these are the questions that separate a serious vendor from a demo:

• What is your measured hallucination and omission rate, on what sample size, and audited by whom?

• Can you provide accuracy data disaggregated by patient race, primary language, and accent? [9]

• How long is audio retained, how long are intermediate transcripts retained, and who can access them?

• Is our data used to train your models, and can we opt out in the contract?

• What happens to note quality when you push a model update, and will you notify us before you do?

• What does your indemnification actually cover if a documentation error contributes to a claim?

• Will you support a local validation audit at go-live and at defined intervals afterward? [12]

A vendor who answers the first two questions with a case study instead of a number has told you something.

Where this is heading

Ambient scribes are not regulated as medical devices. They are generally marketed as administrative or HIPAA-eligible services, which keeps them outside formal FDA evaluation despite their direct effect on the clinical record [9]. Researchers at Columbia and the University of Eastern Finland argue that adoption has outrun validation, transparency, and oversight, and it is hard to read the deployment numbers and disagree [9][15]. That gap will close eventually. Until it does, the validation layer is you.

Worth knowing what you are actually buying, too. An ambient scribe transcribes and drafts; a clinical copilot reasons and suggests. The safety questions are related but not identical, which we unpack in ambient AI scribes versus clinical copilots (/clinical/ambient-scribe-vs-copilot). For the broader picture, start with our practical guide to AI in clinical medicine (/clinical/ai-in-clinical-medicine).

The 2026 AMA survey found that 81 percent of physicians now use AI professionally, up from 40 percent in 2023, and that clear liability frameworks ranked highest among the regulatory actions physicians said would build their trust [16]. Translation: we adopted first and are asking who holds the bag second. Not a great order of operations, but it is the one we are in.

No AI tool replaces clinical judgment. Verify all outputs against primary sources and your own assessment.

Ready to see what doctor-in-the-loop actually looks like? Learn how Physicians' Copilot keeps you in the loop.

Key takeaways

• About 95 percent of ambient AI notes in the largest published safety audit were free of significant errors, but 5.3 percent contained an error rated capable of serious or imminent harm if uncorrected [1].

• Omission is more common than hallucination. In that audit, 18 percent of notes dropped something discussed, versus 11.5 percent that invented something [1].

• Time savings are real but modest: roughly 13 to 16 fewer minutes per 8 scheduled patient hours across five academic health systems, and one of two products in a randomized trial showed no significant benefit [2][3].

• Accuracy is not evenly distributed. Error rates run higher for Black speakers, non-standard accents, limited English proficiency, and impaired speech [8][9].

• Your signature is your attestation. Liability does not transfer to the vendor, and consent litigation under state wiretapping statutes is already active [10][14].

REFERENCES

1. Taylor SL, Jost M, MacDonald S, Ren Y, Hilton S, Davenport S, Aizenberg D, Hall B, Lyles CR, Adams JY. "Quality of Clinical Notes Created by Ambient Listening Generative AI: Pragmatic Prospective Pilot Study." JMIR Medical Informatics, April 17, 2026. https://doi.org/10.2196/86474 (accessed September 16, 2026). Retrieved via PubMed.

2. Rotenstein L, et al. Multisite study of ambient AI documentation across five academic health systems. JAMA, April 1, 2026. DOI 10.1001/jama.2026.2253. Findings summarized in: "Large AI scribe study finds modest time savings, inconsistent use," STAT News, April 1, 2026. https://www.statnews.com/2026/04/01/ai-ambient-scribes-modest-time-savings-clinical-documentation/ and HIT Consultant, April 1, 2026. https://hitconsultant.net/2026/04/01/jama-ai-scribe-study-ehr-time-savings-burnout-reality-check/ (accessed September 16, 2026).

3. Lukac PJ, Turner W, Vangala S, Chin AT, Khalili J, Shih YT, Sarkisian C, Cheng EM, Mafi JN. "Ambient AI Scribes in Clinical Practice: A Randomized Trial." NEJM AI, published online November 26, 2025; 2(12). https://doi.org/10.1056/aioa2501000 (accessed September 16, 2026). Retrieved via PubMed.

4. Tierney AA, Gayre G, Hoberman B, Mattern B, Ballesca M, Wilson Hannay SB, Castilla K, Lau CS, Kipnis P, Liu V, Lee K. "Ambient Artificial Intelligence Scribes: Learnings after 1 Year and over 2.5 Million Uses." NEJM Catalyst Innovations in Care Delivery, May 2025; 6(5). https://doi.org/10.1056/CAT.25.0040 (accessed September 16, 2026).

5. Preiksaitis C, Alvarez A, Winkel M, Karamatsu M, Brown I, Sama N, Morris L, Lee JY, Gubbels A, Wahl E, Frye A, Schoeffler A, Gharahbaghian L, Rose C. "Ambient AI Scribes and Emergency Department Documentation Burden: Retrospective Cohort Study." JMIR AI, July 2, 2026; 5:e92193. https://doi.org/10.2196/92193 (accessed September 16, 2026). Retrieved via PubMed.

6. Chowdhury A, Casey M, Wilson J, Pollak KI, Goldstein BA, Bedoya A, Poon EG. "Comparing ambient scribes: a randomized crossover clinical trial addressing ambient scribe technologies' impact on physician burnout." Journal of the American Medical Informatics Association, May 1, 2026; 33(5):990-999. https://doi.org/10.1093/jamia/ocag018 (accessed September 16, 2026). Retrieved via PubMed.

7. Song J, et al. "Do nurses document all discussions of patient problems and nursing interventions in the electronic health record? A pilot study in home healthcare." JAMIA Open, 2022; 5:ooac034. Cited in reference 9. (accessed September 16, 2026).

8. Koenecke A, Choi ASG, Mei KX, Schellmann H, Sloane M. "Careless Whisper: Speech-to-Text Hallucination Harms." Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT '24), June 2024. https://doi.org/10.1145/3630106.3658996 (accessed September 16, 2026).

9. Topaz M, Peltonen LM, Zhang Z. "Beyond human ears: navigating the uncharted risks of AI scribes in clinical practice." npj Digital Medicine, September 24, 2025. https://www.nature.com/articles/s41746-025-01895-6 (accessed September 16, 2026).

10. Berg S. "Using health AI in the exam room: What doctors should consider." American Medical Association, October 9, 2025. https://www.ama-assn.org/practice-management/digital-health/using-health-ai-exam-room-what-doctors-should-consider (accessed September 16, 2026).

11. Brady M. "Why AI scribes are a malpractice risk, according to experts." Healthcare Dive, September 3, 2026. https://www.healthcaredive.com/news/why-ai-scribes-malpractice-risk/829541/ (accessed September 16, 2026).

12. "Joint Commission and Coalition for Health AI (CHAI) Release Initial Guidance to Support Responsible AI Adoption Across U.S. Health Systems." Guidance on the Responsible Use of AI in Healthcare (RUAIH). The Joint Commission and CHAI, September 17, 2025. https://www.chai.org/news/joint-commission-and-coalition-for-health-ai-chai-release-initial-guidance (accessed September 16, 2026).

13. "Using AI medical scribes: Risk management considerations." Texas Medical Liability Trust (TMLT), July 30, 2025. https://www.tmlt.org/resource/using-ai-medical-scribes-risk-management-considerations (accessed September 16, 2026).

14. Felz D, Everett J, Pike J, Pullen S. "Your AI Scribe May Be Taking Notes (and Plaintiffs Are Too)." Alston & Bird Privacy, Cyber & Data Strategy Blog, April 27, 2026. https://www.alstonprivacy.com/your-ai-scribe-may-be-taking-notes-and-plaintiffs-are-too/ (accessed September 16, 2026).

15. "Health Care's Rush to AI Scribes Risks Patient Safety, Researchers Warn." Columbia University School of Nursing, October 8, 2025. https://www.nursing.columbia.edu/news/health-cares-rush-ai-scribes-risks-patient-safety-researchers-warn (accessed September 16, 2026).

16. "More than 80% of physicians use AI professionally: AMA survey." 2026 Physician Survey on Augmented Intelligence, American Medical Association, March 12, 2026. https://www.ama-assn.org/practice-management/digital-health/more-80-physicians-use-ai-professionally-ama-survey (accessed September 16, 2026).

FAQs

No. Published error rates for ambient scribes cluster around 1 to 3 percent overall, and a formal audit of 356 notes found 5.3 percent contained errors judged capable of serious harm if left uncorrected [1][9]. That rate is low enough to be useful and high enough that unreviewed signing is not defensible.