analysis of texts · Journal of evaluation in clinical practice · la publicación, 1 ago 2026 · gratis
Un médico firma, una máquina escribe: quién responde por lo que dice su expediente
En foros de profesionales, los clínicos describen notas generadas por inteligencia artificial que suenan correctas pero contienen errores. Es un estudio de conversaciones, no de leyes: no dice qué exige su país.
Versión breve · la versión detallada sigue, unos 7 min
Pregunte a weeklyAI
Pregúnteme por este estudio: a quiénes se estudió, qué encontró y qué no dice.
Las conversaciones se guardan mientras exista weeklyAI, para mejorar la publicación. Se responde en el idioma en que usted escribe.
- El estudio, de un vistazo
- Quiénes
- Comentarios de profesionales de la salud en foros de Reddit
- Cuántos
- 484 comentarios de 120 hilos
- Dónde
- Ocho comunidades de profesionales de salud en inglés (Estados Unidos)
- Cuándo
- De octubre de 2020 a febrero de 2026
- Tipo de estudio
- análisis de conversaciones en línea
- Quién lo hizo
- Investigadores de Aston University y University of Wolverhampton
- El límite que importa
- Son comentarios de foros, no leyes ni sentencias: no dicen qué exige su país
Los pacientes leen las notas y detectan errores con frecuencia, pero carecen por completo de rutas prácticas para corregirlos.

En Estados Unidos, los pacientes pueden leer las notas clínicas de sus visitas. Ese derecho se ha ampliado por normas federales sobre acceso a la información. Y en esos mismos expedientes, según relatan los propios clínicos, cada vez hay más texto escrito por programas de inteligencia artificial que escuchan la consulta y redactan la nota.
Un grupo de investigadores estudió cómo los clínicos hablan de esto entre ellos. Usaron un programa informático para extraer conversaciones de Reddit y analizaron 484 comentarios de 120 hilos en ocho comunidades de profesionales de la salud, entre octubre de 2020 y febrero de 2026.
Encontraron cinco problemas. El primero: al firmar la nota, el clínico asume como propio lo que escribió la máquina. Varios comentarios señalan que revisar bien lleva tiempo, y que los sistemas que premian la productividad desalientan esa revisión. El segundo: los errores no se ven como fallas reparables, sino como algo inevitable de estos programas. Las notas inventan datos que suenan razonables. Un caso citado: un programa cambió un medicamento por otro de nombre parecido y de uso completamente distinto. Otro: confundió quién había dicho cada cosa y puso en boca del médico una pregunta del paciente.
El tercero: la nota debe servir a dos lectores a la vez, al colega que continúa la atención y al paciente que la lee. Los participantes describen notas que "no suenan a uno mismo", más largas y parecidas entre sí. El cuarto: los acuerdos de privacidad existentes no encajan con empresas que usan los datos para entrenar sus modelos. El quinto: algunos perciben que la documentación inflada podría encarecer el cobro, aunque eso se plantea como sospecha, no como hecho comprobado.
Conviene ser claro sobre qué es esto. Son comentarios de clínicos en internet, no sentencias ni normas. Muestran lo que ellos dicen que ocurre, no lo que la ley de su país ordena. El estudio analizó conversaciones de foros en línea de profesionales de la salud, no describe el sistema de salud de América Latina.
Tampoco es prueba de que una factura o un código concretos estén mal. Los propios autores señalan que el encarecimiento por documentación excesiva fue percibido por los participantes, no demostrado. Y como es un análisis de conversaciones, no puede afirmarse que estos programas causen daños.
Lo que sí deja son preguntas útiles para llevar a su próxima consulta. ¿Quién revisa la nota antes de firmarla? ¿Se graba la conversación y dónde queda esa grabación? Si usted encuentra un error, ¿ante quién lo presenta y en cuánto tiempo le responden?
Y una para su autoridad de protección de datos: cuando una clínica usa uno de estos programas, ¿qué le exige la ley sobre el uso de su voz y su información?
Pregunte en su próxima cita: "¿Quién revisa esta nota antes de firmarla y cómo puedo corregirla si está mal?"
Qué significa para usted
Ese estudio no describe su ley ni la de su país: son comentarios de clínicos en foros, sin verificación de daños ni cifras. Aun así, le deja preguntas concretas para su próxima cita, sobre quién revisa la nota antes de firmarla, si se graba la conversación y ante quién se corrige un error. Para saber qué exige su autoridad, consulte las normas y decisiones oficiales.
Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516
Quién pagó: Los autores declaran que no tienen nada que informar sobre financiación y que no existen conflictos de interés; el artículo no nombra ningún financiador, equipo ni proveedor de software.
No tome esto como consejo médico profesional.
Los hallazgos de otros estudios que aquí se mencionan los conocemos por este documento, que fue el que leímos; no abrimos cada uno de esos estudios.
Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1399 palabras · unos 7 minLeerla →Cerrar
Las notas clínicas escritas por IA y firmadas por su médico: quién responde cuando el texto se equivoca
Un análisis de conversaciones entre clínicos describe cinco problemas de gobierno en el expediente que el paciente ya puede leer.

El artículo describe un estudio en el que los scribes de inteligencia artificial —programas que escuchan la consulta y redactan la nota clínica— se están desplegando con fuerza para aliviar la carga de documentación en el expediente electrónico1. El problema cambia de tamaño cuando esas notas llegan al paciente: una limitación técnica se convierte de inmediato en una crisis de transparencia de cara al público2. Un estudio cualitativo publicado en 2026 analizó 484 comentarios considerados relevantes dentro de 120 hilos de ocho subreddits de profesionales de salud, extraídos de un total de 3,577 comentarios recolectados en esos mismos hilos, con discusiones que van de octubre de 2020 a febrero de 2026. Los autores declaran no tener conflictos de interés y no reportan financiamiento. No se evaluó ningún producto concreto ni se midió ningún resultado en pacientes.
El hallazgo central es de reparto de responsabilidad. Los participantes describen la firma como el punto exacto en que el error pasa a ser suyo: "en cuanto firmas, esa alucinación se vuelve tu responsabilidad"3. El estudio llama a esto un vacío de accountability: la firma funciona como amortiguador legal de un riesgo que no generó el clínico. Entre los errores descritos hay uno especialmente incómodo: el programa pone en boca del médico algo que dijo el paciente. Un comentario lo cuenta así: la nota decía que él había dado cierto porcentaje de probabilidad de respuesta o cura, cuando en la grabación era el paciente quien especulaba o preguntaba, no él4. Otro caso, contado por una persona que se identifica como paciente, relata que por su acento la IA malinterpretó la mitad de lo que dijo y nadie corrigió nada: hospitalizada por herpes zóster diseminado, el texto hablaba de VIH y de posibles mecanismos de transmisión5. Una nota puede sonar razonable y estar lejos de lo que ocurrió. Eso es lo que los autores llaman riesgo epistémico.
El estudio tiene límites que conviene decir en voz alta. Son comentarios de foros, no sentencias ni resoluciones de reguladores: muestran lo que los clínicos dicen que pasa, no lo que la ley de su país exige. El corpus viene de ocho comunidades en inglés, y el artículo no dice nada sobre normas latinoamericanas. El propio artículo aclara que el upcoding —el registro inflado que puede subir la facturación— fue percibido, no demostrado. El muestreo fue intencional, con tope de 120 hilos y máximo de 40 por comunidad, y la adecuación se justificó por la riqueza temática, no por la representatividad estadística. Los puntajes de Reddit se guardaron solo como contexto, no como medida de verdad. Y los datos completos no se comparten por razones éticas y de privacidad.
Vale la pena entender cómo llegamos aquí, porque no es un accidente. El artículo describe un estudio en el que, en Estados Unidos, el marco contra el bloqueo de información asociado a la ley 21st Century Cures reforzó la expectativa de acceso gratuito a los datos, y la guía federal posterior aclaró mecanismos estrictos de sanción para quien interfiera indebidamente6. Los portales de pacientes dejaron de ser una innovación local discrecional: son la lógica operativa obligatoria de la atención digital moderna7. Es decir, el paciente que lee su nota no está ejerciendo una curiosidad: está usando un derecho que el sistema le debe. Cuando encima el texto lo redactó una máquina, la conversación deja de ser sobre comodidad administrativa.
Aquí aparece la asimetría que más debería importar al lector. El artículo describe un estudio en el que los pacientes leen las notas y detectan errores con frecuencia, pero carecen por completo de rutas prácticas para corregirlos8. En los hilos, la pregunta básica —si los pacientes saben que se usa un scribe y si deben aceptarlo antes— aparece formulada y queda sin respuesta9. Y hay una percepción de inevitabilidad: un comentario, hablando como paciente, compara la situación con la de un profesor que hoy no puede impedir del todo que sus alumnos usen IA10. El resultado es una transparencia que muestra sin dar herramientas.
Los otros trabajos que leímos apuntan en la misma dirección, y conviene decirlo con cuidado. Podríamos leer solo el resumen; el texto completo está detrás de una suscripción. Son señales de un mismo patrón: la tecnología llegó antes que las reglas.
Así lo leemos nosotros. Cuando una institución exige una marca personal —una firma, un nombre— sobre un texto que en realidad produjo una máquina, esa marca deja de certificar quién dijo qué y pasa a repartir responsabilidades y dar apariencia de autenticidad. Es lo que cabe esperar que ocurra donde las normas y los contratos con proveedores se apoyen en la firma del profesional como única ancla, sin obligar al fabricante a registrar de dónde salió cada frase. Sabríamos que nos equivocamos si aparecieran reglas que exijan identificar la versión y el origen de cada fragmento generado, que permitan al paciente pedir una rectificación con respuesta obligatoria, y que hagan responder al proveedor por los errores que su herramienta introduce en el expediente. Mientras eso no exista, usted puede pedir en su próxima consulta que le expliquen si se usó un scribe, quién revisó el texto y cómo se corrige un error, y guardar copia de la nota para compararla con lo que recuerda de la conversación.
Hay una segunda cosa que vemos, y es sobre la lectura misma. Quien lee un texto sin conocer sus reglas lo interpreta con su propia clave: miedo, vergüenza, sospecha sobre seguros o trámites. Por eso cabe esperar que muchos pacientes discutan frases que el sistema considera neutrales, sobre todo cuando el texto lo redactó una máquina sin contexto de la conversación. En los hilos aparece un ejemplo pequeño y elocuente: una paciente pidió que se cambiara "bebe 3 cervezas al día" por "bebe socialmente", el profesional se negó respetuosamente y ella se enfureció11. También se describe que a los pacientes les ofende que lo documentado no coincida con su recuerdo, y que reaccionan a palabras como obesidad o a fórmulas habituales como "el paciente niega" o "el paciente se queja"12. Nos equivocaríamos si resultara que los pacientes, en general, entienden las notas tal como el equipo las redactó, que casi no hay disputas y que las que hay se resuelven rápido con una explicación. Usted puede anotar de antemano qué quiere que quede registrado sobre su salud y sus hábitos, y revisar la nota después para detectar frases que no reconoce como propias antes de que circulen a seguros o a otros médicos.
Una tercera cosa merece una línea, porque afecta el bolsillo. Los datos pueden usarse de formas que el paciente no autorizó: hay políticas que permiten usar información desidentificada o agregada para "mejorar la plataforma", compartirla con terceros o con proveedores en el extranjero13. Antes de firmar cualquier formulario, pregunte por dónde entra un reclamo sobre la nota, quién responde y en cuánto tiempo, y pida que la respuesta quede por escrito.
No queremos dejarle una impresión más negra de la que corresponde.
Sobre lo que tendría que pasar después, el artículo es explícito. Los autores piden pasar de una transparencia pasiva a una transparencia condicionada, con mecanismos autoritativos y exigibles de trazabilidad del origen del texto, impugnación de errores y responsabilidad del proveedor14. Piden vías claras para que pacientes y clínicos disputen inexactitudes, atribuciones equivocadas y contenidos en discusión15. Y piden dientes: auditoría, responsabilidad exigible del fabricante y estándares robustos de manejo de datos, seguridad clínica y condiciones de revisión16. Nada de eso existe todavía en la mayoría de los lugares.
Para una familia en América Latina o en América del Norte, la pregunta correcta no es si la IA escribe mejor o peor. La pregunta es quién puede probar qué se dijo en esa consulta, y quién tiene la obligación de corregirlo. Hoy, en la mayoría de los sistemas, la respuesta recae sobre la persona que firma. Si eso le parece poco, tiene razón en preguntarlo.
La próxima vez que le entreguen o le abran una nota clínica, pida que le digan si la escribió una persona o un programa, quién la revisó y a dónde se dirige un reclamo si algo no coincide con lo que usted vivió. ¿Le darán una respuesta por escrito?
De dónde sale cada dato de contexto, y cuánto leímos de cada documento
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - el artículo del que trata esta nota — el artículo completo — el pasaje: "AI scribes are being deployed aggressively to address the well‐documented burden of electronic health record documentation."
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Once machine‐generated notes are shared directly with patients, technical limitations immediately escalate into patient‐facing transparency crises."
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Users note that “as soon as you hit sign, that hallucination becomes your responsibility” (Q1), marking the exact transfer of institutional liability to the individual practitioner."
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - el artículo del que trata esta nota — el artículo completo — el pasaje: "“I've had the AI scribe say that I said there was xyz% chance of response/cure, but when I reviewed the recording/transcript, it was the patient speculating/asking a question — not anything that I said.”"
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - el artículo del que trata esta nota — el artículo completo — el pasaje: "“As a patient: I have an accent and AI misinterpreted what I said half the time, and no one bothered to correct it. I was hospitalized for disseminated zoster, but the AI wrote a whole thing about HIV and possible mechanisms of transmission.”"
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The information blocking framework associated with the 21st Century Cures Act in the United States has strengthened expectations for free data access, while concurrent federal guidance has clarified strict enforcement mechanisms for actors who improperly interfere"
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Patient portals are no longer discretionary local innovations. They are the mandated operating logic of modern digital care."
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Patients frequently read notes and identify mistakes while entirely lacking practical routes to correction."
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - el artículo del que trata esta nota — el artículo completo — el pasaje: "“Are your patients aware that you're using Abridge? Do they need to agree before you can use it?”"
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - el artículo del que trata esta nota — el artículo completo — el pasaje: "“I don't think we, speaking as patients, can totally stop our physicians from using AI, in as much as no teacher nowadays can stop pupils from using AI to some degree.”"
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - el artículo del que trata esta nota — el artículo completo — el pasaje: "“She asked him to change his note from, “drinks 3 beers per day” to “drinks socially.” He respectfully declined and she was livid."
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - el artículo del que trata esta nota — el artículo completo — el pasaje: "“Patients find it offensive when what is documented does not match their recollection of the visit. They may also react to words like obesity and commonly used phrases like “patient denies” or “patient complains.”"
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - el artículo del que trata esta nota — el artículo completo — el pasaje: "“Their policy lets de‐identified or aggregated patient data be used to “improve the platform.” Data can be shared with third‐party or overseas providers. Patient data might be handled in ways your patients haven't consented to.”"
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - el artículo del que trata esta nota — el artículo completo — el pasaje: "This requires establishing authoritative, enforceable mechanisms for provenance tracking, error contestation, and vendor accountability."
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - el artículo del que trata esta nota — el artículo completo — el pasaje: "That shift requires explicit VOICE mechanisms through which patients and clinicians can contest inaccuracies, misattributions, and disputed content using clear and responsive institutional pathways."
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - el artículo del que trata esta nota — el artículo completo — el pasaje: "It also requires TEETH in the form of provenance tracking, auditability, enforceable vendor accountability, and robust standards for data stewardship, clinical safety, and review conditions."
Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516
Quién pagó: Los autores declaran que no tienen nada que informar sobre financiación y que no existen conflictos de interés; el artículo no nombra ningún financiador, equipo ni proveedor de software.
No tome esto como consejo médico profesional.
Los hallazgos de otros estudios que aquí se mencionan los conocemos por este documento, que fue el que leímos; no abrimos cada uno de esos estudios.
analysis of texts · Journal of evaluation in clinical practice · the paper, 1 Aug 2026 · free
Clinicians Say AI Scribes Are Writing Notes Patients Can Read — And They Don't Trust the Result
In a small analysis of online medical forums, doctors describe a gap between what a machine records and what they can vouch for.
Short version · the longer version follows, about 7 min
Ask weeklyAI
Ask me about this study: who was studied, what it found, and what it does not say.
Conversations are saved for as long as weeklyAI exists, to improve the publication. Answers come in the language you write in.
- The study at a glance
- Who
- clinicians posting in online medical forums
- How many
- 484 relevant comments from 120 threads
- Where
- United States (Reddit communities)
- When
- October 2020 to February 2026
- Kind of study
- analysis of what people did
- Who did it
- Aston University and other institutions
- The limit that matters
- It is talk among practitioners online, not a law, court ruling or regulator's decision.
If the answer is unclear, that gap is the story.

In the United States, patients can now open their medical notes and read what their doctor wrote about them. Increasingly, clinicians posting in online forums say a machine is writing part of that record first.
Researchers at Aston University and other institutions analysed those conversations in a qualitative study. They extracted Reddit threads with a Python script and examined 484 relevant comments from 120 threads across eight clinician-oriented communities, from October 2020 to February 2026. The forums included family medicine, medicine, health technology, residency, emergency medicine, medical school, hospitalist and nursing groups.
Five concerns emerged. First, clinicians described their signature as the point where responsibility for a machine-written error lands on them, with little real ability to check everything. Second, they described errors as an inevitable property of the technology rather than a bug to be fixed. One commenter reported an AI scribe attributing a patient's speculation to the doctor. Another described a note that merged a wife's condition with her husband's recent condition into one plausible-sounding but false story. Third, they said the notes no longer sound like them, and that patients find some phrasing offensive. Fourth, they worried that recordings and data flow to companies under privacy rules that no longer fit. Fifth, they suspected the software inflates notes in ways that could affect billing.
This is talk among practitioners online, not a court ruling, a law or a regulator's decision. It cannot tell you what your country requires of an AI scribe vendor.
The study used publicly accessible Reddit content from eight clinician-oriented subreddits. One Canadian participant did comment that AI scribes 'don't save me noticeable time personally,' linking this to a setting where routine encounters do not require extensive documentation for reimbursement.
And on billing, the worry was perceived by participants, not demonstrated. The authors state plainly that upcoding was not shown to have occurred.
What the study does offer is a set of questions worth asking where you live. If a clinic near you uses an AI scribe, you can ask who reviews the note before it is signed, whether your conversation is recorded, where that recording goes, how long it is kept, and whether it trains the company's models. You can ask who to contact if the note contains something that never happened.
If the answer is unclear, that gap is the story.
What this means for you
For you, this study is not a court ruling or a regulator's decision, so it cannot tell you what your country allows an AI scribe vendor to do with your voice or your record. What it does give you is a set of questions to carry into your next visit: who reviews the note before it is signed, where the recording goes, and who to contact if something in it never happened.
Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516
Who paid: The authors state they have nothing to report on funding, and they declare no conflicts of interest; the article names no funder, equipment or software provider.
Do not take this as professional medical advice.
The findings of other studies mentioned here are known to us through this document, which is the one we read; we did not open each of those studies.
The longer versionThe passages copied from the paper, the pictures, and every source with how much of it we read · 1343 words · about 7 minRead it →Close
Your Visit Note May Have Been Written by Software. Here's What Clinicians Say Happens When It's Wrong.
Clinicians in eight online forums describe signing notes they cannot fully verify — and patients with no clear route to correct them.

Artificial intelligence is now writing parts of the medical record, and patients in the United States can read those notes. A study published in the *Journal of Evaluation in Clinical Practice* asked how clinicians talk about this when they think patients are watching. The authors — Samuel Atiku of Aston University and the University of Staffordshire, and Olufisayo Olakotan of University Hospitals Leicester and the University of Wolverhampton — analysed 484 comments from 120 threads across eight clinician-oriented Reddit communities, posted between October 2020 and February 2026. Aston University granted ethical approval; the authors report no funding and no conflicts of interest.
The tool at the centre of this is an ambient AI scribe: software that listens during a visit and drafts the clinical note, which a clinician then reviews and signs1. That signature is where the trouble starts. In the threads, clinicians described it as the exact point where a machine's error becomes their personal responsibility: "as soon as you hit sign, that hallucination becomes your responsibility"2. The word "hallucination" here means the software inventing something that was never said. One clinician described an AI scribe recording a survival estimate as if the doctor had said it, when the recording showed the patient was the one speculating3.
The study's design matters for what it can and cannot tell you. It is a qualitative analysis of online discussion — not a trial, not a review of medical records, not a survey. The authors searched Reddit using a custom script, collected 3,577 comments across 120 threads, and manually narrowed that to 484 relevant comments. Reddit scores were kept only as context, not as a measure of truth. The authors note the full dataset is not shared because of privacy concerns around potentially identifiable online content. The article does not state that the corpus is mostly United States settings, nor does it discuss a country-level sampling frame. The data come from Reddit communities without stated geographic composition.
Five themes emerged. The first is an accountability vacuum: clinicians described proofreading as mandatory in principle but undermined in practice, because "proofreading is disincentivized in any model which rewards productivity"2. The second is hallucination as a permanent condition rather than a fixable bug. The third is a dual-audience problem — the note must serve both the next clinician and the patient, and AI output often fails both. The fourth is privacy: clinicians worried that de-identified data could be shared with third parties or overseas providers "in ways your patients haven't consented to"4. The fifth is fiscal: some clinicians said AI notes were "almost a little too good," raising the concern of over-documentation that could inadvertently inflate billing. The authors state upcoding was perceived, not demonstrated.
Patient-facing access is not optional in the United States. The article describes a study in which the information-blocking framework tied to the 21st Century Cures Act strengthened expectations for free data access, with enforcement mechanisms for actors who improperly interfere5. As the authors put it, patient portals "are no longer discretionary local innovations" but "the mandated operating logic of modern digital care"6.
The study found a gap between visibility and correction. Patients can read the note; the threads describe little evidence they can get it fixed. One patient commenter described an AI scribe mishearing their accent and writing about HIV transmission when they had been hospitalised for a different condition, with no correction made7. Another clinician noted that patients "find it offensive when what is documented does not match their recollection of the visit"8 — and a third described a patient asking that "drinks 3 beers per day" be changed to "drinks socially," which the clinician declined9. The article describes a study in which patients "frequently read notes and identify mistakes while entirely lacking practical routes to correction"10.
The questions clinicians asked each other reveal how unsettled consent still is. One asked plainly: "Are your patients aware that you're using Abridge? Do they need to agree before you can use it?"11 — and in that thread, the question went unanswered. One patient commenter compared the situation to classrooms: "no teacher nowadays can stop pupils from using AI to some degree"12.
Here is how we read it. When a record about you is drafted by a system built to produce fluent text rather than to be checked by the person it describes, you lose the ability to tell which words were the clinician's and which were the machine's. We expect that a patient who finds something wrong in an AI-written note will often not know which sentence to challenge, and will end up asking a receptionist or a nurse to explain the record rather than filing a formal correction. We would be wrong if patients routinely identified the specific wrong sentence, named who wrote it, and received a documented correction without help from anyone inside the clinic. When you read a visit note, ask in writing which parts were drafted by software and which by the clinician, and keep your own copy of what you actually said. Ask how a correction gets attached to the record and who signs off on it.
We also read other work on this family of problems. One study of governance across European health systems found that health-specific AI strategies and liability standards remain rare, according to the summary — we could read only the abstract, the full paper sits behind a subscription. A separate review of AI in surgery described risks that can be grouped into three kinds — those inside the system, those introduced by the clinician using it, and those created by how an institution deploys it — with liability potentially reaching developers and hospitals, again from the abstract only. That is boundary evidence: it concerns whether findings of this kind carry over to other settings, not whether the Reddit analysis happened.
The pattern we notice is that records acquire momentum. Once a clinic has bought a tool and trained staff on it, undoing that is harder than installing it was, and an entry can outlast the reason it was collected — resurfacing in insurance, disability, or later care. We expect a clinic that has already invested in an AI scribe will keep using it even after complaints, and that a wrong entry will stay in the chart long after the visit. We would be wrong if clinics promptly switched tools or reverted to human notes after complaints, and old machine-written errors were removed or clearly marked as disputed within a short, stated time. Ask how long the clinic keeps the AI-drafted text and whether it is ever deleted or flagged. Ask for any correction to be dated and visible in future notes. If you change clinics or insurers, request your record and check whether the disputed entry travelled with you.
The last thing we take from this is that complaining alone rarely changes how an institution treats you. What tends to produce a written answer is a formal procedure, kept in writing, with dates — and sometimes the same issue raised by more than one person. We expect individual objections to be met with verbal reassurance, while a written records request or a regulator's process is more likely to produce an actual amendment. We would be wrong if single, unassisted complaints reliably produced written corrections or policy changes. Put your objection in writing, ask for a written response, and keep the dates. If the answer is unsatisfactory, ask your country's data protection or health regulator what procedure applies where you live — the study says nothing about Latin American, Canadian or other national rules, so that question is one only your own authority can answer.
The question to bring to your next appointment: which parts of this note did a machine draft, and how do I get a correction attached if something in it is wrong?
Where each piece of context comes from, and how much of it we read
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - the article this story is about — the whole article — the passage: "AI scribes are being deployed aggressively to address the well‐documented burden of electronic health record documentation."
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - the article this story is about — the whole article — the passage: "Users note that “as soon as you hit sign, that hallucination becomes your responsibility” (Q1), marking the exact transfer of institutional liability to the individual practitioner."
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - the article this story is about — the whole article — the passage: "“I've had the AI scribe say that I said there was xyz% chance of response/cure, but when I reviewed the recording/transcript, it was the patient speculating/asking a question — not anything that I said.”"
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - the article this story is about — the whole article — the passage: "“Their policy lets de‐identified or aggregated patient data be used to “improve the platform.” Data can be shared with third‐party or overseas providers. Patient data might be handled in ways your patients haven't consented to.”"
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - the article this story is about — the whole article — the passage: "The information blocking framework associated with the 21st Century Cures Act in the United States has strengthened expectations for free data access, while concurrent federal guidance has clarified strict enforcement mechanisms for actors who improperly interfere"
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - the article this story is about — the whole article — the passage: "Patient portals are no longer discretionary local innovations. They are the mandated operating logic of modern digital care."
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - the article this story is about — the whole article — the passage: "“As a patient: I have an accent and AI misinterpreted what I said half the time, and no one bothered to correct it. I was hospitalized for disseminated zoster, but the AI wrote a whole thing about HIV and possible mechanisms of transmission.”"
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - the article this story is about — the whole article — the passage: "“Patients find it offensive when what is documented does not match their recollection of the visit. They may also react to words like obesity and commonly used phrases like “patient denies” or “patient complains.”"
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - the article this story is about — the whole article — the passage: "“She asked him to change his note from, “drinks 3 beers per day” to “drinks socially.” He respectfully declined and she was livid."
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - the article this story is about — the whole article — the passage: "Patients frequently read notes and identify mistakes while entirely lacking practical routes to correction."
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - the article this story is about — the whole article — the passage: "“Are your patients aware that you're using Abridge? Do they need to agree before you can use it?”"
- Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516 - the article this story is about — the whole article — the passage: "“I don't think we, speaking as patients, can totally stop our physicians from using AI, in as much as no teacher nowadays can stop pupils from using AI to some degree.”"
Atiku, S., Olakotan, O. (2026). An Evaluation of AI‐Generated Clinical Notes in the OpenNotes Era: A Thematic Analysis of Clinician Discourse. Journal of Evaluation in Clinical Practice. https://doi.org/10.1111/jep.70516
Who paid: The authors state they have nothing to report on funding, and they declare no conflicts of interest; the article names no funder, equipment or software provider.
Do not take this as professional medical advice.
The findings of other studies mentioned here are known to us through this document, which is the one we read; we did not open each of those studies.