Ambient AI physiotherapy scribes are the fastest adopted technology in clinical practice since the electronic health record.The problems built into it are arriving at the same speed.
There is a pattern in healthcare technology that repeats itself.
A new tool arrives. It promises to solve a real and pressing problem. It is adopted quickly — enthusiastically, at scale — before the full implications of its design are understood. And then, gradually, the problems that were always built into it begin to surface. Not hypothetical problems. Structural ones.
Ambient AI documentation is in that phase right now.
What it promises
The pitch is straightforward. A device listens to the clinical appointment as it happens. It transcribes the conversation. It generates a clinical note automatically. The clinician reviews and signs it. The documentation burden — the notes written under pressure, the records reconstructed from memory — is solved.
It is a compelling pitch. The problem it is solving is real. Documentation burden is one of the most well-documented sources of clinician burnout in the published literature. Any tool that addresses it deserves serious consideration.
The question is not whether the problem is real. The question is whether this is the right solution to it.
The consent moment
Before the recording starts — before the ambient AI has captured a single word — something has to happen.
The patient has to consent.
Not in an abstract, form-based way. In the room. Face to face. At the start of the appointment, before the clinical conversation has begun, the clinician has to explain that the session will be recorded by an AI system, that the audio will be transmitted to a third-party vendor, that it will be processed and returned as a draft clinical note.
And then they have to ask the patient whether that is acceptable.
Published research from a JAMA Network Open quality improvement study found that patient comfort with ambient documentation varied based on trust in their clinician, understanding of the tool, and perceived benefits or risks, and that the informed consent conversation is a critical but nuanced touchpoint that shapes the adoption and acceptance of technology by both clinicians and patients.
In other words — consent is not a simple tick. It is a conversation – adding to the admin tasks associated with the Practice Trap. It is a consent conversation that happens at the moment the patient is least prepared for and the clinician is most pressed for time on.
The clinician who has used ambient AI for more than a few weeks knows what this moment actually looks like in practice. The explanation. The pause. The patient who nods because declining feels awkward, not because they genuinely understand what they are agreeing to. The appointment that has already consumed two minutes before the clinical work has begun.
This is not a minor friction. It is a structural addition to every appointment.
The hallucination problem
Once consent is obtained and the recording begins, a different problem emerges.
Ambient AI does not transcribe what was said. It generates a note based on what it interpreted from the audio. These are not the same thing.
Peer reviewed studies show that up to 31% of AI notes contain at least one hallucination — the AI inserts findings, details, or clinical content that was never actually said. On complex multi-problem presentations, on visits with ambiguous audio, on cases requiring precise clinical language — the hallucination rate increases.
AI-generated documentation can hallucinate findings, mis-attribute statements, or drop context, so a clinician must read and correct the draft before it becomes part of the legal record.
The clinical note that contains a finding the clinician never made — and that the clinician signs without catching — is not a documentation tool working as intended. It is a new category of clinical risk. Malpractice carriers are now actively flagging AI-documentation hallucinations as an emerging liability. If an AI scribe inserts or alters clinical content and the clinician approves it without proper review, the malpractice exposure sits entirely with the clinician. It is not a hypothetical. It is already happening.
Editing is not eliminating
The standard response to the hallucination concern is that the clinician reviews the draft before signing it. Which is true. And which is precisely the problem.
A draft note that requires careful review and correction is not a documentation solution. It is a documentation relocation. The work has moved from writing to editing — and on complex cases, the editing takes longer than the writing would have.
Studies report frequent documentation omissions and occasional clinically significant hallucinations, and implementation remains a sociotechnical challenge involving workflow redesign and medico-legal considerations.
The time saved on routine appointments is real. The time spent on complex ones — reviewing, correcting, verifying that the AI did not insert a finding that wasn’t there — is equally real. The net saving, across a full clinical week, is considerably less dramatic than the marketing suggests.
Ambient AI did not eliminate the documentation burden. It relocated it. Typing became editing. The cognitive load of producing an accurate clinical record did not disappear — it shifted to a different point in the workflow, with a different kind of attention required and a different category of risk attached to getting it wrong.
The legal landscape
Beyond the clinical risks, the legal exposure attached to ambient AI is significant and growing.
Recent lawsuits in California and Illinois allege that health systems used ambient scribing without obtaining informed consent from patients, potentially violating state wiretapping statutes and confidentiality protections when audio is transmitted to third-party vendors for processing.
The most damaging allegation in one class action — with a proposed class of over 100,000 patients — was that the electronic health record contained boilerplate language stating patients had been advised and consented to recording when, according to the complaint, no such consent was obtained. The statutory damages in that case run to $5,000 per violation.
This is not an edge case. It is the foreseeable consequence of deploying a recording technology in a clinical setting where the consent process is conducted verbally, face to face, under time pressure, at the start of every appointment.
The reliability question
The clinical record is a legal document. It is the contemporaneous evidence of what happened in the appointment, what the clinician observed, what they decided, and why.
A clinical record that depends on a system that loses recordings, crashes mid-workflow, changes its output behaviour after unannounced backend updates, or generates findings that were never made — is not a clinical record the clinician can stand behind with confidence.
Real world user feedback from clinicians using ambient AI systems consistently identifies lost transcripts, app instability, and unannounced degradation of template behaviour as ongoing problems. Not isolated incidents. Recurring structural weaknesses in systems that were adopted before they were fully stable.
In any other professional context these would be product failures. In a clinical context, where the note is a legal record and the consequences of an inaccurate one can be severe, they are something more serious.
The problem that was never solved
Ambient AI addressed one dimension of clinical documentation — the note written after the appointment from memory or from shorthand. It left the rest of the Practice Trap entirely untouched.
The history still gathered in the room. The examination still conducted under time pressure. The clinical thinking still divided between the patient and the administrative layer surrounding them.
The appointment itself — the moment that actually matters — was never the thing being optimised. What ambient AI optimised was what happened after it.
A tool that records the appointment and turns it into a note is solving for the end of the process. The problem begins at the start.
HAve you read the Practice Trap series? Find it here https://vocaform.ai/the-practice-trap/
