A patient in San Diego discovers, months after a routine appointment, that his consultation was recorded by his doctor's phone, sent to a cloud server run by an artificial intelligence vendor, and turned into a clinical note he never agreed to. The medical record says he consented. He did not. That single discrepancy, buried in a patient portal, is now the basis of a federal lawsuit against Sharp HealthCare, one of several filed this year in California against health systems using Abridge's ambient documentation tools. The claims do not challenge the technology itself. They challenge the fiction that consent was obtained when it was not, and they land the argument for the first time on individual clinical conduct rather than institutional procurement.

This distinction matters more to the NHS than it might first appear. Ambient voice technology, the software that listens to a consultation and drafts the note so a clinician does not have to type while talking, has been moving through English hospitals and general practice for two years, sold on the promise of giving time back to overstretched staff. NHS trusts have piloted it under various governance frameworks, and national bodies have issued guidance on data handling and clinical safety. What has not yet been tested, in any serious way, is what happens when a patient asserts that a clinician switched on a listening tool without proper explanation, or leaned on an AI-generated summary without checking it against what was actually said. The American cases suggest that the gap will not stay theoretical for long.

The Medical Protection Society has already named the underlying problem, describing doctors as at risk of becoming a liability sink for AI errors they had little hand in designing. Its argument, that the Consumer Protection Act 1987 was drafted for kettles and cars and struggles to accommodate software, is not abstract lawyering. It reflects a real asymmetry: vendors write the tools, trusts procure them, and clinicians remain the only party a negligence claim can straightforwardly reach. The Law Commission is reviewing product liability with this in mind, and the UK Jurisdiction Taskforce has been preparing a legal statement on AI harms. NHS Resolution, which handles negligence claims across English health services, is drafting its own guidance on how liability should be apportioned when an algorithm is involved. None of this activity has produced a settled answer. All of it signals that the question of who is accountable when an AI tool goes wrong, or when consent is assumed rather than secured, has moved from seminar rooms into the offices that will eventually have to defend or discipline someone.

This lands at an awkward moment for NHS leadership. Sir Jim Mackey's push to hold trusts and individual leaders to tighter account for operational performance has reinforced a culture in which responsibility is meant to sit with identifiable people rather than diffuse systems. AI adoption pulls in the opposite direction: it is procured centrally, trained on external data, and updated by a company the clinician has never met, yet its output ends up in a record the clinician signs. Regulators have so far framed AI as an assistive tool that leaves the doctor fully responsible for the decision, which is a defensible position when the tool suggests a diagnosis but a much harder one to sustain when the tool has silently recorded and interpreted an entire conversation.

The lesson from California is not that ambient AI has no place in British medicine. It is that consent cannot be treated as a box ticked once in a data protection policy and forgotten. Every NHS trust rolling out a listening tool needs a consent process a patient would recognise as real, and every clinician using one needs to know, precisely, where their liability starts and where a vendor's obligations should begin. The first complaint of this kind in England is not a matter of if. The institutions best placed to answer it are the ones building the answer now, before a patient in Leeds or Leicester finds the same discrepancy in their own notes.