Most accurate AI medical scribe in 2026

أبرز النقاط

• We played the same 13 recorded consultations into Hanah, Heidi, CliniScripts, Lyrebird, Freed, PatientNotes, Arlo and Preve, and graded every note, summary and transcript blind in the same packets.
• Hanah's notes captured 93.1% of the required clinical details. The average across the other seven scribes was 82.3%.
• Hanah had 4 major hallucination flags across 23 notes, summaries and letters. The other scribes had between 5 and 56.
• Hanah's transcripts had 0.85 clinically significant errors per consultation against an average of 2.17.
• Every recording, transcript, note and grade is published on our benchmark page, with the reason for each grade.

An AI scribe is only useful if you can trust what it drafts. A note that reads well and gets the drug wrong costs you more time than writing the note yourself, because you now have to find the mistake. We tested every scribe against that standard.

How the test was run

We used the 13 recordings in our accuracy corpus. They cover an ICU handover dictation with and without ward noise, physiotherapy, dietetics, occupational therapy, myotherapy, community nursing, a child with asthma whose parent speaks limited English, a quad bike fall recorded outdoors, two rugby injuries, and a jaw surgery review spoken mostly in Mandarin.

Each scribe ran in its own web app in Chrome on a standard account. The same audio file was played through the microphone for every scribe, and each recording's prompts were given to the scribe word for word, for example "Generate SOAP notes please" and "Summarise this consult in two sentences for a referral letter". Arlo has no free-text box for documents, so each prompt went into Arlo's own template generator word for word and the template it built was used unedited.

On 20 October 2026 every scribe's output was graded again together, with Arlo added. For each recording, all eight scribes' notes were shuffled behind letters with product names removed and graded by Claude, Anthropic's model, against the reference transcript:

• Each required clinical detail is marked yes, partial or no.
• Every statement the recording doesn't support is a hallucination flag. A flag is major when it could change care or misstate the clinical picture: a wrong drug, side or dose, an invented finding, diagnosis or plan, or a possibility written as fact.
• Each transcript is compared with the reference, and every difference that could mislead a clinician is counted as a clinically significant transcript error.

The full rubric is on our methodology page.

النتائج

Freed produced no transcript or note on the shortest recording, so it is scored on 12. Preve follows a custom instruction only after a sample PDF is uploaded, so it was scored on its standard clinical note, and it produced no note for its three shortest recordings.

Hanah led or tied every column. It captured 13% more of the required clinical details than the average, and its transcripts had 61% fewer clinically significant errors.

Major hallucinations across every document

A note is only part of what a scribe drafts. Clinicians also use it for referral summaries, handover notes and letters, and an invented detail in a letter leaves the clinic. This table counts major hallucination flags in every document each scribe drafted.

All four of Hanah's major flags came from the Mandarin jaw surgery review, which I cover below. Hanah had no major flags in the other 12 consultations.

What a major hallucination looks like

These are taken directly from the graded notes and summaries. Each one is on the benchmark page with the recording it came from.

Several scribes wrote the same kind of invented negative in the child with asthma consultation. The clinician asked about neck stiffness and the parent never answered, but the note records it as absent. That is a red flag written down as checked when nobody checked it.

Summaries and letters that follow the request

The second prompt for most recordings asked for something short and specific: two sentences for a referral letter, a handover note, or a letter to the treating surgeon.

Most of Arlo's summaries were written to the patient, opening "You came in today", and ran to four or more sentences. Several added advice the clinician never gave: rest, ice, avoiding activities and when to seek medical attention.

Transcripts on difficult audio

On short consultations in a quiet room, most scribes transcribe within a point or two of each other. The difference shows up when the audio is hard.

The figures are word error rates against the verbatim reference. We count every word the same, fillers included, so a lower number means fewer words to correct. The per-consultation rates for every scribe are on the benchmark page.

The Mandarin consultation

The jaw surgery review was spoken mostly in Mandarin, with English clinical terms. Heidi made the fewest clinically significant transcript errors on it (2), then Hanah (5) and CliniScripts (6). Arlo and Preve each made 18, and neither has an input language setting.

Hanah heard "chlorhexidine" as "Hexadene" and "pus" as "pass", and both errors carried into the note, the plan and the letter to the surgeon. Hanah still captured 81% of the clinical details in this consultation, against an average of 53% for the other seven scribes.

What sets Hanah apart

Hanah drafts what was said. All four of Hanah's major flags were words misheard in one recording. None was an exam finding nobody performed, a negative nobody confirmed, a diagnosis nobody made or advice nobody gave, and every other scribe in this test wrote at least one of those. I explain why we hold to this in accuracy over inference.

Hanah follows the request. When you ask for two sentences for a referral letter, you get two sentences with the chief complaint and the plan in them.

Hanah transcribes difficult audio accurately. Background noise, a parent speaking a second language, a 50 minute session or a recording made outdoors all change how much a clinician has to correct. Hanah's error rate stayed low on all of them.

Results also vary between runs of the same recording. We re-tested Hanah on the short knee pain consultation three times: two runs heard "I hurt it" as "I heard it" and one heard it correctly, and every one of Hanah's notes recorded the injury correctly. We publish the run we used and the variation, and our repeated-run analysis found Hanah varied least of the six scribes it compared.

تحقّقوا بأنفسكم

Every recording, reference transcript, note, summary and grade is on the benchmark page, with the reason each grade was given. You can play the audio and read each scribe's output next to what was actually said. Hanah drafts and the clinician reviews, and that review is quicker when there is less to fix.

الأسئلة الشائعة

Which AI scribes were tested?

Hanah, Heidi, CliniScripts, Lyrebird, Freed, PatientNotes, Arlo and Preve, on the same 13 recordings.

Who graded the notes?

Claude, Anthropic's model, graded every scribe's output with product names removed and the outputs shuffled. All eight scribes were graded together on 20 October 2026 with the same rubric.

What counts as a major hallucination flag?

A statement the recording doesn't support that could change care or misstate the clinical picture, such as a wrong drug, dose or side, an invented finding, diagnosis or plan, or a possibility written as fact.

Does Hanah make mistakes?

Yes. Hanah had four major flags in this test, all in the Mandarin consultation, and they are published with the rest. Every draft should be reviewed by the clinician before it is filed.