---
title: "AI scribe accuracy benchmark: WER and hallucinations — Hanah"
description: "Hanah's published AI scribe accuracy benchmark, graded per region for Word Error Rate and hallucinations: 1.6% Australia, 5.9% UK."
source: https://hanah.health/transparency/
---

# AI scribe accuracy & hallucination benchmark

1.6% Australia

graded 30 July 2026

5.9% United Kingdom

graded 29 June 2026

Word Error Rate against verbatim ground truth, corpus v1. Each region runs its own speech and language stack and is graded on its own run, months apart and on a slightly different set of recordings, so the two numbers describe two stacks rather than a trend.

Every Hanah region is graded by the same harness, the same ground-truth corpus, and the same published methodology: **Word Error Rate** against verbatim transcripts, plus an **LLM jury panel** grading coverage and hallucination. Each region runs its own speech and language stack, so each is graded — and published — separately. The headline numbers below render from the exact data each regional report publishes; open a region for the per-recording transcripts, generated documents, and per-juror verdicts behind them.

Headline numbers are rendered client-side from each region's published harness aggregate — the same file that drives the full report, so this page can never disagree with it. A hallucination is only counted when at least two independent jurors agree.

Looking for security controls, subprocessors and data residency instead? Those are documented per region in the [Trust Center](https://hanah.health/trust/).