Most Latin American Emergency Doctors in This Survey Use AI at Work. Almost None of Their Hospitals Have Rules for It.

survey · International journal of critical illness and injury science · la publicación, 1 jul 2026 · gratis

Sólo 7 de cada 100 médicos encuestados dijeron que su hospital tiene reglas claras para la inteligencia artificial

Una encuesta a 123 médicos de 16 países encontró que 78 de cada 100 lo usaron en el trabajo y que sólo 7 de cada 100 dijeron que su institución tiene normas claras. No puede decir qué pasa en su país ni si algún paciente salió mejor o peor.

Versión breve · la versión detallada sigue, unos 7 min

Pregunte al chat de weeklyAI

Pregúnteme por este estudio: a quiénes se estudió, qué encontró y qué no dice.

Las conversaciones se guardan mientras exista weeklyAI, para mejorar la publicación. Se responde en el idioma en que usted escribe.

El estudio, de un vistazo
Quiénes: Profesionales de medicina de urgencias de América Latina
Cuántos: 123 encuestados de 16 países
Dónde: 16 países de América Latina
Cuándo: Entre abril y octubre de 2025
Tipo de estudio: survey
Quién lo hizo: Universidad Nacional Pedro Henríquez Ureña, Santo Domingo
El límite que importa: No es representativa de ningún país ni de la región; no midió resultados en pacientes.

¿Su institución tiene políticas claras sobre el uso de IA?

Sí tiene políticas claras7.4%

No existe ninguna orientación77%

No sabe15.6%

Porcentaje de los 123 profesionales de urgencias encuestados según lo que reportaron sobre las políticas de IA de su institución; la encuesta no es representativa de ningún país ni de la región.

La pregunta que puede llevarse a casa es sencilla: ¿quién revisa lo que la máquina sugiere, y dónde queda escrito que alguien lo revisó?
Lectura de weeklyAI

Así podría verse · ilustración generada por weeklyAI.watch, no es una fotografía

Usted llega a una sala de urgencias en América Latina. El médico que lo atiende puede estar usando ChatGPT para redactar notas clínicas.

Una encuesta hecha entre abril y octubre de 2025 a 123 profesionales de urgencias de 16 países latinoamericanos encontró que 78 de cada 100 habían usado inteligencia artificial generativa en su trabajo. La herramienta más usada fue ChatGPT, con 81 de cada 100. La usaron sobre todo para revisar investigaciones y literatura médica (66 de cada 100), preparar material educativo (59 de cada 100) y redactar notas clínicas (41 de cada 100).

Los mismos médicos dijeron que casi nadie les ha puesto reglas. Sólo 7 de cada 100 dijeron que su institución tiene políticas claras sobre el uso de IA. 77 de cada 100 dijeron que no existe ninguna guía, y 16 de cada 100 no sabían.

La encuesta no es un censo. Se armó con contactos profesionales y redes sociales, así que probablemente participaron médicos con más acceso a la tecnología y más interés en el tema. No representa a todos los médicos de urgencias de la región, ni de ningún país en particular. No puede decirle qué hacen los médicos de su ciudad. Y no midió si la IA ayudó o perjudicó a algún paciente: no revisó historias clínicas ni resultados de salud.

Más de la mitad de los encuestados señaló dos riesgos: la información falsa o inventada por la herramienta y el debilitamiento del juicio clínico, con 56 de cada 100 cada uno. 54 de cada 100 mencionaron sesgos en los algoritmos, y 46 de cada 100, responsabilidad legal.

88 de cada 100 dijeron que querían capacitación formal, y 70 de cada 100 la preferían en línea. 66 de cada 100 apoyarían que la IA se integrara a las guías clínicas oficiales.

Los autores trabajan en la Universidad Nacional Pedro Henríquez Ureña, en Santo Domingo, y no recibieron financiamiento. Es un retrato de un momento, no una predicción.

Cuando vuelva a sentarse en una sala de urgencias, pregunte qué normas existen en ese hospital sobre el uso de estas herramientas. Esa pregunta, hoy, casi nadie la ha respondido.

Qué significa para usted

Usted puede preguntar en su hospital qué reglas existen sobre el uso de estas herramientas, porque la encuesta encontró que casi ninguna institución las tiene. No concluya que la inteligencia artificial mejoró o empeoró la atención: el estudio no revisó historias clínicas ni resultados de pacientes. Lo que sí aparece es que muchos médicos quieren capacitación formal sobre el tema.

Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26

Quién pagó: Los autores declaran que no hubo financiamiento (Funding Statement: Nil) ni conflictos de interés; el artículo no indica que ningún financiador, empresa o institución haya tenido injerencia en el estudio ni haya prestado equipos o programas.

No tome esto como consejo médico profesional.

Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1366 palabras · unos 7 minLeerla →Cerrar

En una encuesta a 123 médicos de urgencias, 78 de cada 100 dijeron que usan la IA en el trabajo

Una encuesta a 123 profesionales de 16 países encontró que tres de cada cuatro la usan, mientras apenas un 7.4% dice que su institución tiene políticas claras al respecto.

Un estudio publicado en el *International Journal of Critical Illness and Injury Science* encuestó a 123 profesionales de medicina de urgencias de 16 países de América Latina. La encuesta fue voluntaria, anónima, en línea y en español, y se aplicó entre abril y octubre de 2025. De los encuestados, el 78.3% dijo haber usado herramientas de inteligencia artificial generativa en su práctica profesional, y el 76.9% reportó un nivel de familiaridad de moderado a alto con estas herramientas¹. La plataforma más usada fue ChatGPT, mencionada por el 81.0%, seguida de Gemini (31.9%), Copilot (24.1%) y Perplexity AI (11.2%)².

Los usos principales que reportaron fueron la investigación y revisión de literatura (66.1%), la generación de material educativo (59.1%), la documentación clínica y resúmenes (40.9%), el apoyo al razonamiento diagnóstico (38.3%) y la comunicación o traducción con pacientes (32.2%)³. Casi la mitad valoró la utilidad de la herramienta como alta, y el 51.7% notó mejoras en su productividad. Al mismo tiempo, los riesgos más mencionados fueron la información errónea y el deterioro del juicio clínico (56.2% cada uno), el sesgo algorítmico (53.7%) y la responsabilidad legal (46.3%)⁴. Cabe subrayar que estos son porcentajes de un grupo de 123 encuestados, no de todos los médicos de urgencias de la región.

El hallazgo más llamativo tiene que ver con las reglas: solo el 7.4% de los encuestados dijo que su institución tiene políticas claras sobre el uso de IA, mientras que el 77.0% indicó que no existe ninguna orientación institucional y el 15.6% no estaba seguro⁵. A la vez, el 88.4% mostró interés en recibir formación formal sobre IA, con preferencia por formatos virtuales (69.7%) o híbridos (26.2%)⁶, y el 65.6% apoyó integrar estas herramientas en protocolos o guías clínicas oficiales⁷.

Conviene detenerse en los límites de este trabajo. La muestra se reclutó por redes profesionales y redes sociales, lo que probablemente sobrerrepresenta a médicos con mayor conexión digital, urbanos y afiliados a instituciones, y puede inflar las cifras de adopción y familiaridad⁸. Además, aunque abarca 16 países, no es estadísticamente representativa de ningún país en particular ni de la región en su conjunto, y los tamaños por país fueron pequeños⁹. El diseño es transversal: no permite establecer causas ni seguir la evolución en el tiempo. Tampoco midió resultados clínicos objetivos, la calidad real de lo que produce la IA ni indicadores de seguridad del paciente. Los autores declaran no tener conflictos de interés y no haber recibido financiamiento.

Para entender qué es esta tecnología: cuando usted escribe una pregunta o un fragmento de texto en una herramienta como ChatGPT, el sistema no busca la respuesta en un libro ni consulta una base de datos médica verificada. Lo que hace es generar, palabra por palabra, la continuación más probable según los patrones que aprendió de enormes cantidades de texto. Por eso puede producir frases que suenan completamente seguras y correctas aunque el dato sea inventado: a eso se le llama alucinación. Esa forma de funcionar explica por qué los encuestados, a la vez que la usan, desconfían: la herramienta es útil para redactar, resumir o buscar bibliografía, pero no tiene manera de saber si lo que dice es verdad³⁴.

Según el resumen de ese estudio —pudimos leer solo el resumen, el texto completo está detrás de una suscripción—, la clasificación automática consideró "buenas" el 88.6% de las conversaciones, pero una revisión independiente de 100 conversaciones elegidas al azar dentro de ese grupo "bueno" encontró cuatro con errores clínicamente críticos que la clasificación automática no detectó¹⁰¹¹¹². El resumen señala que entre las conversaciones revisadas se encontraron errores clínicamente críticos¹³. El mismo resumen señala que detectar esas fallas exigió revisión conversación por conversación con criterio clínico experto, algo que las métricas automáticas por sí solas no logran¹⁴.

En Europa, según el resumen de una encuesta de la Organización Mundial de la Salud —también leímos solo el resumen—, de los 50 Estados miembros que respondieron, el 8% tiene una estrategia de IA específica para salud y el 14% está desarrollándola; menos del 10% ha desarrollado estándares de responsabilidad legal para la IA, y solo 14 miembros han emitido guías sobre las implicaciones éticas de usar IA en salud¹⁵¹⁶¹⁷. El resumen concluye que la gobernanza de la IA en salud está poco desarrollada en la Región Europea de la OMS¹⁸. La encuesta estuvo disponible para los 53 Estados miembros entre junio de 2024 y marzo de 2025¹⁹.

Según el resumen de una revisión sobre IA en diabetes pediátrica —leímos solo el resumen—, los sistemas automáticos de administración de insulina mejoran el control de la glucosa, aumentan el tiempo en rango y reducen la carga de manejar la enfermedad, y hoy se recomiendan como estándar de cuidado para niños con diabetes tipo 1²⁰²¹. Esa misma revisión advierte que la exposición a la IA, tanto médica como de consumo, plantea desafíos de desarrollo, privacidad, sesgo algorítmico y acceso desigual, y que los efectos de largo plazo siguen entendiéndose mal²²²³²⁴.

Así lo leemos nosotros. Cuando las instituciones no dan reglas claras, la gente resuelve sus necesidades por vías informales, y esas prácticas se vuelven costumbre antes de que alguien las regule. Es probable que los médicos de urgencias sigan usando estas herramientas sin esperar políticas hospitalarias, y que la costumbre se consolide como parte normal del trabajo, incluso si después aparecen normas que no coincidan con lo que ya hacen. Sabríamos que nos equivocamos si viéramos que los médicos dejan de usar IA al notar que no hay reglas, o que esperan instrucciones formales antes de adoptarla. Lo que usted puede hacer con esto es concreto: pregunte en su hospital o clínica si existe alguna política sobre el uso de IA y, si no la hay, pida que se discuta antes de que la práctica se generalice sin control.

Una segunda cosa que nos llama la atención: aunque los médicos expresen entusiasmo, siguen trabajando dentro de plataformas y flujos que no controlan, y sus quejas o adaptaciones no bastan para cambiar la falta de supervisión institucional. Sabríamos que nos equivocamos si viéramos que los propios médicos logran imponer límites o exigir rendición de cuentas a los proveedores de IA sin depender de que las instituciones lo hagan. Y hay algo más que le toca directamente como paciente: cuando un algoritmo es opaco y sus datos circulan sin que usted sepa cómo se usan, la relación entre quien lo atiende y quien controla la herramienta se vuelve desigual. Usted puede preguntar a su empleador o al proveedor de IA qué datos se recogen, dónde se guardan y quién puede verlos, y pedir que esa información esté por escrito.

¿Qué tendría que pasar para que esto cambie? Los propios autores proponen un modelo por etapas: primero reconocer formalmente el uso informal que ya existe y definir su alcance; luego construir capacidades con educación estructurada y marcos de competencia; después desarrollar políticas institucionales y nacionales; seguir con pilotos monitoreados y auditables; y terminar con evaluación y escalamiento que mida seguridad, calidad y equidad de forma continua²⁵. También sugieren que los cuerpos profesionales regionales emitan guías de consenso, estándares mínimos y principios éticos adaptados a la atención de urgencias²⁶, con supervisión humana permanente, carácter consultivo y no prescriptivo de lo que produce la IA, documentación transparente del razonamiento asistido por IA y mecanismos para reportar y aprender de los errores²⁷.

Aquí está lo que esto le deja a usted como paciente o familiar. La próxima vez que entre a una sala de urgencias, o que acompañe a alguien, tiene derecho a preguntar si el equipo usa herramientas de IA para redactar notas, buscar información o apoyar decisiones, y si el hospital tiene alguna política al respecto. No se trata de desconfiar del médico que lo atiende: se trata de que usted sepa quién responde si algo sale mal. La pregunta que puede llevarse a casa es sencilla: ¿quién revisa lo que la máquina sugiere, y dónde queda escrito que alguien lo revisó?

De dónde sale cada dato de contexto, y cuánto leímos de cada documento

Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - el artículo del que trata esta nota — el artículo completo — el pasaje: "More than three-quarters of respondents (78.3%) had used generative AI tools in professional practice."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - el artículo del que trata esta nota — el artículo completo — el pasaje: "ChatGPT was the most commonly used platform (81.0%), followed by Gemini (31.9%), Copilot (24.1%), and Perplexity AI (11.2%)."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The most frequently reported application was research and literature review (66.1%), followed by educational material generation (59.1%), clinical documentation and summarization (40.9%), diagnostic reasoning support (38.3%), and patient communication or translation (32.2%)."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The most frequently cited perceived risks were misinformation or errors and erosion of clinical judgment (both 56.2%), followed by algorithmic bias (53.7%), legal liability (46.3%), and data confidentiality concerns (33.9%)."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Only 7.4% of respondents reported that their institution had clear policies governing AI use, while 77.0% indicated no institutional guidance existed, and 15.6% were uncertain."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Interest in formal AI training was very high (88.4%), with a strong preference for virtual (69.7%) or hybrid (26.2%) formats."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Importantly, 65.6% of respondents supported integration of generative AI into clinical protocols or official guidelines, indicating readiness for structured, regulated adoption once appropriate safeguards are in place."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Second, the convenience sampling strategy, utilizing professional networks and social media, likely overrepresents digitally engaged, urban, and institutionally affiliated clinicians, potentially inflating adoption and familiarity estimates relative to the broader population of emergency medicine professionals in Latin America."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Third, the sample, while geographically diverse across 16 countries, is not statistically representative of any individual country or the region as a whole, and the small per-country sample sizes limited country-level subgroup analysis."
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 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "We conducted a prospective, observational, conversation-level evaluation of jAImes, a retrieval-augmented AI information system for Parkinson's disease deployed publicly in Germany, across its first 129 days of operation (Nov 11, 2025-Mar 20, 2026), analysing all 2035 conversations (6146 messages)."
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 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "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. 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 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "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 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "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 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "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."
Adib K, Letchford N, Dunning HE, Salama N, Tolias Y, De Barros J, et al. (2026). Governance of artificial intelligence for health systems, WHO European Region. Bulletin of the World Health Organization. 10.2471/blt.25.294978 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Of the 50 Member States responding to the survey, 8% (4/50) have a health-specific AI strategy and 14% (7) are developing one."
Adib K, Letchford N, Dunning HE, Salama N, Tolias Y, De Barros J, et al. (2026). Governance of artificial intelligence for health systems, WHO European Region. Bulletin of the World Health Organization. 10.2471/blt.25.294978 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Less than 10% (4) of Member States have developed liability standards for AI or guidance on the application of existing liability standards."
Adib K, Letchford N, Dunning HE, Salama N, Tolias Y, De Barros J, et al. (2026). Governance of artificial intelligence for health systems, WHO European Region. Bulletin of the World Health Organization. 10.2471/blt.25.294978 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Only 14 Member States have issued guidelines to address the ethical implications of using AI in health or across sectors."
Adib K, Letchford N, Dunning HE, Salama N, Tolias Y, De Barros J, et al. (2026). Governance of artificial intelligence for health systems, WHO European Region. Bulletin of the World Health Organization. 10.2471/blt.25.294978 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "In the WHO European Region, governance of AI in health care is underdeveloped."
Adib K, Letchford N, Dunning HE, Salama N, Tolias Y, De Barros J, et al. (2026). Governance of artificial intelligence for health systems, WHO European Region. Bulletin of the World Health Organization. 10.2471/blt.25.294978 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "The survey was available to all 53 Member States between June 2024 and March 2025."
Khalil H, Lynch JL, Sowan A, Ortega A, Zaitoon H. (2026). Artificial intelligence in pediatric diabetes: clinical benefits, developmental risks, and ethical challenges. European Journal of Pediatrics. 10.1007/s00431-026-07442-z — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Medical AI technologies, particularly AID systems, improve glycemic control, increase time in range, and reduce diabetes-management burden."
Khalil H, Lynch JL, Sowan A, Ortega A, Zaitoon H. (2026). Artificial intelligence in pediatric diabetes: clinical benefits, developmental risks, and ethical challenges. European Journal of Pediatrics. 10.1007/s00431-026-07442-z — solo el resumen - no se pudo obtener el texto completo — el pasaje: "AI-driven diabetes technologies, particularly automated insulin delivery systems, improve 69 glycemic outcomes and are now recommended as the standard of care for children with 70 type 1 diabetes."
Khalil H, Lynch JL, Sowan A, Ortega A, Zaitoon H. (2026). Artificial intelligence in pediatric diabetes: clinical benefits, developmental risks, and ethical challenges. European Journal of Pediatrics. 10.1007/s00431-026-07442-z — solo el resumen - no se pudo obtener el texto completo — el pasaje: "However, dual AI exposure introduces a complex spectrum of developmental, psychological, ethical, and governance-related challenges, including increased screen exposure, continuous surveillance, caregiver anxiety, altered psychosocial development, data privacy vulnerabilities, algorithmic bias, and inequitable access."
Khalil H, Lynch JL, Sowan A, Ortega A, Zaitoon H. (2026). Artificial intelligence in pediatric diabetes: clinical benefits, developmental risks, and ethical challenges. European Journal of Pediatrics. 10.1007/s00431-026-07442-z — solo el resumen - no se pudo obtener el texto completo — el pasaje: "The increasing use of generative AI introduces additional potential concerns related to misinformation, psychological dependence, and altered health behaviors."
Khalil H, Lynch JL, Sowan A, Ortega A, Zaitoon H. (2026). Artificial intelligence in pediatric diabetes: clinical benefits, developmental risks, and ethical challenges. European Journal of Pediatrics. 10.1007/s00431-026-07442-z — solo el resumen - no se pudo obtener el texto completo — el pasaje: "However, the broader developmental, psychosocial, ethical, and societal effects of sustained exposure to medical and consumer AI remain insufficiently understood."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - el artículo del que trata esta nota — el artículo completo — el pasaje: "A staged implementation model could guide this transition: (1) a recognition phase that formally acknowledges existing informal use and defines its scope; (2) a capacity-building phase providing structured education and competency frameworks; (3) a governance phase developing institutional and national policy; (4) a pilot implementation phase with monitored, auditable use cases; and (5) an evaluation and scaling phase with continuous assessment of safety, quality, and equity outcomes."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Regional professional bodies could play a convening role by issuing consensus guidance, minimum standards, and ethical principles tailored to emergency care."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Such guidance should emphasize human-in-the-loop oversight, advisory – not prescriptive – status of AI outputs, transparent documentation of AI-assisted reasoning, and mechanisms for error reporting and learning."

No tome esto como consejo médico profesional.

survey · International journal of critical illness and injury science · the paper, 1 Jul 2026 · free

Most Latin American Emergency Doctors in This Survey Use AI at Work. Almost None of Their Hospitals Have Rules for It.

A survey of 123 emergency medicine professionals in 16 countries found 78.3% used AI tools on the job. The survey cannot say what doctors in your country do, or whether AI helped or hurt anyone.

Short version · the longer version follows, about 7 min

Ask the weeklyAI chat

Ask me about this study: who was studied, what it found, and what it does not say.

Conversations are saved for as long as weeklyAI exists, to improve the publication. Answers come in the language you write in.

The study at a glance
Who: emergency medicine professionals in Latin America
How many: 123 respondents
Where: 16 countries across Latin America
When: April 2025 to October 2025
Kind of study: survey
Who did it: Universidad Nacional Pedro Henríquez Ureña, Dominican Republic
The limit that matters: Volunteers recruited online, likely the most tech-comfortable; not representative of any country.

What the surveyed doctors said about AI rules at their workplace

Institution had clear AI policies7.4%

No institutional guidance existed77%

Did not know15.6%

Shares of the 123 surveyed emergency medicine professionals describing their own institution's AI rules; the survey cannot say what rules exist in any country.

When a tool spreads faster than the rules meant to govern it, the people who operate the tool end up holding responsibility for harms they have no power to prevent.
weeklyAI's reading

How it could look · illustration generated by weeklyAI.watch, not a photograph

The survey found that about 78 in 100 of the doctors who answered had used AI tools at work. ChatGPT was the tool most of them named.

A survey asked 123 emergency medicine professionals across 16 Latin American countries about this. The researchers, from Universidad Nacional Pedro Henríquez Ureña in the Dominican Republic, collected answers online from April to October 2025. The doctors answered in Spanish.

Of those 123, about 78 in 100 said they had used generative AI at work. ChatGPT was the tool most of them named. They used it most for research and reading up on medical literature (about 66 in 100), for making teaching material (about 59 in 100), and for writing clinical notes and summaries (about 41 in 100).

Only about 7 in 100 said their hospital or clinic had clear rules for using AI. Another 77 in 100 said there was no guidance at all. The doctors who answered came from public, private and mixed workplaces, and most had more than ten years on the job.

The survey cannot tell you what most doctors in your country do. It cannot tell you what the rules are where you live. Only 123 people answered, they were recruited through professional networks and social media, and the researchers say such volunteers tend to be the ones most comfortable with technology. No single country had enough answers to speak for itself. The survey also did not check whether any patient was helped or harmed, because it did not look at patient outcomes at all.

More than half of the doctors who answered said they worried about false or made-up information from these tools (about 56 in 100), and the same share worried that doctors would lean too little on their own judgment (about 56 in 100). A little over half named bias in the software (about 54 in 100), and about 46 in 100 named legal responsibility. Most of them, about 88 in 100, said they wanted formal training.

So the question to carry is this: what, if anything, does your hospital, your health ministry, or your country's law actually say about a machine helping decide your care?

What this means for you

What the survey shows is that a hospital's own rules, not the country, are what you can check up on. Ask your clinic whether it has a written policy on AI tools, and who is responsible when one is wrong. Only 9 of the 123 doctors said their hospital had a written policy on AI tools.

Who paid: The authors declare no funding (Funding Statement: Nil) and no conflicts of interest; the article does not say that any funder, company, or institution had a say in the study or lent equipment or software.

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 · 1426 words · about 7 minRead it →Close

Latin American Emergency Doctors Are Already Using AI. Almost No Hospital Has a Rule for It.

A survey of 123 clinicians in 16 countries found ChatGPT in the room — and no policy on the wall.

In a survey of 123 emergency medicine professionals in 16 Latin American countries, most reported using generative AI tools in their work, and very few said their institution had rules for it. That is the central finding of a cross-sectional survey published in the *International Journal of Critical Illness and Injury Science* by Amado Alejandro Baez, Claudia Acra and Brailin Difo, all of Universidad Nacional Pedro Henríquez Ureña in Santo Domingo, Dominican Republic. The study was approved by the ethics review board of Universidad Nacional Pedro Henríquez Ureña, which determined it was of minimal-to-no risk and exempt from full review. No funding was received, and the authors declared no conflicts of interest.

The survey was voluntary, anonymous and online, run in Spanish from April to October 2025 and pilot-tested with five emergency clinicians before it went out. Responses were analyzed from 123 people practicing in 16 countries. Most worked in public or mixed public–private settings (71.3 percent) and most had more than ten years of experience (61.3 percent). The largest groups practiced in the Dominican Republic (24.8 percent), Mexico (about 19.5 percent), Colombia (14.2 percent) and Bolivia (6.2 percent).

More than three-quarters of respondents — 78.3 percent — said they had used generative AI tools in professional practice. ChatGPT was the platform most reported, at 81.0 percent, followed by Gemini (31.9 percent), Copilot (24.1 percent) and Perplexity AI (11.2 percent). The most common uses were research and literature review (66.1 percent), educational material (59.1 percent), clinical notes and summaries (40.9 percent), diagnostic reasoning support (38.3 percent) and patient communication or translation (32.2 percent). Radiology interpretation support came last, at 16.5 percent. Nearly half rated the overall usefulness as high (49.6 percent), and 51.7 percent reported a positive impact on productivity and work quality.

Their worries were not small. Misinformation or errors and erosion of clinical judgment were each cited by 56.2 percent, algorithmic bias by 53.7 percent, legal liability by 46.3 percent and data confidentiality by 33.9 percent. The biggest barrier to using these tools was lack of formal training, named by 54.2 percent. On governance, the picture was close to blank: 7.4 percent said their institution had clear policies on AI use, 77.0 percent said no guidance existed, and 15.6 percent did not know. Demand for change was loud: 88.4 percent wanted formal training, most of them preferring it online (69.7 percent) or in a mix of online and in person (26.2 percent), and 65.6 percent supported writing AI into official clinical protocols.

The study's own limits matter here. The authors say the convenience sample — recruited through professional societies, hospital networks and social media — likely overrepresents digitally engaged, urban, institutionally affiliated clinicians, which may inflate the adoption and familiarity figures. The sample is not statistically representative of any single country or of the region. Small numbers per country prevented country-level analysis. Because the survey was taken at one moment, it cannot show cause and effect, and it did not measure whether any patient was helped or harmed, or whether the AI answers were any good. The questionnaire was pilot-tested but never formally validated. The authors also note that country of origin was not significantly associated with familiarity, frequency of use, platform choice or policy status.

What these clinicians are using is a large language model: software trained on enormous amounts of text that produces new text by predicting what words tend to follow others. It does not look anything up unless it is built to, and it can generate confident, fluent, wrong answers — the failure the field calls hallucination, and the one the respondents themselves named most often as a risk¹. That is why the study's authors, in their recommendations, insist that AI output be advisory rather than prescriptive, that a human stay in the loop, and that AI-assisted reasoning be documented transparently, with a way to report errors and learn from them².

This is not only a Latin American story. A study of a publicly deployed patient-facing chatbot for Parkinson's disease in Germany — we could read only the summary; the full paper is behind a subscription — followed the system across its first 129 days and all 2,035 conversations³. Automated triage rated 88.6 percent of those conversations as good, 11.0 percent as partially adequate and 0.4 percent as inadequate⁴. Then clinicians re-read 100 conversations that the automated screen had called good, and found four containing clinically critical errors the screen had missed⁵. The confirmed problems included a wrong drug recommendation⁶. The authors' conclusion is blunt: the failures were invisible to automated quality measures and only surfaced when clinicians read the conversations one by one⁷. For emergency medicine, where a wrong answer arrives in minutes rather than days, that is the risk the surveyed doctors appear to already sense.

Compare that with where the rules stand. The survey was open to all 53 member states between June 2024 and March 2025¹⁰. Its authors describe AI governance in health care in that region as underdeveloped¹¹ and call for adaptive legal and policy mechanisms¹². So the gap the Latin American clinicians report is not a regional peculiarity; it is the international norm, and the Latin American numbers — 7.4 percent with policies — sit inside it rather than outside¹³.

The trade-off is visible in the numbers themselves. These tools appear to buy time and reach: literature synthesis, teaching material, note-writing, translation for patients whose language the clinician may not share¹⁴. What is given up, for now, is any assurance about what happens when the output is wrong. The clinician who pasted a summary into a chart is the one who signed it. And the study's authors point out where that leaves responsibility in a region where the guidance is absent: on individual clinicians, in high-acuity settings, with documented risks of fabrication¹⁵. What would have to happen next, in their account, is staged: first formally acknowledge that informal use is already happening; then build capacity with structured education and competency frameworks; then develop institutional and national policy; then run monitored, auditable pilots; then evaluate and scale while watching safety, quality and equity¹⁶. They suggest regional professional bodies could convene the process and issue consensus guidance with minimum standards¹⁷.

Here is how we read it. When a tool spreads faster than the rules meant to govern it, the people who operate the tool end up holding responsibility for harms they have no power to prevent. That is the shape of what this survey describes: thousands of decisions a day, made with a tool nobody trained them on, under policies that mostly do not exist. If that holds, we would expect the first serious AI-related error in a Latin American emergency department to land on the clinician who typed the prompt, not on the hospital that never wrote a rule and not on the company that sold the model. We would also expect the platforms in use to have been built for other languages, other diseases and other health systems, with their errors falling on patients who had no part in shaping them. We could be wrong about both: we would need to see hospitals or health authorities already assigning responsibility for AI-related errors to the institution or the tool provider rather than the individual clinician, or clinicians being formally protected when they report an AI-related problem.

As we read it, rights in this area are not handed down; they are won when ordinary people learn the facts and press those in authority to justify their decisions. Nothing in this survey suggests Latin American clinicians are waiting passively — the demand for training and the support for written protocols are the opposite of that¹⁸¹⁹. But the institutions above them have not answered yet. That is where a reader comes in, whether they are a doctor or a patient. Ask your hospital or clinic, in writing, who is responsible when an AI tool gives wrong information. Ask whether the tool was tested for your language and your condition, and who to contact if you believe it got something wrong. If you belong to a professional association, a union or a community group, ask your health ministry what AI rules exist and ask them to publish the answer. The answer may be that there are none — which is itself the most useful thing you can learn before you rely on one.

Where each piece of context comes from, and how much of it we read

Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - the article this story is about — the whole article — the passage: "The most frequently cited perceived risks were misinformation or errors and erosion of clinical judgment (both 56.2%), followed by algorithmic bias (53.7%), legal liability (46.3%), and data confidentiality concerns (33.9%)."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - the article this story is about — the whole article — the passage: "Such guidance should emphasize human-in-the-loop oversight, advisory – not prescriptive – status of AI outputs, transparent documentation of AI-assisted reasoning, and mechanisms for error reporting and learning."
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: "We conducted a prospective, observational, conversation-level evaluation of jAImes, a retrieval-augmented AI information system for Parkinson's disease deployed publicly in Germany, across its first 129 days of operation (Nov 11, 2025-Mar 20, 2026), analysing all 2035 conversations (6146 messages)."
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. 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."
Adib K, Letchford N, Dunning HE, Salama N, Tolias Y, De Barros J, et al. (2026). Governance of artificial intelligence for health systems, WHO European Region. Bulletin of the World Health Organization. 10.2471/blt.25.294978 — only the abstract - the full text could not be fetched — the passage: "Of the 50 Member States responding to the survey, 8% (4/50) have a health-specific AI strategy and 14% (7) are developing one."
Adib K, Letchford N, Dunning HE, Salama N, Tolias Y, De Barros J, et al. (2026). Governance of artificial intelligence for health systems, WHO European Region. Bulletin of the World Health Organization. 10.2471/blt.25.294978 — only the abstract - the full text could not be fetched — the passage: "Less than 10% (4) of Member States have developed liability standards for AI or guidance on the application of existing liability standards."
Adib K, Letchford N, Dunning HE, Salama N, Tolias Y, De Barros J, et al. (2026). Governance of artificial intelligence for health systems, WHO European Region. Bulletin of the World Health Organization. 10.2471/blt.25.294978 — only the abstract - the full text could not be fetched — the passage: "The survey was available to all 53 Member States between June 2024 and March 2025."
Adib K, Letchford N, Dunning HE, Salama N, Tolias Y, De Barros J, et al. (2026). Governance of artificial intelligence for health systems, WHO European Region. Bulletin of the World Health Organization. 10.2471/blt.25.294978 — only the abstract - the full text could not be fetched — the passage: "In the WHO European Region, governance of AI in health care is underdeveloped."
Adib K, Letchford N, Dunning HE, Salama N, Tolias Y, De Barros J, et al. (2026). Governance of artificial intelligence for health systems, WHO European Region. Bulletin of the World Health Organization. 10.2471/blt.25.294978 — only the abstract - the full text could not be fetched — the passage: "Adaptive legal and policy mechanisms are needed to respond effectively to the complex and evolving challenges of AI integration in health systems."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - the article this story is about — the whole article — the passage: "Only 7.4% of respondents reported that their institution had clear policies governing AI use, while 77.0% indicated no institutional guidance existed, and 15.6% were uncertain."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - the article this story is about — the whole article — the passage: "The most frequently reported application was research and literature review (66.1%), followed by educational material generation (59.1%), clinical documentation and summarization (40.9%), diagnostic reasoning support (38.3%), and patient communication or translation (32.2%)."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - the article this story is about — the whole article — the passage: "Latin American emergency clinicians are rapidly adopting generative AI with perceived benefits across multiple domains, but almost entirely without governance frameworks or structured training, highlighting an urgent need for regional policy and education infrastructure."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - the article this story is about — the whole article — the passage: "A staged implementation model could guide this transition: (1) a recognition phase that formally acknowledges existing informal use and defines its scope; (2) a capacity-building phase providing structured education and competency frameworks; (3) a governance phase developing institutional and national policy; (4) a pilot implementation phase with monitored, auditable use cases; and (5) an evaluation and scaling phase with continuous assessment of safety, quality, and equity outcomes."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - the article this story is about — the whole article — the passage: "Regional professional bodies could play a convening role by issuing consensus guidance, minimum standards, and ethical principles tailored to emergency care."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - the article this story is about — the whole article — the passage: "Interest in formal AI training was very high (88.4%), with a strong preference for virtual (69.7%) or hybrid (26.2%) formats."
Baez, A. A., Acra, C., Difo, B. (2026). Adoption, perceived value, and barriers to generative artificial intelligence in emergency medicine: A cross-sectional survey of Latin American clinicians. International Journal of Critical Illness and Injury Science. https://doi.org/10.4103/ijciis.ijciis_43_26 - the article this story is about — the whole article — the passage: "Importantly, 65.6% of respondents supported integration of generative AI into clinical protocols or official guidelines, indicating readiness for structured, regulated adoption once appropriate safeguards are in place."

Do not take this as professional medical advice.