A University of Edinburgh study has found that ambient AI scribes now used by around 40 per cent of UK GPs can miss vital information and risk “cognitive offloading” that erodes clinical memory and reasoning.

Edinburgh Study Finds AI Scribes Could Miss Vital Clinical Information

Ambient AI scribes have had a remarkably smooth run through the NHS adoption curve. Roughly 40 per cent of GPs already use the tools, NHS England has backed them as a flagship productivity measure, and the Medicines and Healthcare products Regulatory Agency has just told the sector precisely where the regulatory line sits for the technology. 

What The Edinburgh Study Actually Found

Academics at the university’s centre for biomedicine, self and society reviewed cases involving ambient scribes and identified genuine benefits alongside serious gaps. The tools free clinicians from manual notetaking, allowing more focus on complex tasks during a consultation. But the researchers also found instances where clinicians did not recognise their own AI-generated notes, and where written summaries prioritised clinical detail at the expense of a patient’s own account of their illness and circumstances. Facial expressions, gestures and emotional state, the texture that experienced clinicians rely on to read a room, do not translate into a transcript.

More troubling for a care sector built on disclosure and trust, the study found patients may be less forthcoming about sensitive matters, including substance use, domestic abuse or mental health struggles, once they know a conversation is being recorded and processed by AI. Dr Lucas Seuren, of the University of Edinburgh’s centre for biomedicine, self and society, said the experiences of patients had been poorly considered against the excitement clinicians feel about reduced paperwork, and warned that “patients’ stories are lost” in ways that could further disadvantage people already facing marginalisation in health and social care services.

Cognitive Offloading And The Deskilling Risk

The study’s more structural finding concerns what researchers term cognitive offloading. Manual notetaking is not simply administrative overhead, it serves a genuine function in how clinicians reason through and reflect on what a patient has told them. Remove that function and, the researchers suggest, memory recall and skill development can suffer alongside it. That risk sits uncomfortably close to home for a workforce already stretched thin, where the promise of AI has consistently been framed as giving time back to staff, not eroding the clinical judgement those staff are meant to exercise.

None of this is presented as a case against ambient scribes outright. The researchers were careful to note that clinicians can use freed-up time for more meaningful conversations and complex decision-making, precisely the trade-off NHS England has been banking on. But the study is explicit that design matters enormously, and that systems built primarily to serve documentation efficiency risk overwriting the patient’s own voice rather than preserving it.

A Widening Gap Between Evidence And Enthusiasm

The timing is instructive. Only weeks after the MHRA drew its regulatory line around ambient voice technology, and shortly after NHS England confirmed how £10 billion in technology funding would flow toward AI triage and ambient notetaking, this study lands as the first serious independent counterweight to what has, until now, been an overwhelmingly promotional evidence base. Trust-level results cited elsewhere, 47 minutes saved per shift at one London hospital, nearly a quarter more clinician time freed at Great Ormond Street, have understandably driven the enthusiasm. The Edinburgh findings do not contradict those productivity figures. They complicate what “productivity” means when the thing being optimised away is the clinician’s own memory of the patient in front of them.

That tension is unlikely to slow investment. The Treasury’s newly opened £100 million Sovereign AI procurement scheme names clinical decision-support and workflow automation among its priority challenges, and ambient documentation sits squarely inside that ambition. The study is not likely to reverse that direction of travel, but it does hand commissioners a genuine evidence base for asking harder questions before their next AVT contract is signed.

Implications For Domiciliary And Community Care

For registered care providers, the study’s findings translate directly. Electronic care planning tools increasingly draw on the same natural language processing capabilities as clinical AI scribes, and the same disclosure risk applies, arguably more acutely, in a domiciliary or supported living setting where a care worker is often the first person a service user tells about abuse, neglect or a deteriorating mental state. Under CQC inspection frameworks and the Care Act 2014, providers are expected to understand precisely what a technology captures, what it omits and how it shapes the written record a subsequent professional will rely on. This study gives that expectation sharper teeth. A tool that reliably transcribes medication changes but consistently underweights a person’s fear, confusion or reluctance to engage is not a neutral efficiency gain. It is a documentation system with a bias, and providers evaluating AVT products now have peer-reviewed grounds to ask suppliers exactly how that bias is being addressed.

The researchers call for further work on the medium and long-term use of these tools in specific settings, and for particular caution where technology built for one country’s healthcare system is deployed in another. For a UK care sector still working out how ambient documentation fits alongside human judgement, that caution is worth heeding before adoption outpaces the evidence a second time.