Can You Trust AI Scribe Notes? What the Research Actually Says About Accuracy and Hallucinations

AI medical scribes work. That's not in dispute — peer-reviewed studies consistently show they reduce documentation time by 20-30% and cut after-hours charting meaningfully. But "works" and "can be trusted blindly" are different claims, and the gap between them is where clinicians need to pay attention.

The uncomfortable truth is that AI scribes make errors that human documentation usually doesn't — including a category of error unique to generative AI called hallucination, where the system produces confident, fluent, professionally-worded text that is simply false. In most industries, a 1-3% error rate is excellent. In clinical documentation, where a single fabricated medication or misattributed symptom can affect patient care, billing, and liability, that same rate demands a different relationship with the tool.

This article looks at what the actual research says about AI scribe accuracy — not vendor marketing, not fear-mongering, but peer-reviewed studies and documented incidents — and lays out a practical review workflow that lets you capture the time savings without inheriting the risk.

What the research actually shows

Let's start with the numbers, because they're more nuanced than either AI optimists or skeptics usually admit.

A 2025 review published in npj Digital Medicine (part of the Nature portfolio) synthesized the evidence on ambient AI scribes. The key findings:

• Modern ambient AI scribes using large language models report overall error rates of approximately 1-3%.
• For comparison, automated speech-recognition dictation systems (the previous generation of documentation tech) have error rates of 7-11%, owing to the complexity of medical jargon and accent variability.
• Human medical scribes, in randomized trials, are more than four times as likely to produce notes physicians rate as "accurate" compared with standard physician self-documentation.

So the headline is genuinely good: modern AI scribes are more accurate than the dictation tools many practices used before, and they free up significant time. The npj Digital Medicine review noted that across a quality improvement study of 45 clinicians from 17 specialties, ambient AI scribes reduced documentation time by a median of 2.6 minutes per appointment and cut after-hours EHR work by 29.3%.

But the same body of research is clear that the 1-3% error rate isn't the whole story, because AI errors are categorically different from the errors clinicians are used to catching.

The failure modes that matter

The reason a 1-3% error rate deserves more attention in healthcare than in other fields is the type of errors AI scribes make. Research analyzing real-world clinician feedback has identified several distinct categories of patient safety concern:

Hallucinations. The AI generates information that was never said. The most cited real-world example comes from a clinician study where an internist discovered the AI had listed atorvastatin as a current medication for a patient who had never been on it. The danger here is specific: the fabricated entry is fluent, plausible, and indistinguishable from a real entry on a quick read.

Incorrect medication names and dosages. Medical terms that sound alike — "Lipitor" vs "Lisinopril," for instance — can be transcribed or generated incorrectly. In a medication list, a single wrong drug or dose is a genuine safety event.

Misattribution between patient and clinician. The AI assigns a statement to the wrong speaker — recording the patient as having said something the clinician said, or vice versa. In behavioral health and complex histories, this can materially distort the record.

Critical omissions. Perhaps the most underappreciated failure mode. One analysis found that approximately 50% of patient problems discussed aloud were never documented even by human clinicians — and AI scribes with unclear filtering rules may make this worse, dropping clinically important details the model judged less relevant.

Contextual misinterpretation. The AI captures the words but misunderstands the clinical meaning — documenting a hypothetical ("if your chest pain returns, go to the ER") as a current finding, or missing negation ("no shortness of breath" becoming "shortness of breath").

The through-line is that these errors don't look like errors. A traditional software glitch produces obviously broken output. An AI hallucination produces output that reads like a competent clinician wrote it. That's exactly what makes it dangerous, and exactly why the review step matters more than vendors sometimes emphasize.

A real regulatory incident

This isn't purely theoretical. In May 2026, Ontario's auditor general documented problems with AI transcription tools used in the province's medical system. The report found the system was inadequately evaluated, with accuracy issues and hallucinations occurring during both testing and actual use — including fabricated information, incorrect drug data, and missing mental health details.

The Ontario case is instructive because the failures were exactly the failure modes the research predicts: fabrication, incorrect medication data, and critical omissions. It's also instructive because of the mitigating factor the report noted — medical practitioners review the notes before decisions are made. That review step is what stands between an AI error and actual patient harm. It's not optional. It's the core safety mechanism of the entire workflow.

Who carries the liability

Here's the part that should focus every clinician's attention: under current regulatory frameworks, the clinician — not the AI vendor — is generally liable for errors in AI-generated documentation that they sign.

AI scribes generate draft notes. The clinician reviews and approves them. Once you sign a note, it's your clinical documentation, regardless of which tool drafted it. If a hallucinated medication entry leads to harm and you signed the note, the liability framework treats that essentially the same as if you'd typed it yourself.

This is why the regulatory gap matters. The technology has outpaced the oversight — there's no comprehensive FDA framework specifically validating ambient AI scribes for accuracy the way there is for many medical devices. Vendors are not generally required to publish accuracy data or demographic performance reports. That leaves clinicians in the position of relying on tools whose error rates they can't independently verify, while bearing the legal responsibility for the output.

The practical implication: treat every AI-generated note as a draft from an extremely fast but occasionally unreliable junior colleague. You wouldn't sign a resident's note without reading it. The same standard applies here, and arguably more so.

The demographic accuracy problem

One finding from the research deserves separate attention because it's easy to miss: AI scribe accuracy is not uniform across patient populations.

Transcription accuracy can vary with accent, dialect, speech patterns, and language. A scribe trained predominantly on certain demographic speech patterns may perform measurably worse for patients with accents, non-native English speakers, elderly patients with speech changes, or patients with speech impairments. The npj Digital Medicine research and related studies specifically flag the need for equitable AI scribe performance and demographic performance reporting — data most vendors don't currently publish.

For practices serving linguistically diverse populations, this is a real consideration. It's also an argument for tools with strong multilingual support (Heidi Health, Nabla Copilot) if your patient population is diverse, and for extra review diligence on notes from encounters where transcription was more likely to struggle.

A practical review workflow

None of this means you should avoid AI scribes. The time savings are real and well-documented, and the tools are more accurate than the dictation systems they replace. It means you need a review workflow that catches the specific errors AI scribes make. Here's a practical one.

1. Review immediately, while the encounter is fresh. The single most effective safeguard is reviewing the note right after the visit, when you still remember what was actually said. A hallucinated medication is obvious if you remember the conversation; it's invisible three days later. Don't batch your note review for the end of the day or week.

2. Check the medication list against your memory and the chart, every time. Medications are the highest-risk category for AI errors with the most direct patient safety implications. Verify every medication name and dose in the AI-generated note against what you know to be true. This is the non-negotiable step.

3. Scan for negation errors. Look specifically for places where the note might have flipped a negative to a positive or vice versa. "Denies chest pain" becoming "reports chest pain" is a classic AI error. Pay attention to the review of systems and any symptom documentation.

4. Verify speaker attribution in complex encounters. In behavioral health, family medicine with multiple people in the room, or any encounter with back-and-forth, confirm that statements are attributed to the right person.

5. Check for omissions of anything clinically important. AI scribes sometimes drop details they judge less relevant. If you discussed something that matters — a medication change you're considering, a red-flag symptom you're monitoring, a referral you promised — confirm it made it into the note.

6. Confirm the plan is accurate and complete. The assessment and plan is the most clinically consequential part of the note. Read it carefully. This is where contextual misinterpretation does the most damage.

7. Never sign without reading the whole note. This sounds obvious, but the entire risk of AI scribes comes from the temptation to trust fluent output without verification. The note reads well, so you sign it. That's the trap. The fluency is exactly what you can't trust.

This workflow adds 1-2 minutes per note. Given that AI scribes save far more than that, you still come out dramatically ahead on time — but with the safety mechanism intact.

Questions to ask your AI scribe vendor

Beyond your own review workflow, push your vendor for transparency. Good questions:

• What is your documented error rate, and how was it measured? Vendors who can answer specifically are more trustworthy than those who deflect.
• Do you publish demographic performance data? Accuracy across accents, languages, and age groups matters.
• How do you handle negation and uncertainty? Ask specifically how the system documents "no" findings and hypotheticals.
• What's your hallucination mitigation approach? Some vendors ground notes more tightly in the transcript than others.
• Do you retain the source audio, and for how long? Audio retention (offered by tools like OrbDoc with up to 7-year retention) provides an evidence trail if a note's accuracy is ever questioned.
• Is the model trained on my clinical data? This is a privacy question, but also a quality question — and it should be answered in the BAA.

Choosing tools with accuracy in mind

Different tools make different tradeoffs that affect accuracy:

• Tools with evidence-linking (like OrbDoc, which links every statement to the audio moment it came from) make verification faster — you can check the source rather than rely on memory.
• Tools with live transcript panels (like Nabla Copilot, which shows AI edits in real time) give you more oversight during generation.
• Specialty-tuned tools generally produce fewer contextual errors within their specialty than generalized tools, because they understand the terminology and expected structure. A behavioral health tool (Mentalyc, Upheal) is less likely to misinterpret therapy content than a general medical scribe.
• Tools with strong multilingual support (Heidi Health, Nabla) reduce the demographic accuracy gap for diverse patient populations.

None of these eliminate the need for review. But they can make review faster and reduce the error rate you're reviewing against.

The honest bottom line

AI scribes are a genuine advance. The research supports their use: they save meaningful time, reduce burnout, and are more accurate than the dictation tools many practices relied on before. The clinicians using them well are getting real benefit.

But "use them well" is doing important work in that sentence. The 1-3% error rate is excellent for most applications and not good enough to trust blindly in clinical documentation. The errors AI scribes make — hallucinations, medication mistakes, misattributions, omissions — are precisely the errors that fluent, confident output hides best. The clinician who signs AI-generated notes without reading them carefully is taking on liability for errors they didn't catch and can't blame on the tool.

The solution isn't to avoid AI scribes. It's to use them with a clear-eyed understanding of what they do well (drafting fast, structuring notes, capturing the bulk of an encounter) and what they don't (guaranteeing accuracy, understanding clinical nuance, catching their own mistakes). Pair the tool's speed with your judgment, build a review workflow that targets the specific failure modes, and you get the time savings without the risk.

The technology is a powerful junior colleague, not an infallible one. Treat it accordingly.

For a directory of AI scribes with details on their accuracy features, audio retention, and specialty fit, see our full directory. For tools designed for specific specialties — which generally produce fewer contextual errors within their domain — see our specialty pages. For more on the compliance side of AI tools, see our guide to what HIPAA compliance actually means for AI tools.

This article is informational only and does not constitute medical, legal, or compliance advice. It summarizes published research and documented incidents as of mid-2026. Clinicians remain responsible for reviewing and approving all AI-generated documentation. Consult qualified counsel and your professional liability carrier for guidance specific to your practice.