AI scribe accuracy: where errors come from and how to catch them
An AI scribe turns speech into text and then text into a note. Errors can enter at either step, and a few kinds of error carry most of the clinical risk. Knowing where they come from makes your review faster and more targeted, and it helps you judge any tool fairly.
Two steps, two kinds of error
The first step is speech to text. Mistakes here come from the room: background noise, people talking over each other, a soft voice at the far side of the desk, strong accents, unusual drug names and numbers said quickly. The transcript may hear 15 as 50, or miss a short word like no.
The second step is text to note. A language model reads the transcript and writes a structured note. Here the risks are different: it can leave out something that was said, attribute a statement to the wrong speaker, summarise too hard, or fill a gap with something plausible that nobody said. Good tools are built to avoid invention, but no tool removes the risk entirely.
The errors that matter most
- Negations. A denied symptom written as present, or the reverse.
- Laterality. Left and right swapped in an examination or a plan.
- Numbers and units. Doses, frequencies, blood pressure, laboratory values, durations.
- Who said it. A relative's worry written as the patient's symptom, or the clinician's question written as history.
- Examination findings. A finding written as normal when it was never examined.
- The plan. Tests or referrals listed that were only mentioned as possibilities.
A note can read fluently and still contain one of these. Fluency is not a sign of accuracy, so check these six areas on every note, even when the rest looks right.
Habits that reduce errors
- Control the room. Close the door, turn off the television in the waiting area if you can hear it, and put the device between you and the patient rather than behind a monitor.
- Say the examination out loud. A scribe cannot see. A short spoken summary such as 'chest clear, heart sounds normal, no ankle swelling' gives it something real to write.
- Speak numbers clearly. Say doses and values with their units, and repeat anything important once.
- Close with a spoken plan. Summarising the plan to the patient at the end helps them and gives the scribe a clean source for the plan section.
- Review while the visit is fresh. Edits take seconds straight after the visit and much longer at the end of the day, when memory has faded.
Measuring accuracy in your own practice
Published accuracy figures come from settings that may not match yours. A better test is your own. Take twenty consecutive notes and, as you review each one, tally the edits by type: wording you simply prefer, missing facts, wrong facts and invented facts. Wording preferences cost time; wrong or invented facts are safety issues.
If wrong facts cluster in one place, such as numbers or a particular language, change the habit that feeds it, or switch the tool for those visits. Repeat the tally after a few weeks. This takes little effort and tells you more than any marketing page.
How ClearPass approaches accuracy
ClearPass separates speakers in the transcript and writes the note from what was said. It then runs a numbers check: any number in the note that never appeared in the transcript is marked for your review, because invented or misheard numbers are among the most dangerous errors. You can choose a Concise or Detailed note, and use the Ask box to correct or rewrite a section in plain words.
Quality varies by language and accent, as it does with every scribe, so test the visits you actually see during the 7-day trial. Every note is a draft for you to review and correct. See how to review an AI note for a fast review order.
What ClearPass does not do
- It does not replace your judgement. Every note is an AI draft that you review and correct before it is used.
- It is not a medical device and does not make clinical decisions. It is a documentation aid.
- It does not connect to your EMR. The note is copy-ready text that pastes into any system.
Privacy
Clinical note content is not stored on ClearPass servers: it is processed to write the note, then discarded. Consultation audio is transcribed and discarded, never stored. Your note history stays in your browser, on your device, and you choose how long it is kept. Common identifiers are removed from typed and transcribed text before it reaches the AI; that is a best-effort safeguard, so leave names, ID numbers and file numbers out where you can. Processing goes through the AI providers named in our Privacy Policy, which sets out exactly what each one receives and how long it keeps it. ClearPass does not sign HIPAA business associate agreements today.
Questions
Are AI scribes accurate enough to use without review?
No. Every AI note is a draft. The clinician who signs it is responsible for its content, so read it before it is used.
Do accents and mixed languages lower accuracy?
Often, yes, and by different amounts for different languages. Test your own patient mix during a trial instead of relying on general claims.
What is the single most useful check?
Numbers and negations. Check every dose, value and duration, and make sure each symptom is written as present or absent the way the patient described it.
Can a scribe add things that were never said?
It can. Well-built tools are designed to avoid it, but it still happens, which is why reviewing the examination and plan sections matters.
What does ClearPass cost?
The 7-day trial is free, needs no card and simply ends after 7 days, with no automatic charge. After that there is one plan: $78 a month or $858 a year, with every visit transcribed.
Try it on your next clinic day
The full product for 7 days, with no card and no automatic charge. Then $78 a month or $858 a year. Pricing details.
Start free — 7 days, no card