analysis of texts · Frontiers in Pharmacology · la publicación, 4 nov 2026 · gratis
El modelo sacó más puntos que los médicos, y en 19 de 70 inventó datos del paciente
En un hospital de Beijing se compararon textos del chatbot con los de los médicos. Los del chatbot cubrieron más el riesgo y las alternativas, pero uno de cada cuatro afirmó cosas que la historia clínica no respaldaba.
Versión breve · la versión detallada sigue, unos 6 min
Pregunte al chat de weeklyAI
Pregúnteme sobre la IA en su vida: le señalo las notas que weeklyAI ha publicado sobre el tema, esta semana y antes.
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
- Adultos que recibieron terapias dirigidas en reumatología
- Cuántos
- 70 casos, 140 textos
- Dónde
- Hospital Tongren de Pekín, China
- Cuándo
- Pacientes tratados durante 2025
- Tipo de estudio
- analysis of what people did
- Quién lo hizo
- Hospital Tongren de Pekín, Universidad Médica Capital
- El límite que importa
- Un solo hospital, en chino, con escala hecha por los propios autores; nunca se mostró a pacientes.
Porcentaje de los 70 textos de cada grupo que contenían afirmaciones sin respaldo en el expediente; son textos de investigación, nunca mostrados a pacientes.
Textos del modelo frente a textos de médicos
Calificación mediana de 60 contra 50
Nivel de lectura de 7.87 contra 7.10; ambos por encima de secundaria alta
Datos inventados en 27.1% contra 0.0%
Omisiones en 8.6% contra 20.0%, sin diferencia clara
Cuando le entreguen un consentimiento para un tratamiento dirigido, revise si cada frase sobre sus riesgos y su historia aparece en su expediente.
Diecinueve consentimientos de setenta. En cada uno, la máquina afirmó como hecho algo sobre el paciente que su expediente no decía. Ningún texto de médico lo hizo.
El estudio se hizo en el servicio de reumatología del Hospital Tongren de Beijing, en China. Setenta adultos recibieron durante 2025 alguno de catorce tratamientos dirigidos ya aprobados para sus enfermedades. Cada uno firmó un consentimiento redactado por su médico.
Para cada caso, los investigadores dieron a un modelo de lenguaje, llamado DeepSeek-V4-pro, el expediente clínico sin nombres y una instrucción fija. El modelo escribió su propio texto. Fueron 140 textos, y tres reumatólogos los calificaron sin saber cuál había hecho quién.
El modelo sacó más puntos en total: mediana de 60 contra 50. Cubrió mejor los riesgos, los beneficios y las otras opciones. En la información básica, la lógica y el estilo de escritura no hubo diferencia clara entre los dos. En la fidelidad al expediente, la diferencia entre los dos grupos no quedó clara después de ajustar por comparaciones múltiples.
También resultó más difícil de leer: 7.87 contra 7.10 en una escala que mide nivel de lectura. Los dos pasaron del nivel de secundaria alta. Un puntaje más alto quiere decir más carga para el lector, no más comprensión.
Los textos del modelo se produjeron después, solo para la investigación. Nunca se le mostraron a un paciente ni se usaron en su atención. Nadie preguntó a los pacientes si entendieron el tratamiento, los riesgos o las alternativas.
Es un solo hospital, en chino, calificado con una escala que armaron los propios autores. El tamaño de la diferencia puede no repetirse en otro país, otro idioma u otro hospital.
Cuando le entreguen un consentimiento para un tratamiento dirigido, revise si cada frase sobre sus riesgos y su historia aparece en su expediente. Y pregunte quién redactó el borrador y quién lo revisó línea por línea.
Qué significa para usted
Cuando le entreguen un consentimiento para un tratamiento dirigido, puede preguntar quién redactó el borrador y quién lo revisó línea por línea. También vale revisar si cada frase sobre sus riesgos y su historia aparece en su expediente. Ningún país ha decidido todavía que estos textos puedan firmarse sin esa revisión médica.
Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826
Quién pagó: Los autores declararon que no recibieron apoyo financiero para este trabajo ni para su publicación, y que no tenían relaciones comerciales o financieras que pudieran constituir un conflicto de interés; el artículo no dice quién proporcionó el modelo DeepSeek-V4-pro, que los autores describen como de despliegue local.
No tome esto como consejo médico profesional.
Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1183 palabras · unos 6 minLeerla →Cerrar
La máquina redactó el formulario de consentimiento; en una de cada cuatro veces inventó datos del paciente
Un estudio en un hospital de Pekín comparó textos escritos por un modelo de lenguaje con los que redactaron los médicos. El modelo cubrió mejor los temas, pero se inventó datos que ningún expediente respaldaba.
Piense en el formulario que le ponen enfrente antes de una terapia dirigida, esos medicamentos que atacan una parte precisa del sistema inmune. Ahí deben aparecer cuatro cosas: la información básica del tratamiento, los riesgos, los beneficios y las alternativas. En el Departamento de Reumatología del Hospital Tongren de Pekín tomaron 70 casos reales de pacientes tratados durante 2025 y, para cada uno, juntaron dos textos: el que había escrito el médico de cabecera y otro que generó un modelo de lenguaje a partir de los mismos datos clínicos, sin nombres. Tres reumatólogos calificaron los 140 textos sin saber cuál era de quién.
El texto de la máquina sacó una calificación mediana de 60 sobre el puntaje compuesto; el del médico, 50. La diferencia fue mayor en riesgos, beneficios y alternativas. Ahí se acaba la ventaja tranquila: los mismos evaluadores encontraron invenciones en 19 de los 70 textos generados por el modelo, el 27.1 por ciento, y en ninguno de los escritos por médicos. El artículo las llama alucinaciones, y son justo lo que suena: deducciones que el modelo presenta como hechos del paciente, sin respaldo en el expediente.
Hubo otros dos resultados que conviene no mezclar. Los textos del modelo pedían más nivel de lectura que los de los médicos, y ambos pasaban del nivel de bachillerato, según la escala que usaron. En errores por datos mal copiados y en omisiones, las diferencias no alcanzaron significancia estadística: no se puede decir que la máquina se equivocara menos ni que olvidara menos.
Los límites son grandes y hay que decirlos completos. Fue un solo hospital, en chino, con una escala que armaron los propios autores. Los textos del modelo se produjeron después, para la investigación: nunca se le mostraron a un paciente ni se usaron en la atención. A nadie se le preguntó si entendió el tratamiento, los riesgos o las alternativas. No hubo cálculo previo del tamaño de muestra; el número salió de los casos disponibles en el periodo. Cuatro de los trece criterios de calificación quedaron por debajo del acuerdo entre evaluadores que los autores se habían fijado. Y no se midió cuánto trabajo le costaría al médico corregir el borrador.
Veamos cómo funciona la cosa, paso a paso. Al modelo se le entregó un conjunto fijo de datos del caso, sin acceso al expediente completo1. Con instrucciones idénticas para todos los casos, redactó el texto2. Lo curioso es que esa misma disciplina explica las dos caras del resultado: un texto que sigue un orden fijo cubre los temas predefinidos, y también rellena los huecos que nadie le dio. Cuando al modelo le falta un dato, no siempre se detiene: lo completa con algo verosímil3. Los autores piden que, cuando la información no alcance, el texto conserve frases como "información no proporcionada" o "pendiente de confirmación por el médico tratante", y que el modelo no las rellene por su cuenta4.
El fenómeno no es exclusivo de este estudio. Los propios autores señalan que la evidencia sobre modelos de lenguaje para redactar consentimientos de tratamientos rutinarios es escasa, y que faltaban justamente comparaciones ciegas con casos reales y textos de médicos en reumatología5. En trabajos anteriores sobre consentimientos quirúrgicos y de ensayos clínicos se vio un desempeño parecido en cobertura de información y orden6. Y en resúmenes de alta hospitalaria, otros investigadores encontraron que los textos de modelos de lenguaje tienden más a los errores de hecho que los de médicos7.
En trabajos anteriores sobre consentimientos quirúrgicos y de ensayos clínicos se vio un desempeño parecido entre los textos de los modelos y los de las personas en cobertura de información y orden6. Es decir, en la misma dirección que la ventaja de cobertura que reportó el estudio chino de reumatología. Eso no desmiente este resultado: muestra que la ventaja de la máquina no viaja sola a otros países, idiomas o especialidades.
En este estudio no se preguntó a los pacientes si entendieron el tratamiento, los riesgos o las alternativas.
Así lo leemos nosotros: cuando una decisión delicada pasa por un programa, la gente tiende a aceptarla como un hecho consumado. Es probable que muchos pacientes firmen un borrador automático sin preguntar quién lo escribió ni de dónde salió cada dato de su caso, y que tomen por cierto cualquier detalle clínico que el texto haya inventado. Nos equivocaríamos si viéramos lo contrario: pacientes que preguntan de forma espontánea por el origen del texto, que piden señalar qué frases vienen de su expediente y que cuestionan lo que no reconocen, aunque el formulario se vea completo y ordenado.
En trabajos anteriores sobre consentimientos quirúrgicos y de ensayos clínicos se vio un desempeño parecido entre los textos de los modelos y los de las personas en cobertura de información y orden. Eso habla de si el resultado se puede trasladar a otros lugares, no de lo que ocurrió en este hospital.
Hay algo más que nos parece la advertencia central: un texto ordenado y completo puede parecer confiable sin serlo. Lo que el estudio sugiere es que conviene revisar frase por frase los datos personales, y que esa revisión no la puede hacer el programa que redactó el borrador. Los propios autores lo plantean así: el médico tratante debe verificar diagnóstico, esquema de tratamiento, factores de riesgo individuales, indicaciones, contraindicaciones y alternativas, y confirmar que cada dato del paciente se pueda rastrear a un expediente, una etiqueta del medicamento o una guía clínica10.
Usted puede hacer algo concreto la próxima vez que le pongan un formulario enfrente, sea de una terapia dirigida o de cualquier otro tratamiento. Antes de firmar, pregunte quién redactó el documento y pida que le señalen, frase por frase, de dónde sale cada dato sobre su salud. Si algo no aparece en su expediente, pida que se corrija o que se marque como no confirmado. Pregunte también si el borrador lo hizo una persona o un programa, y quién lo revisó después; pida que esa revisión quede por escrito y que le expliquen qué se cambió respecto del borrador original. Y si el texto es difícil de leer, pida que se lo expliquen en palabras sencillas y repita con sus propias palabras los riesgos y las alternativas. Si algo no le queda claro, pida que se lo aclaren y que quede anotado.
¿Cómo sabría que nos estamos equivocando? Si los pacientes empiezan a preguntar de dónde salió cada frase y los equipos médicos advierten de entrada que un borrador automático puede contener errores, la desconfianza útil habrá llegado antes que el problema.
El formulario sigue siendo el mismo papel, con el mismo encabezado y la misma línea para la firma. Lo que cambió es lo que usted ya puede preguntar antes de levantar la pluma: ¿quién escribió esto, y de dónde sacó lo que dice de mí?
De dónde sale cada dato de contexto, y cuánto leímos de cada documento
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The LLM did not access or browse the complete EHR and received only the predefined case-information set."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - el artículo del que trata esta nota — el artículo completo — el pasaje: "LLM texts were generated using DeepSeek-V4-pro based on identical de- identified case data and standardized prompts, and corresponding physician texts were extracted from electronic records."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - el artículo del que trata esta nota — el artículo completo — el pasaje: "In LLM-generated texts, inaccuracies primarily reflected factual discrepancies when restating medical record information. Hallucinations mainly manifested as unsupported inferences presented as patient-specific facts, whereas omissions typically involved incomplete reporting of prior diagnoses or individualized risk factors related to the proposed treatment."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - el artículo del que trata esta nota — el artículo completo — el pasaje: "For content with insufficient data, statements such as “information not provided” or “to be further confirmed by the treating physician” should be retained, and should not be autonomously completed by the model."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - el artículo del que trata esta nota — el artículo completo — el pasaje: "At present, empirical evidence regarding the use of LLMs to assist in generating informed consent forms for routine clinical treatment remains limited. In particular, blinded comparative studies based on real patient cases and texts routinely prepared by physicians are lacking in the context of targeted therapies in rheumatology."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Previous studies on informed consent texts for surgical treatment and clinical trials have similarly observed comparable LLM performance in information coverage and structured expression (Decker et al., 2023; Shi et al., 2025)."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Previous research on generating discharge summaries has similarly indicated that LLM- generated medical texts are more prone to factual errors compared with those drafted by physicians (Williams et al., 2025)."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - el artículo del que trata esta nota — el artículo completo — el pasaje: "After LLM generation of a draft, the treating physician should verify patient diagnosis, treatment regimen, individualized risk factors, indications, contraindications, risk statements, and alternative treatment options item by item, and confirm that patient-specific information is traceable to medical records, drug labels, clinical guidelines, or other explicit evidence."
Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826
Quién pagó: Los autores declararon que no recibieron apoyo financiero para este trabajo ni para su publicación, y que no tenían relaciones comerciales o financieras que pudieran constituir un conflicto de interés; el artículo no dice quién proporcionó el modelo DeepSeek-V4-pro, que los autores describen como de despliegue local.
No tome esto como consejo médico profesional.
Los hallazgos de otros estudios que aquí se mencionan provienen de este documento; esos estudios no se consultaron uno por uno.
analysis of texts · Frontiers in Pharmacology · the paper, 4 Nov 2026 · free
A Machine Wrote 70 Consent Forms. One in Four Invented a Medical Fact About the Patient.
At a single Beijing hospital, chatbot drafts beat doctor drafts on a scoring sheet. No patient ever read one.
Short version · the longer version follows, about 7 min
Ask the weeklyAI chat
Ask me about AI in your life: I point you to the stories weeklyAI has published on it, this week and before.
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
- adults receiving targeted therapy for rheumatic diseases
- How many
- 70 patients, 140 texts
- Where
- Beijing Tongren Hospital, China
- When
- 2025
- Kind of study
- analysis of what people did
- Who did it
- Beijing Tongren Hospital, Capital Medical University
- The limit that matters
- One hospital, in Chinese, scored on the authors' own scale; no patient read the drafts.
Out of 70 texts in each group, the share containing a statement about the patient that the medical record did not support; these drafts were made for research and never shown to patients.
Machine-written consent drafts against doctor-written drafts, same 70 patients
Higher overall score, 60 against 50, on a scale the authors built themselves
Invented patient-specific facts in 27.1% against 0.0%
Harder to read, 7.87 against 7.10; both above senior-secondary-school level
Missing information in 8.6% against 20.0%, a gap the authors report as not significant
When you are handed a consent form, look for each personal statement about your risks and history and ask whether it traces to your own record.
Nineteen of 70 consent texts written by a chatbot contained a statement about the patient that the medical record did not support. That is 27.1 percent of the chatbot texts. No patient ever saw these drafts, because they were written for research only.
The texts came from Beijing Tongren Hospital. Seventy adults received any of 14 approved targeted therapies for rheumatic diseases during 2025, and each had a consent form their own physician wrote. For each case, researchers fed a large language model, DeepSeek-V4-pro, the same stripped-down medical record and the same fixed instructions, and it wrote its own version. That made 140 texts. Three rheumatologists scored them without knowing who wrote what.
Consider the score sheet. The model's texts came out higher overall: 60 points against 50 on a scale the authors built themselves. They covered risks, benefits and alternatives more fully. On basic information, logic and writing style, the two versions were about equal.
Then the record check. Nineteen model texts invented a patient-specific fact the record did not support. Physician texts did none. The study reports no clear difference in plain errors or in missing information.
The model's texts were also harder to read, at 7.87 against 7.10 on a reading-level scale. Both sat above senior-secondary-school level. A higher number here means more burden on the reader.
Three limits sit close together. This was one hospital, in Chinese, scored with the authors' own scale, and the gap may not hold in your country or your language. No patient was shown the drafts or asked whether they understood anything. And the model's texts were never used in care.
When you are handed a consent form, look for each personal statement about your risks and history and ask whether it traces to your own record. Then ask who wrote the draft and who checked it, line by line.
What this means for you
One Chinese hospital, one language, and a scoring sheet the authors built themselves: the study cannot tell you what a consent form in your own country would say, or whether any patient understood it better. What it does show is a machine that fills in personal detail it was never given. So when a form mentions your history or your risks, it is fair to ask who wrote that line and which record it came from.
Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826
Who paid: The authors declared that no financial support was received for this work or its publication, and declared no commercial or financial relationships that could be a conflict of interest; the article does not say who supplied the DeepSeek-V4-pro model, which the authors describe as locally deployed.
Do not take this as professional medical advice.
The longer versionThe passages copied from the paper, the pictures, and every source with how much of it we read · 1341 words · about 7 minRead it →Close
A Machine Wrote the Consent Form. One in Four Invented Facts About the Patient.
In a Beijing rheumatology department, computer-drafted consent texts outscored doctors' own on a local rubric. They also made up personal details no record supported.
Consider what a consent form for a targeted therapy actually says. It names the treatment, the expected benefit, the main risks, the precautions, and the alternatives if the patient says no. In China, a doctor with prescribing authority drafts that document after the consent conversation, enters it in the electronic record, and the patient signs.
A study published on 4 November 2026 in *Frontiers in Pharmacology* asked a large language model to write those drafts, then set them beside the doctors' versions for the same 70 patients. The patients were adults treated in the rheumatology department at Beijing Tongren Hospital during 2025, five for each of 14 targeted drugs approved in China. Their median age was 47; 41 were women. The authors are researchers from the hospital's pharmacy, rheumatology, medical affairs and rehabilitation departments, led by Li Cui. They declared no funding and no conflicts of interest.
The model was DeepSeek-V4-pro, run locally. For each case it received a fixed, de-identified set of chart information and the same standardized prompt, and it had no access to the full medical record12. The doctors' texts came from the same records. Three rheumatologists, blinded to which text was which, scored all 140 documents on a rating scale the authors built for the study. On the overall score the model's drafts did better: a median of 60 against 50 for the doctors. They also scored higher on risks, benefits and alternatives. On basic information, the two were level.
Then the reviewers checked every clinical statement against the chart. Nineteen of the 70 model drafts — 27.1 percent, roughly one in four — contained what the authors call a hallucination: an unsupported inference presented as a patient-specific fact. None of the doctors' texts did3. The model's errors were mostly of that kind, while its omissions were fewer than the doctors' (8.6 percent against 20.0 percent), a gap the authors report as not statistically significant4. Inaccuracies were also similar between the two, and again not significantly different5. The model's drafts were harder to read: both arms sat above senior-secondary-school level, and the model's mean was higher56. The authors' own conclusion is a warning, not a sales pitch: such drafts may help with structure and coverage, but must not become final consent documents without a physician checking every item78.
What the study did not do matters as much as what it did. The model's texts were generated after the fact for research and never shown to a patient9. Nobody asked a patient whether they understood anything. It was one hospital, in Chinese, scored on a scale the authors wrote themselves. Four of the thirteen scoring items fell below the reliability threshold the authors had set, so some item-level comparisons rest on weaker agreement. There was no advance calculation of how many cases would be needed. And documents with core sections entirely missing were excluded, which may have skewed the sample.
Here is how the thing works, step by step. A language model does not retrieve sentences from a chart. It predicts, word by word, what text plausibly comes next, given the prompt and the case summary it was handed2. A fluent document is exactly what that process produces. A true one is not guaranteed by it. Fluent writing and verified writing are two different products, and only one of them is what a signature is supposed to certify.
The trade-off is visible in a single line of the results. Give the model a template and it covers every heading in order: risks, benefits, alternatives, all present, all in place. Give the same template to a doctor and the coverage varies, because doctors adjust the level of detail to the patient's condition and treatment context10. What the model gains in completeness, it may lose in fidelity: its invented details looked medically plausible, which is precisely why a tired reader would not flag them11. The doctors' own failures ran the other way, toward thin or generalized descriptions of a patient's individual risks12. Neither arm was clean.
Comparison is where this study sits. Its authors write that evidence on using language models for routine consent forms is thin, and that blinded comparisons against real physician texts for rheumatology targeted therapies were missing13. Earlier work on surgical and trial consent found similar model strength in coverage and structure14. Work on discharge summaries found models more prone to factual error than doctors15. Separately, a review of consent documents for AI-related clinical trials — which we could read only in summary — found that more than half failed to disclose the AI's type or intended use, and that 18.4 percent left out risks altogether16. The paperwork problem predates the machine.
The wider picture is a health system adopting these tools faster than it can check them. A real-world evaluation of a patient-facing AI information system for Parkinson's disease in Germany — summary only — followed its first 129 days and 2,035 conversations. Automated triage rated most conversations good, yet expert review of flagged cases confirmed five critical events, and a re-review of 100 randomly sampled "good" conversations found four with clinically critical errors the triage had missed171819. One involved a recommendation that should not have been made; another involved a message signalling suicidal thinking that did not trigger the intended emergency response20. The lesson the authors draw is that critical failures can hide behind good automated scores21.
In the United States — summary only — a review of 16 national surveys of adults found that more than four in ten Americans who used AI for a health question did so partly because they could not access or afford care22. State lawmakers have been busy on AI oversight, accuracy and privacy, but not on that access gap23. The machine is filling a hole in the system, not only a hole in the paperwork.
Here is how we read it. A consent form is a document almost nobody reads closely. It is long, it is dense, and the signature usually follows a conversation, not a careful reading. A draft that reads harder than the doctor's version makes that shortcut more likely, not less. So our expectation — not this study's finding — is that a patient handed a fluent, well-organized machine draft would sign it and trust the spoken explanation, and would not catch an invented personal detail on the page. We would be shown wrong if patients given such drafts actually read them line by line, or spotted the unsupported statements unprompted.
That is the reverse of what the Beijing scores showed. It is a caution about carrying one hospital's result to another country, language or specialty — not a reason to doubt that this result happened.
The checking is the part that gets sold as automatic and is not. A polished draft looks verified because it is fluent, while the work of tracing each personal statement to a record is done by hand, line by line, after the draft appears. Ask your clinician whether they checked each personal statement in the form against your record. Do not treat a well-organized page as proof that anyone did.
So what should a reader take from a form they are about to sign? The authors' own instruction is concrete: the treating physician should verify diagnosis, regimen, individual risk factors, indications, contraindications, risk statements and alternatives item by item, and confirm each personal claim traces to a record, a drug label or a guideline7. Where data are missing, the draft should say so rather than fill the gap itself8.
The consent form is still a piece of paper with your name at the bottom. What has changed is who may have typed it first — and the one question worth asking before you sign: who wrote this draft, and can you show me where each statement about me comes from?
Where each piece of context comes from, and how much of it we read
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - the article this story is about — the whole article — the passage: "LLM texts were generated using DeepSeek-V4-pro based on identical de- identified case data and standardized prompts, and corresponding physician texts were extracted from electronic records."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - the article this story is about — the whole article — the passage: "The LLM did not access or browse the complete EHR and received only the predefined case-information set."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - the article this story is about — the whole article — the passage: "Hallucinations were present in 19 LLM-generated texts (27.1%) but were absent from all physician-authored texts (0.0%), a difference that remained statistically significant after adjustment (adjusted p < 0.01)."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - the article this story is about — the whole article — the passage: "Omission errors occurred in 6 LLM-generated texts (8.6%) and 14 physician-authored texts (20.0%); after adjustment, this difference was not statistically significant (adjusted p = 0.23)."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - the article this story is about — the whole article — the passage: "The LLM arm also exhibited higher reading level scores (7.87 vs. 7.10; p < 0.01) and a higher rate of hallucinations (27.1% vs 0; p < 0.01). No statistically significant between-arm differences were found for inaccuracies or omissions."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - the article this story is about — the whole article — the passage: "The mean scores in both arms exceeded 7, indicating a reading level at or above the senior secondary school level."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - the article this story is about — the whole article — the passage: "After LLM generation of a draft, the treating physician should verify patient diagnosis, treatment regimen, individualized risk factors, indications, contraindications, risk statements, and alternative treatment options item by item, and confirm that patient-specific information is traceable to medical records, drug labels, clinical guidelines, or other explicit evidence."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - the article this story is about — the whole article — the passage: "For content with insufficient data, statements such as “information not provided” or “to be further confirmed by the treating physician” should be retained, and should not be autonomously completed by the model."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - the article this story is about — the whole article — the passage: "Second, patients were not involved in the research-based evaluation of the written texts. Although routine clinical informed consent was conducted by treating physicians and signed by the patients as part of clinical care, this study did not assess patient comprehension, decision quality, or other patient- centered outcomes."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - the article this story is about — the whole article — the passage: "The LLM arm achieved a higher median overall quality score than the physician arm (60 vs. 50; adjusted mean difference 9.95, 95% CI 9.36 to 10.54; p < 0.01), with higher scores for risks, benefits, and alternatives (all adjusted p < 0.01)."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - the article this story is about — the whole article — the passage: "In LLM-generated texts, inaccuracies primarily reflected factual discrepancies when restating medical record information. Hallucinations mainly manifested as unsupported inferences presented as patient-specific facts, whereas omissions typically involved incomplete reporting of prior diagnoses or individualized risk factors related to the proposed treatment."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - the article this story is about — the whole article — the passage: "In physician-authored texts, inaccuracies primarily involved inconsistencies in describing existing diagnoses or patient risk information, as well as overgeneralization of specific high-risk factors. Omissions were mainly characterized by incomplete reporting of individualized risk factors pertinent to treatment. No hallucinations were identified in physician-authored texts."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - the article this story is about — the whole article — the passage: "At present, empirical evidence regarding the use of LLMs to assist in generating informed consent forms for routine clinical treatment remains limited. In particular, blinded comparative studies based on real patient cases and texts routinely prepared by physicians are lacking in the context of targeted therapies in rheumatology."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - the article this story is about — the whole article — the passage: "Previous studies on informed consent texts for surgical treatment and clinical trials have similarly observed comparable LLM performance in information coverage and structured expression (Decker et al., 2023; Shi et al., 2025)."
- Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826 - the article this story is about — the whole article — the passage: "Previous research on generating discharge summaries has similarly indicated that LLM- generated medical texts are more prone to factual errors compared with those drafted by physicians (Williams et al., 2025)."
- Su H, Xiao F, Chau H, Tong Y, Han S, Cheng X, et al. (2026). Informed Consent Disclosures and Minimum Requirements in AI Clinical Trials: Cross-Sectional Analysis. Journal of Medical Internet Research. 10.2196/94504 — only the abstract - the full text could not be fetched — the passage: "More than half (66/114, 58%) of ICs failed to disclose the AI type or its intended use, and 18.4% (n=21) omitted risks entirely."
- Lange F, Mardi S, Binder T, Reich MM, Odorfer T, Volkmann J. (2026). Real-world use and evaluation of a generative AI chatbot for Parkinson's disease information: a prospective observational study. The Lancet Regional Health - Europe. 10.1016/j.lanepe.2026.101866 — only the abstract - the full text could not be fetched — the passage: "AI-assisted triage classified 1803/2035 conversations (88·6%) as good, 224/2035 (11·0%) as partially adequate, and 8/2035 (0·4%) as inadequate."
- Lange F, Mardi S, Binder T, Reich MM, Odorfer T, Volkmann J. (2026). Real-world use and evaluation of a generative AI chatbot for Parkinson's disease information: a prospective observational study. The Lancet Regional Health - Europe. 10.1016/j.lanepe.2026.101866 — only the abstract - the full text could not be fetched — the passage: "Expert review of 45 flagged conversations confirmed five critical events."
- Lange F, Mardi S, Binder T, Reich MM, Odorfer T, Volkmann J. (2026). Real-world use and evaluation of a generative AI chatbot for Parkinson's disease information: a prospective observational study. The Lancet Regional Health - Europe. 10.1016/j.lanepe.2026.101866 — only the abstract - the full text could not be fetched — the passage: "Independent re-review of 100 randomly sampled 'good'-rated conversations identified four (4/100, 4%; 95% CI 1·6-9·8) with clinically critical errors missed by the triage-a conditional false-negative rate within the 'good' stratum, not an overall critical-event rate."
- Lange F, Mardi S, Binder T, Reich MM, Odorfer T, Volkmann J. (2026). Real-world use and evaluation of a generative AI chatbot for Parkinson's disease information: a prospective observational study. The Lancet Regional Health - Europe. 10.1016/j.lanepe.2026.101866 — only the abstract - the full text could not be fetched — the passage: "Confirmed critical events-five flagged, four sampled-spanned three failure classes (knowledge boundary, robustness, and escalation failures) and included an inappropriate memantine recommendation in Parkinson's disease dementia and explicit suicidal ideation that did not trigger the intended emergency response."
- Lange F, Mardi S, Binder T, Reich MM, Odorfer T, Volkmann J. (2026). Real-world use and evaluation of a generative AI chatbot for Parkinson's disease information: a prospective observational study. The Lancet Regional Health - Europe. 10.1016/j.lanepe.2026.101866 — only the abstract - the full text could not be fetched — the passage: "Tightly scoped, retrieval-augmented, patient-facing AI achieved favourable automated triage classifications while still harbouring clinically critical failures invisible to automated quality metrics alone; detecting them required conversation-level evaluation with clinical expert adjudication."
- Salmi L, Norris A, Sarabu C, DesRoches CM. (2026). Americans' views about artificial intelligence in healthcare: 2022-2026. Health Affairs Scholar. 10.1093/haschl/qxag285 — only the abstract - the full text could not be fetched — the passage: "This analysis reveals a significant equity issue: more than 4 in 10 Americans using AI for a health question did so, in part, because they could not access or afford care."
- Salmi L, Norris A, Sarabu C, DesRoches CM. (2026). Americans' views about artificial intelligence in healthcare: 2022-2026. Health Affairs Scholar. 10.1093/haschl/qxag285 — only the abstract - the full text could not be fetched — the passage: "Public concerns about AI are leading to legislative action at the state level, but this activity primarily focuses on oversight, accuracy, and privacy, not on the access gap driving AI use for many Americans."
Xie, Y., Gao, Y., Peng, J. et al. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology. https://doi.org/10.3389/fphar.2026.1958826
Who paid: The authors declared that no financial support was received for this work or its publication, and declared no commercial or financial relationships that could be a conflict of interest; the article does not say who supplied the DeepSeek-V4-pro model, which the authors describe as locally deployed.
Do not take this as professional medical advice.