AI Scribes in the NHS: What Patients Actually Want and Why Governance Matters More Than Technology

When Hopkins Van Mil published its public dialogue findings on AI scribes in the NHS in May 2026, the headline that circulated was reassuring: patients are broadly open to the use of AI scribes if certain conditions are met. That framing is accurate but incomplete. The detail of what those conditions are, and what the dialogue revealed about the current state of AI deployment in NHS clinical settings, tells a more uncomfortable story — one that has significant implications for anyone building or procuring AI tools for healthcare.

The dialogue, commissioned by Newton's Tree and conducted with 41 participants broadly representative of the UK population, is the most substantial piece of public engagement on AI scribes in NHS settings conducted to date. It is worth reading carefully rather than summarising optimistically.

The first finding: patients did not know

Before the dialogue began, approximately two thirds of participants had never heard of AI scribes. More significantly, very few were aware that AI scribes were already in use in NHS settings they had recently attended. They were not consulted before deployment. They were not informed during appointments. They found out, during a research dialogue, that they had already been recorded.

Their reaction was not primarily alarm. It was frustration at the absence of transparency. The concern was not that AI was being used but that decisions had been made without public awareness or engagement, and that by the time patients were being asked their views, the direction of travel had already been set.

This matters for the governance debate in a specific way. The current deployment model for AI scribes in NHS settings operates on implied consent — an assumption that patients can opt out if they choose, without any systematic effort to ensure they know the option exists. The Hopkins Van Mil participants were clear that this is not acceptable. Most supported an opt-out rather than an opt-in model, but only if accompanied by a genuine national awareness campaign. Implied consent without awareness is not consent at all. It is, as one participant put it, a concept that should not exist.

The second finding: patients are not convinced there is a problem to solve

One of the most striking themes in the dialogue is the difficulty participants had in identifying what problem AI scribes were addressing. Most reported that their current experience of clinical note-taking was satisfactory. They did not identify the GP typing during an appointment as a significant barrier to good care. They were not experiencing the administrative burden that AI scribes are positioned to relieve.

This does not mean patients are wrong or that the administrative burden is not real. The pressure on clinical staff from documentation requirements is well-documented and significant. But it does mean that the primary beneficiary of AI scribes, as participants perceived it, was the health system and its workforce rather than the patient in front of the clinician. Several asked directly: who is this for?

The implication is that the case for AI scribes needs to be made in patient-facing terms, not just efficiency terms. If the benefit to patients is indirect — less stressed clinicians, shorter waiting times, more time available for care — then that argument needs to be made explicitly and evidenced. Asserting that AI will improve the NHS without demonstrating how it will improve the patient's experience is not persuasive to a public that has watched multiple NHS IT programmes fail to deliver.

The third finding: the efficiency calculation is questioned

Participants were not convinced that AI scribes would actually save time in practice. Their reasoning was straightforward: if the AI-generated record has to be reviewed and corrected by the clinician before it enters the patient record, the time saved on transcription is partly consumed by the time spent on verification.

This concern is directly supported by the Brighton and Sussex Medical School study published in the same period, which found that while AI marking of physician associate MSc submissions was fifteen times faster than manual marking, almost half the files required human amendment before release, and four contained clinically important errors. The efficiency gain on one side of the ledger was partially handed back as a moderation cost on the other.

Participants in the Hopkins Van Mil dialogue reached the same conclusion independently. If doctors have to spend time reviewing and correcting transcripts, and if that review time is genuinely engaged rather than a rubber-stamp, the net efficiency gain may be smaller than the headline figures suggest. Several participants noted that accuracy checks sufficient to catch clinically significant errors would necessarily reduce the time savings that justified the technology.

The honest conclusion is that AI scribes are efficient if the error rate is low enough that verification is quick. They are less efficient than claimed if the error rate requires substantive review. Measuring the actual moderation burden — not the time saved on transcription, but the total time including correction — is the evidence base that currently does not exist at scale.

The fourth finding: clinicians must remain responsible

There was no ambiguity among participants on the question of accountability. The clinician using the AI scribe is responsible for the accuracy of the patient record, regardless of whether that record was generated by AI. The AI company is not responsible. The platform is not responsible. The human who signs off the record is responsible.

This is not simply a preference. It is the condition under which patients are willing to accept AI-generated clinical records at all. They want a human to be accountable because they understand that accountability requires the capacity to say sorry, to correct errors, to be held to consequences. AI systems cannot do any of these things. The humans using them can.

The implication for AI system design is significant. Any platform that obscures human responsibility — by making it difficult to identify who reviewed and approved an AI-generated record, or by creating workflows in which verification is nominally present but practically bypassed — is undermining the trust condition that makes AI-assisted clinical documentation acceptable to patients.

The Nursing and Midwifery Council's approach in professional revalidation reflects the same principle. The NMC does not validate AI-generated reflective accounts. It validates accounts that a registered professional takes responsibility for, regardless of how they were produced. The professional's signature is not a formality. It is the accountability mechanism.

The fifth finding: the account layer is ungoverned

Perhaps the most important finding in the Hopkins Van Mil report, and the one least visible in the summary findings, is the set of questions participants asked that nobody could answer. What exactly is in the AI-generated record of my appointment? Can I see it? Can I correct it? What happens if it is wrong? Where does it go? Who has access to it? How long is it kept? If I have already had appointments since AI scribes were deployed, what records exist of those appointments and what do they say about me?

These are not unreasonable questions. Under UK GDPR, patients have the right to access their personal data, the right to rectification of inaccurate data, and the right to understand how their data is processed. The difficulty is that the current deployment of AI scribes in NHS settings does not systematically provide the infrastructure for patients to exercise these rights in relation to AI-generated content specifically.

The account — the AI-produced summary of a clinical consultation — is where the patient's interests are most directly at stake. It is also the layer at which governance is currently weakest. The data security questions — where is it stored, is it in the UK, who can access it — are being addressed, imperfectly but actively. The account integrity questions — does it accurately represent what happened, can it be contested, how does a correction propagate — are not.

This is the problem that Stephen Hall at Digital Narrative Care identifies at the public services level through the Assurance DIAL framework. At Dial 2, the source integrity question: does the account accurately represent its source? In the AI scribes context, the source is the clinical consultation. The account is the AI-generated summary. The gap between them, and the absence of any systematic mechanism for patients to identify and correct that gap, is the ungoverned space that the Hopkins Van Mil participants were circling throughout their discussions without having a framework to name it.

What patients actually want

The participants' recommendations, synthesised into a patient charter at the end of the dialogue, can be stated clearly. A national awareness campaign before any further deployment expansion, so that implied consent operates in an environment of genuine awareness rather than ignorance. An opt-out process that follows the patient across NHS settings so they do not have to repeat their preference at every encounter. Clinicians remaining responsible for record accuracy, with patients having the right to view and dispute AI-generated records. National oversight of AI scribes with independent auditing. Data remaining in the UK. AI scribes trained on diverse voices, accents, and dialects before deployment rather than after.

None of these conditions are technically demanding. They are governance conditions. They require decisions about accountability, transparency, and oversight rather than advances in the underlying technology.

The implications for clinical documentation AI

The Hopkins Van Mil findings have implications beyond AI scribes specifically. They describe the conditions under which patients are willing to accept AI involvement in clinical documentation of any kind — and those conditions apply equally to AI used in clinical education, professional development, and revalidation documentation.

The conditions can be summarised as four principles. Transparency: patients and professionals must know when AI is involved in producing a clinical record and what it contributed. Accountability: a named human professional must take responsibility for the accuracy of any AI-assisted record. Contestability: there must be a mechanism to identify, dispute, and correct errors in AI-generated accounts. Proportionality: AI involvement must be appropriate to the clinical context and should not proceed where patient needs or professional judgement indicate otherwise.

These four principles describe the governance architecture that responsible AI deployment in clinical settings requires. They are not derived from regulatory theory. They are derived from what 41 members of the UK public said they needed in order to trust AI in their healthcare. That is a more robust foundation than any compliance framework.

For organisations building AI tools for clinical settings, the Hopkins Van Mil report is not a market research document. It is a governance specification written by the people whose trust you need to earn. The technology question — can AI produce accurate clinical records — is being answered, imperfectly and with ongoing caveats. The governance question — who is responsible, what can be seen, what can be corrected, and where does the record travel — is the one that will determine whether patients accept AI in their clinical care or resist it.

The answer is not better AI. It is better governance of the accounts that AI produces.

This article is published in accordance with PAIDS™ (Professional AI Documentation Standards) — well-sourced, thoroughly researched, and defensible with verifiable data only.