experiment · Oral and maxillofacial surgery · la publicación, 26 sep 2026 · gratis
Un informe médico en palabras llanas: 82 de cada 100 respuestas correctas frente a 62
En un hospital alemán, versiones de cinco tipos de informe pasadas por ChatGPT mejoraron la lectura y las respuestas, pero todo se hizo en alemán y con cinco informes reales escritos por médicos, sin datos que identificaran al paciente.
Versión breve · la versión detallada sigue, unos 7 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
- Personas sin formación médica que leyeron un informe médico
- Cuántos
- 576 personas
- Dónde
- Maguncia, Alemania
- Cuándo
- datos recogidos del 15 de enero al 2 de junio de 2025
- Tipo de estudio
- experiment
- Quién lo hizo
- Departamento de Cirugía Oral y Maxilofacial, Centro Médico Universitario de Maguncia
- El límite que importa
- Todo se hizo en alemán, en un solo hospital, y no se midió si mejora la salud
Porcentaje de respuestas correctas en preguntas de opción múltiple, contestadas justo después de leer un solo informe; son cinco informes reales en alemán y no prueban mejores decisiones de salud.
Informe simplificado por ChatGPT frente al informe original escrito por el médico
Más aciertos en las preguntas: 82.2% frente a 62.2% en el conjunto
Lectura más fácil: 48.4 frente a 22.9 en la escala Flesch
En patología la mejora fue de 61.2% a 51.3% y pudo deberse al azar
Muestra que es posible.
Tome un informe médico. Puede ser una resonancia magnética, una tomografía, una nota de cirugía, un resultado de patología o una hoja de alta. Está escrito para médicos y muchas veces se entrega al paciente tal cual.
Investigadores del Hospital Universitario de Mainz, en Alemania, tomaron cinco informes así y se los pasaron a ChatGPT. La instrucción era conservar cada dato clínico y escribir para un nivel de lectura de unos 11 o 12 años. Repitieron el ejercicio diez veces por informe y revisaron cada versión contra una lista de contenidos. No faltó información médica en ninguna.
Después, 576 personas sin formación médica leyeron un informe, el original o el simplificado, y contestaron preguntas de opción múltiple. Con los textos simplificados acertaron el 82.2% de las respuestas. Con los originales, el 62.2%.
La resonancia fue donde más se notó: 80.3% de aciertos frente a 53.7%. En los informes de patología la mejora fue menor, de 61.2% a 51.3%, y esa diferencia pudo deberse al azar. El texto más técnico siguió siendo el más difícil.
Todo esto ocurrió en alemán, en un solo departamento de cirugía oral y maxilofacial. Los informes originales eran documentos reales escritos por médicos, sin datos que identificaran al paciente. Las versiones simplificadas sí se hicieron para el estudio. Nadie sabe todavía si un informe en español, inglés o portugués mejoraría lo mismo. Y las preguntas se contestaron justo después de leer. Un puntaje mejor no prueba que el paciente tome mejores decisiones con su salud.
Los propios autores señalan otro punto: ChatGPT puede inventar datos, y aquí solo se comprobó que no lo hiciera bajo instrucciones muy precisas y revisadas. Las versiones del programa cambian con el tiempo.
Mientras eso se aclara, cuando reciba un informe, pida a su médico que se lo explique en palabras sencillas, y pregunte si su clínica ofrece alguna vez una versión en lenguaje claro junto al original.
Qué significa para usted
Cuando reciba un informe de resonancia, tomografía, cirugía, patología o alta, usted puede pedir que le expliquen en palabras sencillas lo que dice. En este estudio alemán, esas versiones claras se leyeron mejor y las personas acertaron más preguntas. Pregunte en su clínica si le pueden dar también una versión en lenguaje llano.
Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4
Quién pagó: Los autores declaran que no recibieron fondos, subvenciones ni otro tipo de apoyo durante la preparación del manuscrito, y declaran no tener conflictos de interés.
No tome esto como consejo médico profesional.
Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1453 palabras · unos 7 minLeerla →Cerrar
Un informe médico que sí se entiende: lo que probó ChatGPT en Alemania, y lo que todavía no
Un estudio con 576 pacientes en Maguncia encontró que las versiones simplificadas por inteligencia artificial se leían mejor y se contestaban mejor. Todo ocurrió en alemán, en un solo hospital.
Hay un momento que casi todos conocemos: el doctor le entrega un papel con el resultado de un estudio, usted lo lee en la sala de espera y no entiende nada. Ese papel es el protagonista de un estudio publicado el 26 de septiembre de 2026 en la revista *Oral and Maxillofacial Surgery*, firmado por Daniel Stephan, Sophia Schumacher, Bilal Al-Nawas, Peer W. Kämmerer y Daniel G. E. Thiem, del Departamento de Cirugía Oral y Maxilofacial del Centro Médico Universitario de Maguncia, en Alemania. Los autores declaran no haber recibido financiamiento y no tener conflictos de interés.
Lo que hicieron fue tomar cinco documentos escritos por médicos: un informe de resonancia magnética (MRI), uno de tomografía computarizada (CT), un reporte quirúrgico, uno de patología y una hoja de alta hospitalaria. Cada uno se reescribió con ChatGPT 4.0 bajo una instrucción fija: conservar toda la información clínica, no inventar nada y usar un lenguaje apto para alguien con una lectura de 11 a 12 años1. Después, 576 personas sin formación médica —289 con la versión simplificada y 287 con la original— leyeron un solo documento, elegido al azar, y contestaron un cuestionario de opinión y cinco preguntas de comprensión. La recolección de datos duró del 15 de enero al 2 de junio de 2025.
Los resultados fueron en la misma dirección en casi todo. Las versiones simplificadas obtuvieron mejores puntajes de legibilidad: 48.4 contra 22.9 en la escala Flesch, donde más alto es más fácil, y 48.3 contra 59.2 en el índice LIX, donde más bajo es más sencillo2. Los pacientes calificaron los textos simplificados como más claros, mejor estructurados y más útiles3. En las preguntas de comprensión, el porcentaje general de respuestas correctas subió de 62.2 a 82.2 por ciento4. En la resonancia magnética fue de 53.7 a 80.3 por ciento; en el reporte quirúrgico, de 75.2 a 92.9; en la hoja de alta, de 61.8 a 88.7; en la tomografía, de 68.8 a 87.95.
Hubo una excepción que el propio estudio reporta. En los informes de patología, la comprensión subió de 51.3 a 61.2 por ciento, pero la diferencia no alcanzó significancia estadística6. Los autores lo atribuyen a que ese tipo de texto usa terminología muy especializada y clasificaciones diagnósticas abstractas que resisten la traducción al lenguaje común. Además, cada documento se simplificó diez veces con la misma instrucción, y en las diez versiones de cada tipo se conservó el contenido clínico completo, sin pérdidas ni invenciones7.
Conviene detenerse en lo que el estudio no hizo. Todo se llevó a cabo en alemán, en un solo departamento hospitalario, lo cual limita la generalización a otros idiomas y sistemas de salud8. No se recabó el nivel educativo de los participantes por el diseño anónimo9. Las preguntas de comprensión fueron de opción múltiple, aplicadas una sola vez, y pueden no captar comprensión profunda ni retención a largo plazo10. No se midió si los pacientes tomaron mejores decisiones, si siguieron su tratamiento o si recordaron la información semanas después.
Piense en el informe como una carta escrita en otro idioma. Los informes médicos están llenos de abreviaturas, términos anatómicos y frases propias de cada especialidad, y para la mayoría de los pacientes esa complejidad es una barrera para entender su estado de salud, sus opciones de tratamiento y las recomendaciones11. El problema no es menor: el artículo describe un estudio en el que muchos pacientes salen de urgencias sin comprender del todo su diagnóstico, su tratamiento o las instrucciones de seguimiento, y no recuerdan los síntomas de alarma ni siguen las recomendaciones de alta, ya sea que la información se dé de palabra o por escrito12.
Los materiales escritos para pacientes suelen estar por encima de lo que se recomienda. El artículo describe un estudio en el que, en oncología y cardiología, por ejemplo, es común que los recursos educativos estén escritos a nivel universitario, aunque las autoridades de salud pública recomiendan un nivel de sexto grado para los materiales dirigidos al público13. La distancia entre quien escribe y quien lee no es un accidente de un solo hospital: es la norma.
La tecnología que se probó aquí no es un programa especializado en medicina. Es un modelo de lenguaje grande, de la familia de ChatGPT, al que se le da una instrucción detallada y produce un texto nuevo. En este caso la instrucción pedía reescribir para lectores sin formación médica, conservar toda la información clínica, no fabricar contenido y mantener la coherencia lógica y estructural del original1. Funciona como un traductor que no cambia el contenido sino la manera de decirlo.
Ya había indicios de que esto podía funcionar. El mismo grupo de investigación había demostrado antes que los informes de radiología simplificados con ChatGPT mejoraban la percepción de claridad, empatía y facilidad de lectura de los pacientes14. Lo que añade este estudio es la ampliación a cinco tipos de documento y, sobre todo, una prueba objetiva de comprensión, no solo la opinión del lector.
Otros trabajos que leímos, y que solo pudimos consultar en su resumen, matizan el entusiasmo. Un ensayo clínico aleatorizado en línea con padres de Estados Unidos, Canadá y el Reino Unido encontró que, a los 45 días, ni el chatbot con GPT-4o ni el de estilo conversacional aumentaban la intención de vacunar contra el VPH, mientras que los materiales de salud pública mantenían un efecto modesto15. En ese mismo ensayo, ninguna intervención aumentó la vacunación reportada16. La lección es que una conversación breve con una máquina no siempre deja huella.
También leímos, solo en resumen, un ensayo pragmático en el que un chatbot automatizado que aplicaba principios de entrevista motivacional logró mejoras en actividad física, consumo de frutas y verduras y compromiso con el cambio a las 12 semanas17. Y un estudio sobre respuestas de plataformas de inteligencia artificial a preguntas de medicamentos encontró que la exactitud y la calidad variaban, y que sigue siendo necesaria la supervisión de especialistas en información de medicamentos1819. Los sistemas sirven para algunas tareas y no para otras.
Así lo leemos nosotros. El patrón que se repite es el de la información que se entrega en un lenguaje que solo domina quien la escribe, y que deja a los demás dependiendo de que alguien se la traduzca. Aquí la traducción la hizo una máquina, y funcionó en el experimento. Pero un informe simplificado que nadie le explica al paciente puede seguir siendo un monólogo: correcto, completo y aun así inútil si nadie le pregunta qué entendió y qué le preocupa.
El propio estudio señala que estos sistemas pueden producir afirmaciones que suenan bien pero son falsas, y que pueden repetir sesgos de los datos con que fueron entrenados. Son advertencias sobre cómo puede fallar este tipo de sistema, no sobre lo que este estudio encontró.
Lo que esto nos hace esperar en casas como la suya es lo siguiente: si la clínica no le entrega la versión sencilla junto con el informe técnico, usted seguirá recibiendo el papel difícil y saliendo sin entender su diagnóstico. Sabremos que nos equivocamos si los pacientes empiezan a pedir la versión simplificada por su cuenta y las clínicas la entregan por defecto. Mientras tanto, hay algo que usted puede hacer.
Al recibir un informe, pida por escrito una versión en lenguaje sencillo y pregunte quién se la puede explicar. Si no existe, pídala como derecho, no como favor. Después de leerla, cuéntesela en voz alta a alguien de confianza y fíjese qué partes no puede explicar: esas son las que hay que preguntar al médico. Y no se conforme con que le entreguen el papel: pida que le pregunten qué entendió y responda con sus palabras. Anote sus dudas antes de la consulta y llévelas escritas.
Volvamos al papel de la sala de espera. La versión que usted tiene en la mano sigue siendo la que escribió el médico, con sus abreviaturas y su vocabulario de especialista. Lo que este estudio muestra es que una versión más clara, hecha por una máquina bajo instrucciones precisas, puede aumentar mucho la probabilidad de que usted acierte lo que dice. No muestra que su clínica vaya a dársela mañana, ni que funcione igual en español, ni que leerla mejore su salud. Muestra que es posible.
La próxima vez que le entreguen un informe, usted ya sabe qué preguntar: ¿hay una versión en lenguaje sencillo, y me la pueden explicar? Esa pregunta, hoy, es lo más concreto que puede llevarse de este estudio.
De dónde sale cada dato de contexto, y cuánto leímos de cada documento
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - el artículo del que trata esta nota — el artículo completo — el pasaje: "For each document, a context-specific prompt instructed ChatGPT to rephrase the text into accessible language suitable for laypersons with a reading level of approximately 11–12 years. The instructions explicitly required preservation of all clinically relevant information, exclusion of fabricated content, and maintenance of the original logical and structural coherence."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Simplified reports achieved significantly higher readability (FRE 48.4 ± 5.0 vs. 22.9 ± 5.2; p < 0.0001; LIX 48.3 ± 3.2 vs. 59.2 ± 2.5; p = 0.004)."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Across all document types, patients rated simplified texts significantly higher in clarity, structure, and usefulness ( p < 0.001), with significantly improved comprehension accuracy (e.g., MRI 80.3% vs. 53.7%; p < 0.001). No loss of medical information was observed."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Overall, participants answered a significantly higher proportion of comprehension questions correctly after reading the simplified texts (Fig. 5 A: 82.2 ± 12.6 % vs. 62.2 ± 10.1 %; p = 0.023; t = 2,782, df = 8; n = 898 / 801)."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Participants demonstrated significantly higher comprehension accuracy for simplified surgical reports (92.9 % vs. 75.2 %; p < 0.001; n = 184 / 157), discharge summaries (88.7 % vs. 61.8 %; p < 0.001; n = 196 / 158), CT reports (87.9 % vs. 68.8 %; p < 0.001; n = 157 / 186), and MRI reports (80.3 % vs. 53.7 %; p < 0.001; n = 188 / 164)."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - el artículo del que trata esta nota — el artículo completo — el pasaje: "For pathology reports, comprehension rates were higher for the simplified version (61.2 % vs. 51.3 %; n = 168 / 136), although this difference was not significant (p = 0.0669)."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - el artículo del que trata esta nota — el artículo completo — el pasaje: "For all text types, each of the ten repetitively simplified versions retained the full medical content of the corresponding original report. The simplification process, performed repeatedly for validation, did not result in any loss of relevant information."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - el artículo del que trata esta nota — el artículo completo — el pasaje: "First, the study was conducted in German within a single linguistic and cultural setting, which may limit the generalizability of the findings to other languages and healthcare systems."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - el artículo del que trata esta nota — el artículo completo — el pasaje: "In addition, information on participants’ educational background was not collected due to the fully anonymous study design. As educational level is closely associated with health literacy, future studies should incorporate this variable to allow a more detailed interpretation of comprehension outcomes."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Moreover, while the multiple-choice comprehension questions provided an objective measure of understanding, they may not fully capture deeper conceptual understanding, inferential reasoning, or long-term information retention."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - el artículo del que trata esta nota — el artículo completo — el pasaje: "However, medical reports are often written in highly specialized language, characterized by abbreviations, anatomical terminology, and discipline-specific phrasing. For most patients, this complexity poses a major barrier to understanding their health status, treatment options, and recommendations."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Research consistently demonstrates that a substantial proportion of patients leave emergency departments without fully understanding their diagnosis, treatment, or follow-up instructions [ 8 , 9 ]. Many patients fail to recall critical warning symptoms or adhere to discharge recommendations, regardless of whether the information is provided verbally or in written form."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - el artículo del que trata esta nota — el artículo completo — el pasaje: "For example, educational resources in oncology and cardiology are commonly written at a university level, despite public health authorities recommending a sixth-grade reading level for patient materials [ 34 ]."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - el artículo del que trata esta nota — el artículo completo — el pasaje: "A previous study from our group demonstrated that ChatGPT-simplified radiology reports significantly improved patient perceptions of clarity, empathy, and readability compared with original AI-generated texts [ 21 ]."
- Sehgal NKR, Rai S, Tonneau M, Agarwal AK, Cappella J, Kornides ML, et al. (2026). Large Language Model Chatbot Conversations vs Public Health Materials and Parental HPV Vaccination Intentions. JAMA Network Open. 10.1001/jamanetworkopen.2026.16822 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "At 45 days, neither chatbot increased intent relative to controls, whereas public health materials maintained modest effects."
- Sehgal NKR, Rai S, Tonneau M, Agarwal AK, Cappella J, Kornides ML, et al. (2026). Large Language Model Chatbot Conversations vs Public Health Materials and Parental HPV Vaccination Intentions. JAMA Network Open. 10.1001/jamanetworkopen.2026.16822 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "No intervention increased self-reported vaccination uptake."
- Wong CHY, Wong MNK, Hou WK, Ting FST, Ng YM, Che Hin Chan C. (2026). Motivational interviewing chatbot improves lifestyle in primary healthcare settings in a pragmatic randomised controlled trial. npj Digital Medicine. 10.1038/s41746-026-02728-w — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Modified intention-to-treat analysis (n = 460) showed significant improvements in the intervention group compared to usual care in physical activity (576 MET-min/week), fruit/vegetable intake (0.27 portions/day), and committed action (0.95 a.u.) at 12 weeks."
- Issa N, Colmenares E, Douglas A, Heindel G, Lee LM, Maxwell E, et al. (2026). Accuracy and Quality Assessment of Artificial Intelligence Platforms and a Hybrid Evidence-Assisted Model in Responding to Drug Information Questions. American Journal of Health-System Pharmacy. 10.1093/ajhp/zxag379 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "AI platforms demonstrated variable accuracy and quality when responding to drug information questions."
- Issa N, Colmenares E, Douglas A, Heindel G, Lee LM, Maxwell E, et al. (2026). Accuracy and Quality Assessment of Artificial Intelligence Platforms and a Hybrid Evidence-Assisted Model in Responding to Drug Information Questions. American Journal of Health-System Pharmacy. 10.1093/ajhp/zxag379 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "Despite strengths in organization and readability, limitations in accuracy, reference identifiability, and evidence synthesis support the continued need for drug information specialist oversight."
Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4
Quién pagó: Los autores declaran que no recibieron fondos, subvenciones ni otro tipo de apoyo durante la preparación del manuscrito, y declaran no tener conflictos de interés.
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.
experiment · Oral and maxillofacial surgery · the paper, 26 Sep 2026 · free
The Same Scan, in Words a Patient Can Use
In one German hospital, 576 people read five medical reports, original or rewritten by ChatGPT. More answered correctly after the plain versions, though the study was in German and the reports were not their own.
Short version · the longer version follows, about 6 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
- people with no medical training reading medical reports
- How many
- 576
- Where
- Mainz, Germany
- When
- between January and June 2025
- Kind of study
- experiment
- Who did it
- University Medical Center Mainz, Germany
- The limit that matters
- All in German, one clinic; not tested in English, Spanish or Portuguese.
Percent of multiple-choice questions answered correctly right after reading, in one German clinic; the pathology gap was small enough to be chance.
Original doctor-written report vs the same report rewritten in plain language
Readers answered more questions correctly after the plain version
Readers rated the plain version clearer, better structured and more useful
No medical information was lost in the plain version
A plain version does not remove that dependence, but it changes what the reader brings to the conversation: questions instead of a blank page.
Take the MRI report. It arrives with the scan, dense with abbreviations and anatomy, written for the doctor who ordered it. The patient holds it and understands little.
Researchers at a university hospital in Mainz, Germany, took five such documents: an MRI report, a CT report, a surgery note, a pathology finding and a discharge summary. They fed each one to ChatGPT with an instruction to keep every clinical fact and write for a reading level of about 11 to 12 years. Then they ran each report through ten times to check nothing was lost. No medical information went missing.
Next, 576 people with no medical training each read one report, the original or the simplified version, chosen at random. They rated how clear it was and answered multiple-choice questions about what they had read.
After the simplified texts, 82.2 percent of the answers were correct. After the originals, 62.2 percent. The discharge summary showed one of the largest gaps: 88.7 percent correct against 61.8 percent. The pathology report barely moved, from 51.3 to 61.2 percent, and that difference could have come from chance.
All of this happened in German, in one department. The reports were anonymized medical documents routinely provided to patients or accessible through electronic health records. So the study cannot say that a clinic in the United States, Canada or Latin America would see the same gains from reports in English, Spanish or Portuguese. The questions were answered right after reading, not weeks later, and better scores do not show that patients make better decisions or fare better in their care.
A report like this one reaches you with a diagnosis on it. One thing to look for: whether your clinic or health service ever offers a plain-language version alongside the original. Until it does, ask your doctor to walk you through the original page by page.
What this means for you
A plain-language copy of your report is something to ask about at your next visit, once the clinic actually offers one. For now, what you can ask is that your doctor go through the original with you, line by line, and explain any word you do not recognize.
Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4
Who paid: The authors state that no funds, grants or other support were received during the preparation of the manuscript, and they declare no competing interests.
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 · 1279 words · about 6 minRead it →Close
A Hospital Report, Rewritten for You
German researchers had ChatGPT turn five kinds of medical reports into plain language. Readers understood more — in that one study, in one language, in one clinic.
Take a discharge summary. It is the piece of paper a hospital hands you on the way out: what was found, what was done, what to watch for, when to come back. It is written by doctors, for doctors. You are given it, and you are not really its reader.
That gap is not a small one. Medical reports are full of abbreviations, anatomical terms and phrasing that belongs to a specialty, and for most patients that complexity is a major barrier to understanding their health status, treatment options and recommendations1. The article describes a study in which patients leave emergency departments without fully understanding their diagnosis, treatment or follow-up instructions, and many cannot recall critical warning symptoms or follow discharge advice, whether it was given spoken or written2. The article describes a study in which patient materials in oncology and cardiology are commonly written at a university level, even though public health authorities recommend a sixth-grade reading level for them3.
Into this gap, a team at the Department of Oral and Maxillofacial Surgery, Facial Plastic Surgery, University Medical Center Mainz, in Germany, put ChatGPT 4.0. They took five anonymized physician-written documents, an MRI report, a CT report, a surgical operation report, a pathology report and a discharge summary, and gave the model a standardized prompt for each: rewrite it for laypeople at a reading level of roughly 11 to 12 years, keep every clinically relevant detail, invent nothing, keep the logical structure4. The study received no funding, and the authors declare no competing interests.
They checked the content first. Each report was simplified ten times with the same prompt, and the versions were compared against a checklist built from the original; across all five document types, the repeated simplifications kept the full medical content and no relevant information was lost5. Readability scores moved in the expected direction: the Flesch Reading Ease score, where higher means easier, went from 22.9 to 48.4, and the LIX score, where higher means harder, fell from 59.2 to 48.36.
Then came the readers. In total 576 people without medical knowledge took part, 289 of them reading a simplified version and 287 the original, randomly assigned, between January and June 2025; people with medical education or healthcare experience were excluded. Across all document types, the simplified texts were rated significantly higher for clarity, structure and usefulness, and comprehension improved: for the MRI report, 80.3 percent of answers were correct against 53.7 percent for the original7. Overall, correct answers rose from 62.2 percent to 82.2 percent8. The gains held for surgical reports (92.9 against 75.2 percent), discharge summaries (88.7 against 61.8 percent) and CT reports (87.9 against 68.8 percent)9.
One report type did not clear the bar. For pathology reports, comprehension was higher with the simplified version, 61.2 percent against 51.3 percent, but the difference was not statistically significant10. Pathology language is dense and abstract, and translating it fully into everyday words is hard.
The limits are as plain as the result. The whole study was conducted in German, in a single linguistic and cultural setting, so how well this would work in Spanish, English or Portuguese is not known11. Participants' educational background was not collected, because the survey was fully anonymous, so its influence on comprehension cannot be assessed12. The comprehension test was a set of multiple-choice questions answered right after reading; such questions may not capture deeper conceptual understanding, reasoning from the text, or what a person remembers weeks later13.
Compare two things the same study set side by side: a report written for colleagues, and the same report restated for a patient. The information is meant to be identical; what changes is who can read it. The article describes the researchers' own earlier work, which had already found that ChatGPT-simplified radiology reports were perceived as clearer, more empathetic and more readable than the original AI-generated texts14. What is new here is that the gain showed up not only in how people felt about the text but in how many questions about its content they answered correctly.
The mechanics are worth seeing. The model was not asked to summarize. It was told to rephrase the whole text, preserve every clinically relevant fact, exclude fabricated content and keep the original order and logic4. That is a narrower job than "answer my medical question," and the authors treat the narrowness as the point: with precise prompts, the model performed a bounded linguistic task reliably, and they argue that standardized prompts and validation protocols will be needed as such tools enter clinical workflows15.
What would have to happen before this reaches your clinic? The authors are explicit that physicians remain responsible for the accuracy and appropriateness of what patients receive, that institutional oversight and transparent review are needed, and that AI tools should be complementary aids rather than replacements for human communication15. They also note that as health systems put more records online, such tools could generate patient-friendly versions of reports with minimal extra effort for clinicians16. That is a possibility they describe, not a service that exists.
The article notes that such tools can still produce fluent but factually wrong statements. Separately, the study's own repeated checks found no fabricated content. The authors also say the model's output may vary with updates, prompt wording or context, so the result is not automatic elsewhere.
Here is how we read it. A report is written once and read by whoever holds it, and the person holding it is often not the person it was written for. When the words belong to a specialty, the reader ends up depending on someone else to say what the document means. A plain version does not remove that dependence, but it changes what the reader brings to the conversation: questions instead of a blank page. We would expect a simplified report to work best in someone's hand at an appointment, not as a substitute for the appointment. We could be wrong: if patients given the plain version asked nothing further and understood it just as well with no one to talk it through, the expectation would fail. What you can do is ask for a plain-language version of any report and check that the findings, the diagnosis and the next steps in it match the technical one; if a line is missing, ask before you act on it.
We also read that a text can feel clear without being understood. Feeling informed and being informed are different things, and the study's own design shows why the distinction matters: it measured both how people rated the text and whether they could answer questions about it, because one does not guarantee the other. So do not treat an easy read as proof of understanding. Before a decision, say the main points back in your own words to your clinician and ask them to correct you.
The discharge summary is still in your hand. What has changed is that you now know it can be rewritten without losing what it says, that readers in one German study understood the rewritten version better, and that nobody has yet shown the same for a report in your language, from your hospital. The next time a clinic hands you a page you cannot read, the question worth asking is a small one: could I have this in plain language, with the original attached?
What would you ask for at the front desk — the plain version, or someone to go through it with you?
Where each piece of context comes from, and how much of it we read
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - the article this story is about — the whole article — the passage: "However, medical reports are often written in highly specialized language, characterized by abbreviations, anatomical terminology, and discipline-specific phrasing. For most patients, this complexity poses a major barrier to understanding their health status, treatment options, and recommendations."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - the article this story is about — the whole article — the passage: "Research consistently demonstrates that a substantial proportion of patients leave emergency departments without fully understanding their diagnosis, treatment, or follow-up instructions [ 8 , 9 ]. Many patients fail to recall critical warning symptoms or adhere to discharge recommendations, regardless of whether the information is provided verbally or in written form."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - the article this story is about — the whole article — the passage: "For example, educational resources in oncology and cardiology are commonly written at a university level, despite public health authorities recommending a sixth-grade reading level for patient materials [ 34 ]."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - the article this story is about — the whole article — the passage: "For each document, a context-specific prompt instructed ChatGPT to rephrase the text into accessible language suitable for laypersons with a reading level of approximately 11–12 years. The instructions explicitly required preservation of all clinically relevant information, exclusion of fabricated content, and maintenance of the original logical and structural coherence."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - the article this story is about — the whole article — the passage: "For all text types, each of the ten repetitively simplified versions retained the full medical content of the corresponding original report. The simplification process, performed repeatedly for validation, did not result in any loss of relevant information."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - the article this story is about — the whole article — the passage: "Simplified reports achieved significantly higher readability (FRE 48.4 ± 5.0 vs. 22.9 ± 5.2; p < 0.0001; LIX 48.3 ± 3.2 vs. 59.2 ± 2.5; p = 0.004)."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - the article this story is about — the whole article — the passage: "Across all document types, patients rated simplified texts significantly higher in clarity, structure, and usefulness ( p < 0.001), with significantly improved comprehension accuracy (e.g., MRI 80.3% vs. 53.7%; p < 0.001). No loss of medical information was observed."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - the article this story is about — the whole article — the passage: "Overall, participants answered a significantly higher proportion of comprehension questions correctly after reading the simplified texts (Fig. 5 A: 82.2 ± 12.6 % vs. 62.2 ± 10.1 %; p = 0.023; t = 2.782, df = 8; n = 898 / 801)."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - the article this story is about — the whole article — the passage: "Participants demonstrated significantly higher comprehension accuracy for simplified surgical reports (92.9 % vs. 75.2 %; p < 0.001; n = 184 / 157), discharge summaries (88.7 % vs. 61.8 %; p < 0.001; n = 196 / 158), CT reports (87.9 % vs. 68.8 %; p < 0.001; n = 157 / 186), and MRI reports (80.3 % vs. 53.7 %; p < 0.001; n = 188 / 164)."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - the article this story is about — the whole article — the passage: "For pathology reports, comprehension rates were higher for the simplified version (61.2 % vs. 51.3 %; n = 168 / 136), although this difference was not significant (p = 0.0669)."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - the article this story is about — the whole article — the passage: "First, the study was conducted in German within a single linguistic and cultural setting, which may limit the generalizability of the findings to other languages and healthcare systems."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - the article this story is about — the whole article — the passage: "In addition, information on participants’ educational background was not collected due to the fully anonymous study design. As educational level is closely associated with health literacy, future studies should incorporate this variable to allow a more detailed interpretation of comprehension outcomes."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - the article this story is about — the whole article — the passage: "Moreover, while the multiple-choice comprehension questions provided an objective measure of understanding, they may not fully capture deeper conceptual understanding, inferential reasoning, or long-term information retention."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - the article this story is about — the whole article — the passage: "A previous study from our group demonstrated that ChatGPT-simplified radiology reports significantly improved patient perceptions of clarity, empathy, and readability compared with original AI-generated texts [ 21 ]."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - the article this story is about — the whole article — the passage: "Physicians remain ultimately responsible for the accuracy and appropriateness of the information provided to patients, underscoring the need for institutional oversight and transparent review mechanisms. AI tools such as ChatGPT should thus be regarded as complementary aids rather than replacements for human communication."
- Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4 - the article this story is about — the whole article — the passage: "As healthcare systems continue to expand digital access to medical records, the integration of AI tools such as ChatGPT into documentation workflows could enable the automated generation of patient-friendly report versions with minimal additional effort for clinicians."
Stephan, D., Schumacher, S., Al-Nawas, B. et al. (2026). Improving patient communication and understanding of medical reports through text simplification with ChatGPT: a comparative study. Oral and Maxillofacial Surgery. https://doi.org/10.1007/s10006-026-01643-4
Who paid: The authors state that no funds, grants or other support were received during the preparation of the manuscript, and they declare no competing interests.
Do not take this as professional medical advice.
The findings of other studies mentioned here come from this document; those studies were not opened one by one.

