Recent headlines about ambient voice technology (AVT) missing nuances in consultations and mis-transcribing drug names and diagnoses have put the safety and potential bias of ambient scribes in the spotlight. But I believe most of the debate is answering the wrong question. Errors are inevitable in any system. What matters is whether they fall evenly across patients or land disproportionately on some. Specifically, whether AI scribes narrow the inequities already built into NHS care, or quietly widen them.
Bias in clinical documentation did not arrive with AI. Decades of evidence show variation in pain assessment, diagnostic thresholds and treatment intensity by patient ethnicity, gender and accent. One study found many US medical students and residents held false beliefs about biological differences between Black and White patients, beliefs that shaped how they said they'd manage pain. In UK psychiatry, Black men are over-represented in schizophrenia diagnoses and under-represented in diagnoses of affective disorders. AVT is built on records shaped by exactly this history. The right question was never ‘is this system unbiased'. It's whether AVT ends up at least as equitable as the pathway it replaces, ideally more so, with visible safeguards where risk remains.
The evidence splits across AVT's two layers: automatic speech recognition, which converts speech to text, and the large language model, which turns that transcript into a structured note. Both carry demonstrated bias. Word error rates for African American speakers have been shown to run roughly double those for White speakers across major commercial systems, a gap newer models, including Whisper, have narrowed but not closed. Whisper has also been shown to hallucinate entire invented sentences, more often for speakers with longer pauses, a pattern more common among people with aphasia and other speech differences. Controlled studies show language models generating different diagnoses, workups and treatment intensity for identical clinical presentations, depending only on a patient's stated race. In both cases, models are reproducing the differential treatment already in their historical training data.
But none of this shows the scale of the problem in the NHS. There's no paired study checking whether an ambient scribe describes the same consultation differently depending on a patient's ethnicity, language or age, a gap that matters given 5.1m people in England and Wales don't speak English as their main language. Interpreter-mediated consultations, disproportionately used by patients from minority ethnic groups, sit almost entirely outside the two-speaker design most AVT systems were built around. We've set out these gaps in more detail in a report with Tandem Health through our AVT Academy. The whole sector should be racing to close them openly, not sitting on them for competitive advantage.
None of this is a reason to pull back. It's a reason to be precise about where responsibility for safe roll out sits. The Commission into the Regulation of AI in Healthcare's recent report draws that line: suppliers must spell out ‘the operational conditions needed for safe deployment and use', contracts should state which controls each side delivers and clinicians must not become ‘liability sinks' for errors the system was built to catch. In practice, that means programme leads tracking editing and correction rates by ethnicity, language and disability from go-live, and sharing those patterns with suppliers. Products should surface risk rather than hide it, with an audit trail that lets patterns be investigated and every note kept a draft until reviewed. Clinicians must still read drafts critically, but stretched staff shouldn't carry the weight of catching what the contract should already assign.
NHS England's AVT Supplier Registry sets a floor. But the MHRA has been clear since July that a scribe which only transcribes and summarises is not a medical device, so the equity expectations the Commission wants for AI devices won't automatically reach most of what AVT does. Today, that assurance falls to the deploying trust. I expect, and welcome, suppliers publishing stratified benchmarks, such as error rates by accent and evidence on whether identical consultations produce differently worded notes. Suppliers, including Accurx, should be judged on whether we do this.
The NHS has strong equity infrastructure, including the Race and Health Observatory and the Equality Act's public sector duty. What's missing isn't the will, but asking suppliers the right questions and holding them to their answers.
What I want, and what I think most people working in this space would want, is straightforward: every NHS patient, whoever they are, wherever they're from and however they speak, should get the same standard of documentation and care. Ambient scribes should help to close the gaps that already exist in the NHS, rather than quietly writing them into the record for good.
