The Third Listener: How AI Scribes Decide What Your Doctor Records

A clinician examines a sore wrist and suggests a thumb spica, the rigid splint that immobilises the base of the thumb. A moment later the clinician changes their mind. Not a spica after all; a wrist brace instead. It is the most ordinary thing in medicine: thinking aloud, revising, settling. The patient hears both versions and understands which one counts.

The AI scribe in the room heard both versions too. Its note recorded the thumb spica as applied.

That example comes from a preprint posted to arXiv on 31 August 2026 by Sebastian Fox, Luke Markham, Ryan Lail and Michael Karotsieris, titled “One note in three: a verified census of three deployed AI scribes, and the instrument that counted it”. The consultation was a test scenario, not a real patient, and the error belonged to one product in one encounter. But it captures, in miniature, the problem underneath the entire ambient scribe boom. A transcript can contain a retraction. A summary has to decide what the retraction means. Once it has decided, the reasoning disappears, and what remains is a clean, confident sentence in an official record stating that something happened which did not.

For most of medicine's history there were two people in the consulting room, and one of them wrote things down. Now there is a third listener. It hears everything, writes fast and never tires. It also decides, on terms nobody in the room set, what the conversation was about. Neither doctor nor patient can see how it made those choices, and its mistakes are written well enough to slip past a tired reader at the end of a long clinic.

So what happens to a patient when the official story of their health is quietly shaped by a party nobody can fully audit? And when the third listener decides something is not worth recording, where does that thing go, and who answers for its absence?

One Note in Three, Under a Stated Instrument

The Fox paper deserves a close reading, because its headline number is both striking and carefully hedged. The authors ran 142 consultations through three commercial AI scribes, anonymised as Scribes A, B and C. Fifty-seven were mock UK primary care consultations from the PriMock57 dataset, 45 were US ambulatory encounters from the ACI-Bench dataset, and 40 were scenarios the authors wrote. The products received replayed audio or transcripts rather than live clinical traffic. No real patient appears in the study. Between them, the products wrote 565 notes.

To find errors, the authors built what they call an instrument. Automated discovery passes proposed 13,678 candidate errors; an adversarial panel of two AI models from different families whittled these down to 618 verified findings. A physician author upheld 20 of 21 sampled findings. An independent clinician with no other involvement in the study upheld 12 of 12.

The result: 177 of the 565 notes, or 31.3 per cent (95 per cent confidence interval 27.0 to 35.6), carried at least one verified failure. Scribe A came in at 21.3 per cent, Scribe B at 40.4 per cent and Scribe C at 42.3 per cent, though the authors stress these figures are not a product ranking.

The failures clustered. There were 111 findings involving allergy status and medication lists, including a note that omitted an allergy status that had been asked about and answered in an encounter where a drug was prescribed, and another where a prescribed emollient became an oral antihistamine. There were 93 findings of invented patient identity: where a name was inaudible, notes supplied one anyway, and in one case Scribe C substituted the brand name of a chemotherapy drug. Scribe A alone dropped a stated working diagnosis 53 times. Relative timings were converted into invented calendar dates 44 times. Across ten telephone consultations, 42 findings recorded what the patient reported as though the clinician had examined it; as the paper notes, “A telephone consultation contains no physical examination.” And 11 findings came from one consultation where a retracted device, the thumb spica, was recorded as delivered care.

Read that list as a patient. Every cluster concerns something later care depends on: what you are allergic to, what you take, who you are, what your doctor thought was wrong, when it started, what was actually examined and what was actually done.

The authors are equally clear about their number's limits. Remove invented identities and dates, which a real patient record would normally prefill, and the rate falls to 24.8 per cent. More strikingly, the rate depends on the standard of review. With identical models and evidence, a strict instruction verified 9.3 per cent of candidate errors and a lenient one 79.0 per cent; depending on the standard, between 28 and 97 per cent of sampled notes carry a verified failure. Hence their warning to health systems: “A rate without its instrument cannot be compared with anyone else's.”

Two further caveats. The corresponding author lists an email address at composo.ai, an AI evaluation company; the paper's competing-interests statement discloses that the authors build evaluation tooling commercially, says no commercial product of theirs is used anywhere in the study's measurement path, and notes that the instrument it recommends is released free. And this is an unreviewed preprint. Those are good reasons not to treat 31.3 per cent as a universal constant. They are not reasons to ignore where the errors cluster.

Nineteen Notes at UC Davis

The second figure in this debate comes from real patients and has passed peer review. In 2026, JMIR Medical Informatics published a pragmatic pilot study from UC Davis Health, led by Sandra L. Taylor with co-authors including Melissa Jost, Scott MacDonald, Courtney R. Lyles and Jason Y. Adams. Over July and August 2024, 31 physicians across several specialties used an ambient AI scribe to help create 7,545 clinic notes.

The details matter, because this study is easy to misquote. Physicians did not review all 7,545 notes. They evaluated 356, or 4.7 per cent, using an instrument covering accidental inclusions, accidental omissions, hallucinations and bias, and rated each error from zero to five for its potential to harm the patient if left uncorrected. Accidental omissions were the most frequent error, with 64, followed by 41 hallucinations and 33 accidental inclusions. Most errors (119 of 142) were mild to moderate. But 19 of the 356 reviewed notes, 5.3 per cent, contained errors rated as posing serious or imminent risk.

Where you stand determines how you read that. A health system executive sees 94.7 per cent of reviewed notes free of significant errors. A patient sees roughly one note in nineteen that the reviewing doctors themselves judged capable of serious harm if nobody fixed it. The authors draw the sensible conclusion that careful clinician review remains imperative. The question this article keeps returning to is what careful review can realistically achieve.

Twenty Scribes in Ontario

Ontario's Auditor General had already run a blunter test. In a special report on public-sector AI released on 12 May 2026, Shelley Spence's office examined testing of the 20 AI scribe vendors the province had approved for doctors, run against simulated doctor-patient recordings.

According to Canadian Healthcare Technology and HR Reporter, nine of the 20 systems fabricated information, such as referrals and test orders that had not occurred. Twelve recorded a different drug from the one prescribed. Seventeen missed important details about patients' mental health that were present in the recordings. Five vendors that had not submitted required threat risk or privacy impact assessments were approved anyway.

According to reporting including The Register, officials responded that doctors must manually review AI-generated notes and that there had been no known reports of patient harm among the more than 5,000 Ontario physicians using scribes. Both claims can be true, and together they expose the core difficulty. A note missing a patient's mental health disclosure does not generate an incident report. It generates nothing. Any harm arrives later, in another consultation, with another clinician who reads a note that seems complete and has no reason to ask about what it leaves out.

A Fivefold Rise Across Canadian Primary Care

None of this has slowed adoption, and there are honest reasons why. Canada Health Infoway, the federally funded digital health agency, offered funded licences for pre-qualified scribes to primary care clinicians. On 23 June 2026 it published six-month results from an independent evaluation by the Centre for Digital Health Evaluation at Women's College Hospital Institute for Health System Solutions and Virtual Care.

The evaluation covered 12,019 clinicians enrolled between June 2025 and February 2026. Monthly AI-assisted encounters rose from just over 244,000 to more than 1.22 million, nearly fivefold. Among survey respondents, nearly 70 per cent reported less administrative burden and almost 80 per cent lower cognitive load. Clinicians saved an average of 1.11 hours a week during working hours. Same-day note completion rose from 55.9 per cent to nearly 75 per cent, and in a subset of Ontario practices, chart review found documentation completeness rising from 92.9 to 95.5 per cent.

Dr Onil Bhattacharyya, the institute's director, said: “Until now, the field has lacked evidence on what AI scribes actually do in everyday primary care practice at scale.” That is fair, and the evaluation is a genuine contribution. But note what it measured: burden, time, perception, completeness, timeliness. The release reports no accuracy findings. A note can be finished on the day, contain every expected heading and still name a drug the doctor never prescribed.

More Than a Transcription Tool

That gap, between what the tool saves and what it writes, is where the ethicists come in. In April 2026, JMIR Medical Informatics published a Viewpoint by Frank Ursin and Sabine Salloch of the Institute for Ethics, History and Philosophy of Medicine at Hannover Medical School, titled “The Ethics of AI Scribes as Epistemic Agents”.

Their core claim is simple: “AI scribes are more than transcription tools.” Transcription turns speech into text. A scribe turns a conversation into a clinical document, which means choosing what goes in, what stays out, which heading each fragment belongs under and how the whole thing reads. In the authors' summary table, the epistemic agent is one that “selects and rearranges information”.

They are careful not to overclaim. “AI scribes do not qualify as full artificial agents in a strong philosophical sense,” they write, since current systems lack intentionality. The scribe wants nothing; its agency comes from developers' design choices and institutions' deployments. But a system can shape knowledge without meaning to, and that is the concern. The abstract names the risks to physicians as “cognitive deskilling, clinical deprofessionalization, and shifts in epistemic accountability”.

Where the Chitchat Goes

Ursin and Salloch's sharpest observation concerns what vanishes. Commercial scribes typically sort conversation into a structured format, usually a version of SOAP (subjective, objective, assessment, plan), and keep what fits. The rest is discarded. “This 'chitchat,' however, is deliberately omitted in AI-generated notes, not as a bug but as a feature of the system.”

Doctors know chitchat is rarely just chitchat. A remark about a son who no longer visits, a joke about not sleeping since the redundancy, a passing mention that a husband “handles the tablets now”: these are the threads from which clinicians learn how a patient actually lives. The authors argue such talk matters for understanding patients and sustaining trust, and warn that clinicians and patients may begin steering towards what seems “recordable”.

A University of Edinburgh review published in BMJ Digital Health and AI in September 2026 reached a similar verdict from another direction. Mapping 27 articles on ambient scribe adoption, the team found the tools capture words but can miss facial expressions, gestures and emotional state, and that summaries favour clinical detail over patients' own accounts. Dr Lucas Seuren, a research fellow at Edinburgh's Centre for Biomedicine, Self and Society, warned that “the experiences of patients are poorly considered, and there are real risks that the patients' stories are lost”.

So the answer to the first question, what happens to information the scribe deems irrelevant, is uncomfortable. In most deployments it is simply not kept anywhere a later clinician will look. Audio may sit on a vendor's server for a while; transcripts may or may not be retained. But the official record, the document that follows a patient from clinic to hospital, holds only what the summary kept. Anything left out of the summary is gone from the record entirely, with nothing to show it was ever said.

Those most exposed are not random. Ursin and Salloch note that “Speech recognition and language models often perform less accurately for speakers with strong accents”, and flag the same risk for people with speech impairments. A study by Reddy and colleagues in Annals of Internal Medicine, reported by EMJ in April 2026, had 30 blinded raters score notes from 11 AI scribes and 18 human note takers across five recorded scenarios. Human notes scored higher overall; in a back pain consultation with background noise, humans averaged 43.8 out of 50 and AI 20.3. AI also scored lower in two scenarios with accented voices, though those gaps were not statistically significant.

Patients Already Changing What They Say

The warning that people will start talking for the machine is no longer hypothetical. In April 2026 YouGov polled 4,039 adults in England for Healthwatch England, the statutory patient watchdog. Almost half (48 per cent) were comfortable with an AI scribe at a routine health check. That fell to 29 per cent for sexual health, 28 per cent for mental health and 23 per cent for domestic abuse. Of the 2,795 respondents who had attended an appointment in the previous year, nearly 90 per cent did not know whether AI scribing had been used. Eighty-one per cent wanted explicit consent.

Stated discomfort is not measured silence, and the poll cannot show anyone actually withheld anything. But set it beside Ontario's finding that 17 of 20 approved scribes missed mental health details, and a troubling symmetry emerges. The topics patients least want to raise in front of a machine are the topics machines are most likely to drop. The Edinburgh review points the same way, noting evidence that patients may hold back about substance use, abuse and mental health when they know an AI system is processing the consultation. A disclosure that took courage can vanish twice: first from what the patient will say, then from what the note keeps.

Consent itself can become something the system writes. In December 2025, KPBS reported a proposed class action filed in San Diego Superior Court by a patient, Jose Saucedo, against Sharp HealthCare. The complaint alleges Sharp used Abridge's ambient AI tool to record exam-room conversations without patients' knowledge, and that charts were automatically populated with statements that patients “were advised” of the recording and “consented”, even where patients say that never happened. It estimates more than 100,000 patients may have been recorded. Sharp said it could not comment on pending litigation, and the allegations are unproven. If they hold, the record documented a conversation that never took place, in the same calm clinical prose as everything else.

Fluent Errors and the Signing Clinician

Every regulator and professional body offers the same answer: the clinician reviews and signs. In July 2026 the UK's Medicines and Healthcare products Regulatory Agency, working with NHS England, published guidance on ambient voice technology. Products “intended solely for transcription, summarising of clinical conversations”, drafting letters or suggesting codes for clinician review are not regulated as medical devices under the current framework; tools that support diagnosis or treatment, or act without clinician review, are. “Clinicians remain responsible for reviewing and verifying AI generated transcripts, summaries and other outputs,” the regulator states.

Avant, the Australian medical defence organisation, agrees. Georgie Haysom, its general manager of advocacy, education and research, wrote in January 2025: “It is your responsibility to file correct medical records, regardless of whether you use an AI scribe or not.” The same article collected errors clinicians had reported: olmesartan recorded instead of irbesartan; a family history noted as “Father happy with progress”; the oral contraceptive pill transcribed for a male patient. One attendee wrote: “Worryingly, one AI interaction made up a whole neurological exam when I didn't perform one.” Another offered a comparison many clinicians now make: “I have used medical students and RMOs as 'real intelligence' scribes and they have made mistakes like this quite often. They need to be checked, and AI is no different.”

That comparison holds only so far. A trainee's errors usually look like a trainee's: hesitant prose, uneven structure, visible gaps that invite suspicion. A scribe's errors arrive in polished, uniform, confident prose. The Fox paper puts it in one line: “The failures easiest to miss are the fluent ones, and an invented calendar date reads as diligence rather than as an error.” An exact onset date looks more careful than “about a week ago”. An examination section looks thorough. A plan with a thumb spica looks complete. Ursin and Salloch have a name for one version of this: “hallucinated compliance”, where standard sections of a note are recorded as completed when they were not. They call such fabricated entries “a distinct epistemic failure”, because they produce a record that looks complete but is not. The Avant attendee's invented neurological exam and the Fox census's telephone “examinations” are the same failure seen from two directions.

The authors admit what they did not measure: “The census counts what the products produced, not what a signing clinician would have caught.” But they are pointed about where the risk sits: “the failures concentrate where the signing clinician is the only remaining check.” Allergy status, identity and what was actually done are precisely what a busy doctor is least likely to check against memory, because they assume the system got them right.

Machines That Cannot See What Is Missing

The obvious fix is a second AI checking the first. The same group tested that in a companion preprint, also posted on 31 August 2026, titled “LLM Judges Verify Presence, Not Absence”. They built 500 note pairs each containing one error: 298 omissions and 202 additions or alterations.

Across eight judge designs, AI judges spotted added or altered content reasonably well, scoring 0.79 to 0.94 on a discrimination measure. On omissions they scored 0.50 to 0.63, where 0.5 is chance. Judging single notes, no design flagged notes with omissions more reliably than perfect ones. Detection only recovered when the task was restructured: list every fact the transcript establishes, then check the note against each.

That mirrors the human problem. Reviewing a draft is recognition: you read what is there and ask if it looks right. Catching an omission is recall: you must remember what is not there. People and machines are both far better at the first. And across this research, from UC Davis to Ontario, omission keeps surfacing as the most common or most consequential failure.

When the Editor Forgets How to Write

Ursin and Salloch's deskilling concern is about what reviewing does to the reviewer. They warn of “clinicians becoming passive recipients of AI-generated notes”, and that “If physicians become passive editors of AI-produced notes”, reflective practice and case memory may weaken. For many clinicians, writing the note is where the reasoning happens. You discover, drafting the assessment, that you are less sure than you thought.

There is real-world evidence that AI dependence can show up in measurable outcomes elsewhere in medicine. A multicentre observational study by Budzyń and colleagues in The Lancet Gastroenterology & Hepatology in August 2025 found that after AI-assisted polyp detection arrived at four Polish centres, experienced endoscopists' adenoma detection rate in standard, non-AI colonoscopies fell from 28.4 to 22.4 per cent. The design cannot prove causation, but it is a signal documentation should heed.

Medical schools are meeting the question early. In a nonrandomised trial at Yale School of Medicine led by Jaideep Talwalkar and published in JMIR Medical Education in 2026, first-year students saw AI scribe output only after writing their own notes, and the resulting hybrid notes showed no loss of quality. But Yale has also reported that in separate work, students who saw the AI output before forming their own view produced weaker assessment and plan sections, a finding Talwalkar described as a cause for concern. Write first, compare second is a deliberate defence against learning to edit before learning to think.

The Edinburgh review adds the balance: scribes may weaken memory and skill development, but could give clinicians more time for complex tasks. The Infoway clinicians reporting lower cognitive load were not imagining it. The open question is whether reclaimed attention goes into listening or into more patients per hour. For the patient, the stakes are concrete: a clinician who no longer writes the reasoning down may be less able to notice when the note's reasoning is not their own.

Who Answers for an Omission

Now the hardest question. If the scribe is an epistemic agent deciding what is clinically relevant, who is responsible when an omission shapes later care?

On paper, it is settled. The MHRA, NHS England and Avant all say the signing clinician owns the note. In the UK, the regulatory line places pure documentation tools outside medical device regulation altogether, so the vendor carries none of the safety and performance obligations device rules would impose. The developer designs the filter, the institution buys and configures it, and the clinician signs whatever emerges.

“Shifts in epistemic accountability” names what that arrangement conceals. Accountability used to follow knowledge: the clinician answered for the note because they produced it from their own perception and reasoning. Now the clinician answers for a selection made by a system whose criteria they cannot inspect, after a conversation they may only partly remember, under exactly the time pressure the scribe was bought to relieve. The signature still sits with the clinician. The decision about what to keep has moved elsewhere.

Omissions are also the hardest errors to prove. A hallucination is a sentence you can point at. An omission requires showing that something was said and then not written, which means retrieving audio or a transcript, if it still exists. In the Sharp case, the clinic reportedly told the plaintiff a recording could remain on the vendor's servers for about a month before deletion. Once the source is gone, the note is the only surviving account of the consultation, and nothing in it shows that anything is missing.

Follow that forward. A specialist reading a referral without the allergy the patient mentioned has no reason to ask again. A new GP reads a history missing the mental health disclosure and treats the visit as routine. A hospital doctor sees a thumb spica recorded as applied and plans around it. Each acts reasonably on the record. If harm follows, the chain will most likely run back to the clinician who signed the original note, who will say, truthfully, that it looked right.

What Patients Could Already Catch

Patients are not powerless, and the pre-AI baseline is instructive. In 2020 Sigall Bell and colleagues reported in JAMA Network Open on 22,889 patients at three US health systems who had read at least one of their own visit notes. About 21 per cent found a mistake, and 42.3 per cent of those judged it serious. The commonest errors involved diagnoses (27.5 per cent), medical history (23.9 per cent) and medications or allergies (14.0 per cent).

Two lessons follow. The categories patients caught in human-written notes overlap strongly with the Fox clusters, so the scribe has not invented new ways for records to fail; it may be changing how often they fail and how convincing the failures look. And patients who read their notes are a real safety check: the one participant present throughout with a stake in the details being right.

That check only works if patients can see the note, know a scribe was involved and can compare the note with what was said. Most reliably have none of the three. Nearly nine in ten recent patients polled by Healthwatch did not know whether a scribe was used, few systems offer patients the transcript, and the filtered material, the aside, the hesitation, the retracted plan, is shown to nobody.

The Case for the Scribe, Stated Fairly

It would be dishonest to cast AI scribes as simply harmful. The problem they tackle is real. Documentation burden drives burnout, and burned-out clinicians make mistakes, leave and see fewer patients. Infoway's findings on cognitive load and time come from thousands of working clinicians. Kaiser Permanente's Division of Research reported that 7,260 physicians at The Permanente Medical Group used an ambient scribe across 2,576,627 encounters between October 2023 and December 2024, saving nearly 16,000 hours of documentation; 47 per cent of 118 patients surveyed said their doctor spent less time looking at the screen, and two-thirds said they were comfortable with the technology, against 8 per cent who reported some discomfort.

Human records were never neutral either. Doctors writing late at night from memory leave plenty out, and their selections are shaped by templates, billing codes and fatigue. A scribe working from a complete recording arguably starts with better evidence than a tired clinician's recollection.

UC Davis found most reviewed notes free of significant errors, and the best product in the Fox census failed at 21.3 per cent under a demanding instrument, a figure better configuration may well push down. The honest position is not that scribes are worse than humans. It is that their failures differ in kind: more fluent, harder to detect, concentrated in specific places and shaped by criteria nobody in the room chose.

What an Auditable Third Listener Would Need

If the third listener is here to stay, and 1.22 million monthly Canadian encounters suggest it is, then the questions it raises point to concrete demands. None requires abandoning the technology. All require treating the record, rather than the clinician's time, as the thing being protected.

Keep the source and let the patient see it. A retained transcript linked to the note is the strongest safeguard against omission; without one, an omission cannot be proved. Ursin and Salloch propose consent-based narrative preservation with time-limited retention and strict limits on commercial reuse, a more sensible balance than either deleting everything or keeping everything forever.

Show what was filtered. A scribe could list what it judged irrelevant in a short “not included” section visible to the clinician and ideally the patient. The omission-blindness work explains why this helps: once a transcript's facts are listed, checking the note becomes recognition rather than recall.

Measure accuracy, not only satisfaction. Evaluations like Infoway's should report error rates under a stated method alongside time saved. Ontario shows procurement testing is feasible, and also that scribes failing on fabrication, medication and mental health can still win approval.

Design for thinking, not just editing. Ursin and Salloch recommend scribes that act as cognitive aids, prompting clinicians to confirm their own assessment rather than presenting a finished one.

Tell patients, and ask them. Healthwatch England has recommended that services tell patients at booking if a scribe will be used. Sharp shows what can follow when consent becomes something the software writes down.

And revisit the regulatory boundary. Treating pure summarisation as outside device regulation assumes clinician review is an adequate net. The evidence here questions that: fluent errors are hard to see, omissions harder, and AI judges nearly blind to absence. If review is the only safeguard, review quality deserves the scrutiny we give the tool.

The Record Is Where the Patient Lives On

A clinical note is not a memo between colleagues. It is how a patient's history outlives the appointment, travelling to specialists, hospitals, insurers, courts and future doctors who will never meet the person in the chair. For most of medicine's history, the selection behind that record was made by a human the patient could see and, in principle, question.

The third listener changes that. It hears more than any human scribe ever did and keeps less of it visible. It decides what mattered using criteria built by developers and accepted by institutions, then hands the result to a clinician who signs under pressure. When it adds something false, a careful reader may catch it. When it leaves something out, there is often nothing on the page to catch.

The thumb spica was never applied. In that test case, a researcher knew, because a researcher wrote the scenario. In a real clinic, the only people who know what was said are the clinician, who may not remember, and the patient, who will probably never read the note. Until the third listener can show its working, an AI-drafted record is best described as a well-written selection, made by a system with no stake in the outcome, that someone has signed.

References and Sources

  1. Fox, Sebastian; Markham, Luke; Lail, Ryan; Karotsieris, Michael. “One note in three: a verified census of three deployed AI scribes, and the instrument that counted it.” arXiv preprint 2608.31017v1, 31 August 2026. https://arxiv.org/abs/2608.31017 (HTML full text: https://arxiv.org/html/2608.31017). Accessed 28 October 2026.
  2. Fox, Sebastian; Markham, Luke; Lail, Ryan; Karotsieris, Michael. “LLM Judges Verify Presence, Not Absence: Omission Blindness in AI Clinical Notes and What Recovers It.” arXiv preprint 2608.31016, 31 August 2026. https://arxiv.org/abs/2608.31016 Accessed 28 October 2026.
  3. Taylor, Sandra L.; Jost, Melissa; MacDonald, Scott; Ren, Yunyi; Hilton, Shelley; Davenport, Sadie; Aizenberg, Debbie; Hall, Bruce; Lyles, Courtney R.; Adams, Jason Y. “Quality of Clinical Notes Created by Ambient Listening Generative AI: Pragmatic Prospective Pilot Study.” JMIR Medical Informatics, 14, e86474, 2026. DOI: 10.2196/86474. https://medinform.jmir.org/2026/1/e86474 (PubMed: https://pubmed.ncbi.nlm.nih.gov/41996389/; eScholarship: https://escholarship.org/uc/item/9r89g3sp). Accessed 28 October 2026.
  4. Ursin, Frank; Salloch, Sabine. “The Ethics of AI Scribes as Epistemic Agents.” JMIR Medical Informatics, 14, e88235, 30 April 2026. DOI: 10.2196/88235. https://medinform.jmir.org/2026/1/e88235 (PMC full text: https://pmc.ncbi.nlm.nih.gov/articles/PMC13132480/). Accessed 28 October 2026.
  5. Canada Health Infoway. “Canada Health Infoway Releases Results of National AI Scribe Program, Demonstrating Meaningful Benefits for Primary Care Clinicians.” News release, Toronto, 23 June 2026. https://www.infoway-inforoute.ca/en/news-events/announcements/news/2026-announcements/canada-health-infoway-releases-results-of-national-ai-scribe-program-demonstrating-meaningful-benefits-for-primary-care-clinicians Accessed 28 October 2026.
  6. The Register. “Sick and wrong: Ontario auditors find doctors' AI note takers routinely blow basic facts.” May 2026. https://www.theregister.com/a/5240771 Accessed 28 October 2026.
  7. Canadian Healthcare Technology. “Ontario AG finds flaws in AI scribes.” 13 May 2026. https://www.canhealth.com/2026/05/13/ontario-ag-finds-flaws-in-ai-scribes/ Accessed 28 October 2026.
  8. HR Reporter. “AI systems for Ontario doctors hallucinate: auditor general.” May 2026. https://www.hrreporter.com/focus-areas/automation-ai/ai-systems-for-ontario-doctors-hallucinate-auditor-general/394430 Accessed 28 October 2026.
  9. University of Edinburgh. “AI scribes may fail to capture patients' experiences.” News release, 4 September 2026 (reporting a review published in BMJ Digital Health and AI). https://www.ed.ac.uk/news/ai-scribes-may-fail-to-capture-patients-experiences (also via EurekAlert: https://www.eurekalert.org/news-releases/1142499). Accessed 28 October 2026.
  10. Medical Economics. “AI scribes fail to note patient experiences, U.K. review finds.” 4 September 2026. https://www.medicaleconomics.com/view/ai-scribes-fail-to-note-patient-experiences-u-k-review-finds Accessed 28 October 2026.
  11. Caldwell, Anna. “Doctors Make Better Clinical Notes than AI Scribes, New Study Finds.” EMJ, 20 April 2026 (reporting Reddy, A. et al., Annals of Internal Medicine, 2026, DOI: 10.7326/ANNALS-25-02772). https://www.emjreviews.com/general-healthcare/news/doctors-make-better-clinical-notes-than-ai-scribes-new-study-finds/ Accessed 28 October 2026.
  12. Healthwatch England. “What does the public think about AI scribe use in healthcare?” 16 July 2026 (YouGov polling of 4,039 adults, 16 to 27 April 2026). https://www.healthwatch.co.uk/blog/2026-07-16/what-does-public-think-about-ai-scribe-use-healthcare Accessed 28 October 2026.
  13. de Marco, Heidi. “Lawsuit claims Sharp HealthCare secretly recorded exam room conversations without patient consent.” KPBS, 11 December 2025. https://www.kpbs.org/news/health/2025/12/11/lawsuit-claims-sharp-healthcare-secretly-recorded-exam-room-conversations-without-patient-consent Accessed 28 October 2026.
  14. Becker's Hospital Review. “Patient sues Sharp HealthCare over ambient AI use.” 2025. https://www.beckershospitalreview.com/healthcare-information-technology/ai/patient-sues-sharp-healthcare-over-ambient-ai-use/ Accessed 28 October 2026.
  15. Medicines and Healthcare products Regulatory Agency. “MHRA clarifies regulatory status of ambient voice technologies used in the NHS.” GOV.UK press release, 29 July 2026. https://www.gov.uk/government/news/mhra-clarifies-regulatory-status-of-ambient-voice-technologies-used-in-the-nhs (guidance PDF: https://assets.publishing.service.gov.uk/media/6a68daaff23ff62c51c1a7a3/MHRA_Guidance_on_Ambient_Voice_Technology-Enabled_Products.pdf). Accessed 28 October 2026.
  16. LexisNexis. “MHRA publishes guidance on regulation of ambient voice technologies.” LNB News, 29 July 2026. https://www.lexisnexis.co.uk/legal/news/mhra-publishes-guidance-on-regulation-of-ambient-voice-technologies Accessed 28 October 2026.
  17. Haysom, Georgie. “AI scribes in practice: common errors to consider.” Avant, 8 January 2025. https://avant.org.au/resources/ai-scribes-in-practice-common-errors-to-consider Accessed 28 October 2026.
  18. Budzyń, K. et al. “Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study.” The Lancet Gastroenterology & Hepatology, 10, 896-903, published online 12 August 2025. DOI: 10.1016/S2468-1253(25)00133-5. https://wrap.warwick.ac.uk/id/eprint/191005 Accessed 28 October 2026.
  19. News-Medical. “AI use in colonoscopy may weaken doctors' detection skills over time.” 13 August 2025. https://www.news-medical.net/news/20250813/AI-use-in-colonoscopy-may-weaken-doctors-detection-skills-over-time.aspx Accessed 28 October 2026.
  20. Talwalkar, J.; Wright, D.; Schwamm, L.; Leydon, G.; Shabanova, V. “Impact of an Ambient AI Scribe on Medical Student Objective Structured Clinical Examination Notes: Nonrandomized Clinical Trial.” JMIR Medical Education, 12, e88264, 2026. DOI: 10.2196/88264. https://mededu.jmir.org/2026/1/e88264 Accessed 28 October 2026.
  21. Yale School of Medicine. “How Ambient AI Scribes Impact Medical Students.” 18 August 2026. https://medicine.yale.edu/news-article/ambient-ai-scribes-impact-medical-students/ Accessed 28 October 2026.
  22. JMIR Publications. “AI Scribes in Med Ed: Why We Shared Yale's Latest Insights.” JMIR blog, 2026. https://blog.jmir.org/ai-scribes-in-med-ed-why-we-shared-yales-latest-insights Accessed 28 October 2026.
  23. Bell, Sigall K. et al. “Frequency and Types of Patient-Reported Errors in Electronic Health Record Ambulatory Care Notes.” JAMA Network Open, 3(6), e205867, 2020. DOI: 10.1001/jamanetworkopen.2020.5867. https://www.opennotes.org/research/frequency-and-types-of-patient-reported-errors-in-electronic-health-record-ambulatory-care-notes/ (AHRQ PSNet summary: https://psnet.ahrq.gov/issue/frequency-and-types-patient-reported-errors-electronic-health-record-ambulatory-care-notes). Accessed 28 October 2026.
  24. Greene, Jan. “AI-assisted notetaking gains steady support from Kaiser Permanente physicians.” Kaiser Permanente Division of Research, 1 April 2025 (reporting Tierney, Aaron et al., NEJM Catalyst, DOI: 10.1056/CAT.25.0040). https://divisionofresearch.kaiserpermanente.org/ai-assisted-notetaking-gains-steady-support-from-kaiser-permanente-physicians/ Accessed 28 October 2026.
  25. Becker's Hospital Review. “16K hours saved: Ambient AI scribes at Kaiser Permanente.” 2025. https://www.beckershospitalreview.com/healthcare-information-technology/ai/16k-hours-saved-ambient-ai-scribes-at-kaiser-permanente/ Accessed 28 October 2026.

Tim Green

Tim Green UK-based Systems Theorist & Independent Technology Writer

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

ORCID: 0009-0002-0156-9795 Email: tim@smarterarticles.co.uk

Listen to the free weekly SmarterArticles Podcast

Support What We Do

Discuss...