survey · Health affairs scholar · la publicación, 25 sep 2026 · gratis
Una llamada en tigriña, un intérprete que nunca llegó y lo que opinan quienes contestan el 9-1-1
Ochenta y nueve despachadores de tres condados de California dijeron qué debería hacer la inteligencia artificial en una emergencia y qué no. Son opiniones recogidas en una encuesta, no pruebas de ningún sistema.
Versión breve · la versión detallada sigue, unos 8 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
- Operadores del 9-1-1 que atienden sobre todo llamadas médicas
- Cuántos
- 131 invitados; 89 contestaron al menos una pregunta abierta
- Dónde
- California, Estados Unidos (condados de Santa Clara, San Mateo y San Diego)
- Cuándo
- Entre febrero y abril de 2026
- Tipo de estudio
- survey
- Quién lo hizo
- Universidad de Stanford y Universidad de California en Los Ángeles
- El límite que importa
- Son opiniones de tres centros urbanos; no se probó ninguna herramienta ni se preguntó a quienes llaman.
Porcentaje de centros que reportaron cada servicio en un sondeo de la asociación nacional de números de emergencia (NENA); son respuestas de los centros, no una medición de qué tan bien traduce cada servicio.
Llamadas que no son de emergencia, según un informe federal de 2025
La inteligencia artificial redujo las llamadas no de emergencia 36% en Monterey y 40% en Jefferson
La pregunta que queda no es si una máquina contestará su emergencia. Es quién, en la agencia de su localidad, decide ponerla a contestar, y quién responde cuando se equivoca.
Alguien llamó al 9-1-1 y dijo que hablaba algo que sonaba a "segrenia". La petición automática de idioma entendió otra cosa. El intérprete nunca se consiguió. El idioma era tigriña, y de la llamada apenas se obtuvo una dirección. Lo contó una persona que atiende llamadas de emergencia.
Investigadores de Stanford y de la Universidad de California en Los Ángeles preguntaron a quienes atienden llamadas de emergencia médica en tres centros de California: los condados de Santa Clara, San Mateo y San Diego. La encuesta se hizo entre febrero y abril de 2026. De 131 despachadores invitados, 89 contestaron al menos una de las tres preguntas abiertas.
La mayoría vio en la inteligencia artificial una ayuda posible. Podría separar las llamadas que no son emergencia, como los avisos de fuegos artificiales un 4 de julio. Los despachadores ven ocasión de que la inteligencia artificial ayude a ubicar al que llama cuando el sistema actual no da con el lugar exacto, por ejemplo en edificios de varios pisos. Podría reconocer el idioma más rápido y acortar las esperas de intérpretes. Una llamada en criollo tardó al menos 20 minutos en dar con una dirección.
También pusieron condiciones. La herramienta debe apoyar al despachador, no sustituirlo, y la decisión final debe quedar en manos de una persona. Temen que la máquina no transmita empatía, y que se pierda lo que se oye de fondo: un ruido que avisa de peligro. Preguntaron quién responde si la herramienta falla y cómo se revisa el error. Una lo dijo así: "Mi mayor preocupación es que si usamos esta tecnología, quién es responsable si falla".
Son las respuestas de 89 despachadores de tres condados urbanos y con muchos residentes que hablan poco inglés. Entre el 61.8% y el 67.9% contestó cada pregunta, así que hay voces que no aparecen. El estudio no probó ninguna herramienta: nada muestra que un programa traduzca o ubique mejor a quien llama. Tampoco dice nada sobre el centro de su ciudad. Si quiere saber qué usa el suyo, pregúntele directamente.
La pregunta que queda no es si una máquina contestará su emergencia. Es quién, en la agencia de su localidad, decide ponerla a contestar, y quién responde cuando se equivoca.
Qué significa para usted
Lo que dicen estos despachadores vale para tres condados de California y para nadie más. Si en su país se anuncia un programa así, pregunte en la agencia que atiende las emergencias si una persona conserva la última palabra y quién responde cuando la herramienta se equivoca. Nada de esto prueba que traduzca o ubique mejor.
Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262
Quién pagó: El estudio fue financiado por la Fundación CARESTAR, que, según los autores, no tuvo ningún papel en el diseño del estudio, la recolección de datos, el análisis, la interpretación de resultados, la preparación del manuscrito ni la decisión de enviarlo.
No tome esto como consejo médico profesional.
Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1532 palabras · unos 8 minLeerla →Cerrar
La IA en el 9-1-1: quién responde cuando la máquina se equivoca
Invitaron a 131 operadores de emergencias en California; 107 completaron casi toda la encuesta y 89 contestaron al menos una pregunta abierta. En general ven la IA como apoyo, no como reemplazo, y preguntan quién responde si la tecnología falla.
En California, un grupo de investigadores preguntó a los operadores del 9-1-1 qué piensan de que la inteligencia artificial entre a sus llamadas. Participaron tres centros de llamadas de emergencia: los condados de Santa Clara y San Mateo, en el norte, y el de San Diego, en el sur. Invitaron a 131 operadores que atienden sobre todo llamadas médicas; 107 completaron casi toda la encuesta y 89 contestaron al menos una de las tres preguntas abiertas. El estudio aparece en la revista Health Affairs Scholar y lo pagó la fundación CARESTAR, que no intervino en el diseño, la recogida de datos, el análisis, la interpretación de los resultados, la preparación del manuscrito ni la decisión de enviarlo a publicar.
Las respuestas dejaron ocho temas. El primero: los operadores ven la IA como una herramienta que acompaña al humano, no que lo sustituye. Lo dijo uno de ellos con claridad: la IA puede ayudar, pero la decisión final debe quedar en personas capacitadas, que puedan corregirla1. Junto a eso aparecieron la preocupación por no perder el trato humano y la empatía, la duda sobre quién es responsable si el sistema falla, el cuidado de los datos personales, el filtro entre llamadas de emergencia y las que no lo son, la ubicación de quien llama, la detección del idioma y las demoras para conseguir un intérprete.
La parte del idioma pesa. Un operador contó que alguien dijo hablar algo que sonaba a "segrenia"; el sistema automático de idioma entendió otra cosa, nunca llegó un intérprete y solo se pudo obtener una dirección. El idioma era tigriña. Otro relató una llamada en criollo que tardó al menos 20 minutos y varias llamadas a otras personas antes de armar una dirección. Una llamada en vietnamita usó la línea de idiomas y nunca se conectó. El estudio no midió si la IA traduce mejor: recogió lo que los operadores opinan.
Hay que decir con claridad qué es este trabajo y qué no. No es una prueba de ningún sistema de IA. No midió tiempos de despacho, ni resultados de salud, ni si una máquina entiende mejor un acento. Son opiniones de operadores de tres centros en zonas urbanas y con mucha diversidad de idiomas, y los propios autores advierten que eso limita lo que se puede decir de otros centros2. Tampoco se preguntó a quienes llaman, ni a los intérpretes, ni a los paramédicos.
El artículo describe un estudio en el que en Estados Unidos se reciben unos 240 millones de llamadas al 9-1-1 cada año, repartidas en 5,806 centros3. Y el artículo describe un estudio en el que en 2024 había 27.1 millones de personas con dominio limitado del inglés; entre ellas, los idiomas más comunes eran el español, el chino —mandarín y cantonés— y el vietnamita45. El artículo describe un estudio en el que, cuando quien llama y quien contesta no comparten idioma, las llamadas duran más, se retrasa la conexión con un intérprete y se retrasa el envío de la ayuda adecuada6.
Lo curioso es que la tecnología ya está en la puerta. El artículo describe un informe de la agencia federal de telecomunicaciones (NTIA) de 2025 en el que la IA redujo las llamadas no de emergencia en 36% en el condado de Monterey, California, y en 40% en el condado de Jefferson, Colorado7. El artículo describe otro sondeo de la asociación nacional de números de emergencia (NENA) en el que 90% de los centros tenía línea telefónica de idiomas, 24% tenía transcripción de mensajes de texto al 9-1-1, 12% tenía transcripción de voz con IA, 5% tenía transcripción de audio en vivo y 5% no tenía ningún servicio de traducción8. El artículo describe que desde marzo de 2026, la oficina del sheriff del condado de San Diego puso en marcha Hyper, un sistema de voz con IA que contesta y canaliza llamadas que no son de emergencia, aunque no se usa para llamadas médicas de emergencia en ese condado9.
En palabras llanas, la IA podría apoyar varias tareas: clasificar si una llamada es urgencia o no, identificar el idioma de quien habla y ampliar los servicios de traducción e interpretación; la ubicación de quien llama ya se hace hoy con programas de localización, y la IA solo aparece como una oportunidad para mejorarla. Los operadores mencionaron los fuegos artificiales como ejemplo de llamada que no es de emergencia y que la IA podría filtrar, y señalaron que la ubicación de quien llama todavía falla en algunos lugares, como aulas u oficinas en edificios de varios pisos. El operador, mientras tanto, escucha lo que pasa de fondo. Eso también es información.
El ruido es el enemigo. Un operador lo dijo sin rodeos: muchas llamadas tienen audio pobre y mucho ruido de fondo, y eso puede volver difícil que un modelo traduzca con confianza10. Otros pidieron que la IA se adapte a acentos, que entienda palabras mal pronunciadas y hasta que cambie el tono o la velocidad para calmar a alguien alterado. La pregunta queda abierta: ¿puede una máquina hacer eso?
Sobre la responsabilidad, los operadores no se quedaron callados. Uno preguntó quién responde si la tecnología falla11. Otro explicó que con un intérprete humano se puede reclamar al servicio y presentar una queja, pero que con la IA no sabe cuál sería el seguimiento12. Los operadores también preguntaron cómo se van a identificar, atender y dar seguimiento a los errores de la IA.
Los autores proponen pasos concretos. Los operadores deberían conservar la autoridad final sobre el procesamiento de la llamada: la clasificación, las instrucciones previas a la llegada y la asignación de recursos13. La responsabilidad debería quedar definida por escrito entre proveedores, centros de llamadas y operadores en los contratos, antes de que el sistema entre en operación y no después de un daño14. Y el artículo describe un hueco legal: los centros que reciben dinero federal tienen obligaciones hacia las personas con dominio limitado del inglés bajo el Título VI, pero ese título no dice nada específico sobre la interpretación hecha por IA15.
Así lo leemos nosotros. El patrón que se repite en muchas instituciones es poner a una máquina en medio de la decisión: aparece limpia, pareja, sin preferencias, y el error se queda sin dueño visible. No decimos que eso pase ya en estos centros; decimos que es lo que hay que vigilar. Si los contratos y los reglamentos no nombran a una persona responsable por cada falla del sistema automático, el ciudadano afectado no tendrá a quién reclamar. Usted lo puede comprobar: si algún día le contesta o le traduce una máquina, pida el nombre o el número de identificación de la persona responsable de esa llamada y anótelo. Y pregunte a la autoridad de emergencias o de telecomunicaciones de su país quién responde legalmente cuando el sistema automático se equivoca.
Hay una segunda cosa que conviene tener presente. Es probable que mucha gente que no habla bien inglés prefiera la traducción automática a esperar veinte minutos por un intérprete humano, y que ese alivio quite presión para exigir intérpretes de carne y hueso. Lo sabremos si los propios usuarios de idiomas minoritarios, organizados, rechazan la traducción automática y exigen por escrito intérpretes humanos con plazos, aun cuando la máquina sea más rápida. Mientras tanto, cuando use un servicio público en su idioma, fíjese si le ofrecen una máquina en lugar de una persona, pregunte si existe la opción de un intérprete humano y pida que quede registrado por escrito que la atención fue automática.
También hay una tercera, más de fondo. Una conversación real exige que las dos partes puedan turnarse, corregirse y responder a lo que la otra acaba de decir. Si la máquina identifica mal el idioma y usted no tiene manera de corregirla en el momento, la llamada avanza con datos equivocados y nadie del otro lado puede confirmarlos ni desmentirlos. Lo sabremos si los sistemas permiten corregir el idioma detectado en voz alta y un operador humano confirma esa corrección antes de despachar ayuda, y si esos casos quedan registrados. Mientras eso no exista, al llamar a emergencias diga primero el nombre de su idioma, despacio y en voz alta, y repítalo si la respuesta no tiene sentido. Si puede, pida que un familiar o vecino bilingüe esté presente o llame junto a usted.
El estudio deja una tarea pendiente que los autores mismos señalan: escuchar a quienes llaman, con y sin dominio del inglés, a los intérpretes, a los paramédicos y a la comunidad. Hasta que eso ocurra, lo que tenemos es la voz de quienes hoy levantan el teléfono: la IA puede ayudar, dicen, pero la llamada sigue siendo de una persona.
La próxima vez que marque el 9-1-1, o el número de emergencias de su país, fíjese quién contesta y qué idioma le ofrecen. Esa llamada también es suya, y tiene derecho a saber quién responde por ella.
De dónde sale cada dato de contexto, y cuánto leímos de cada documento
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - el artículo del que trata esta nota — el artículo completo — el pasaje: "I think AI can help as a tool, but it should be up to people to make the ultimate decisions or be able to override AI, assuming those people have proper training."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - el artículo del que trata esta nota — el artículo completo — el pasaje: "This study was based on survey responses from 9-1-1 telecommunicators employed in call centers in Santa Clara, San Mateo, and San Diego Counties, which may limit the generalizability of the findings to other other 9-1-1 call centers."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Each year, an estimated 240 million 9-1-1 calls are received across 5806 9-1-1 call centers nationwide."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - el artículo del que trata esta nota — el artículo completo — el pasaje: "As of 2024, 27.1 million people in the United States had LEP."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Among individuals with LEP, the most common languages were Spanish, Chinese, including Mandarin and Cantonese, and Vietnamese."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Language discordance between callers and 9-1-1 telecommunicators is associated with longer 9-1-1 calls, delays in connecting with interpreters, and delays in dispatching the appropriate EMS resources."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - el artículo del que trata esta nota — el artículo completo — el pasaje: "In 2025, the National Telecommunications and Information Administration (NTIA) published a report on the current use of AI in 9-1-1 emergency call centers, noting that AI reduced non-emergency calls by 36% in Monterey County, California, and 40% in Jefferson County, Colorado."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Among respondents, 90% reported access to a telephone language line, 24% had text-to-9-1-1 transcription, 12% had AI-based voice transcription, 5% had AI-based live audio transcription, 5% reported no language translation services, and 2% indicated another language service."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - el artículo del que trata esta nota — el artículo completo — el pasaje: "As of March 2026, for example, San Diego County Sheriff's Office launched Hyper, an AI voice processing system to autonomously answer and route non-emergency calls, although this technology is not currently used for emergency medical calls in this county."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Many 9-1-1 calls have less than ideal audio quality with a lot of background noise…and this artifact could easily make it harder for an AI model to reliably translate."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - el artículo del que trata esta nota — el artículo completo — el pasaje: "My biggest concern is that if we are using this technology who is liable if the technology fails."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - el artículo del que trata esta nota — el artículo completo — el pasaje: "I am not comfortable with AI interpreting when I don't know exactly what is being said…If an interpreter does something wrong, we can follow up with the answering service and file a complaint. The idea that AI could make a mistake makes me nervous and I don't know what the follow up would be."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Telecommunicators should retain final authority over 9-1-1 call processing, including call classification, pre-arrival instructions, and resource assignment and dispatch."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Third, participants asked about clear assignment of liability: responsibility should be predefined among vendors, call centers, and 9-1-1 telecommunicators in pre-deployment contracts and state authorizations rather than after an adverse event."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Furthermore, 9-1-1 call centers receiving federal financial assistance carry Title VI obligations toward persons with LEP, yet Title VI does not comment specifically on AI-mediated interpretation."
Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262
Quién pagó: El estudio fue financiado por la Fundación CARESTAR, que, según los autores, no tuvo ningún papel en el diseño del estudio, la recolección de datos, el análisis, la interpretación de resultados, la preparación del manuscrito ni la decisión de enviarlo.
No tome esto como consejo médico profesional.
Los hallazgos de otros estudios que aquí se mencionan los conocemos por este documento, que fue el que leímos; no abrimos cada uno de esos estudios.
survey · Health affairs scholar · the paper, 25 Sep 2026 · free
A Caller Said “Segrenia.” The Language Was Tigrinya.
In a survey of 89 emergency dispatchers in three California counties, most wanted AI to help with language and location, and a few were strongly against adding it at all.
Short version · the longer version follows, about 7 min
Ask the weeklyAI chat
Ask me about AI in your life: I point you to the stories weeklyAI has published on it, this week and before.
Conversations are saved for as long as weeklyAI exists, to improve the publication. Answers come in the language you write in.
- The study at a glance
- Who
- 9-1-1 telecommunicators who handle emergency medical calls
- How many
- 89 answered at least one open-ended question
- Where
- Santa Clara, San Mateo and San Diego Counties, California, United States
- When
- February to April 2026
- Kind of study
- survey
- Who did it
- Stanford University and UCLA
- The limit that matters
- It records opinions, not tests of any AI system.
These are shares of an industry survey of 9-1-1 telecommunicators, not of the 89 dispatchers in this study, and respondents could report more than one service.
The question to carry is not whether AI will one day help answer your emergency call. It is who at your local 9-1-1 agency decides that, and who answers when it goes wrong.
A 9-1-1 telecommunicator in California took a call from someone who said they spoke what sounded like “segrenia.” The automated language request misheard. No interpreter was ever reached. The language was Tigrinya.
That story comes from one of the dispatchers surveyed by researchers at Stanford and UCLA. The team asked 9-1-1 telecommunicators who handle emergency medical calls at three California call centers, in Santa Clara, San Mateo and San Diego Counties, what they thought about adding AI to their work. The survey ran from February to April 2026. Of 131 invited, 89 answered at least one of the three open-ended questions.
Most saw ways AI could help. It could sort out non-emergency calls, like fireworks reports on the Fourth of July, and free the phones for real emergencies. It could help pin down where a caller is. It could identify a caller’s language faster and cut the waits for interpreters. One dispatcher described a Creole call that took at least 20 minutes and several phone calls just to get an address.
They also drew a line. “I think AI can help as a tool, but it should be up to people to make the ultimate decisions,” one wrote. Others worried about the human side: “When someone has an emergency, speaking to a person can be calming whereas speaking to a robot is not.” Some feared for their jobs.
There were practical doubts too. Background noise can carry signals a dispatcher needs, like whether a scene is safe. Accents, poor audio and panicked callers could trip up a machine. And if it fails, who answers? “My biggest concern is that if we are using this technology who is liable if the technology fails,” one wrote.
Weigh these as opinions, not results. The study tested no AI system, so nothing here shows any tool actually translates or locates callers better. It covers three urban, linguistically diverse counties, and says nothing about what your own local 9-1-1 center does or plans. Each question drew answers from about 62 to 68 percent of the 131 dispatchers invited, so some views went unheard.
The question to carry is not whether AI will one day help answer your emergency call. It is who at your local 9-1-1 agency decides that, and who answers when it goes wrong.
What this means for you
What was decided in court, or by a regulator, about AI in your own country is a separate question from this survey, and this survey answers none of it. The dispatchers here spoke for three California counties, and only about what they think. If your local emergency service announces an AI tool, the thing to look for is who keeps the final say, and who is named when it fails.
Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262
Who paid: The study was funded by the CARESTAR Foundation, which the authors say had no role in study design, data collection, analysis, interpretation, manuscript preparation, or the decision to submit.
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 · 1324 words · about 7 minRead it →Close
What 9-1-1 Dispatchers Say About AI Taking the Call
California telecommunicators welcome help with language and non-emergencies, and insist a person stays in charge.
Dispatchers at three California call centers described waiting on hold for an interpreter while a caller with limited English waited on the line.
That wait is the small, everyday problem behind a new study of 9-1-1 telecommunicators in California, published 25 September 2026 in *Health Affairs Scholar*. A research team of eight, with expertise in health services research, emergency medicine, EMS, community-engaged research and health policy, surveyed dispatchers about artificial intelligence in their work, especially for translation and interpretation. The CARESTAR Foundation paid for the study and had no role in how it was designed, analyzed or written, the authors report.
The study surveyed telecommunicators who handle emergency medical calls at three call centers, in Santa Clara, San Mateo and San Diego Counties, between February and April 2026. Of 131 invited, 107 completed at least 97 percent of the survey, and 89 answered at least one open-ended question. The researchers read the written answers closely and drew out eight themes. The central one: dispatchers generally saw "human-centered AI" as a tool to supplement, not replace, human expertise1. One put it plainly: "I think AI can help as a tool, but it should be up to people to make the ultimate decisions or be able to override AI, assuming those people have proper training"2.
The article describes a study in which an estimated 240 million 9-1-1 calls arrive each year at 5,806 call centers across the country3. The article describes a study in which, as of 2024, 27.1 million people in the United States had limited English proficiency, most of them speakers of Spanish, Chinese or Vietnamese45. The article describes a study in which, when caller and dispatcher do not share a language, calls run longer, interpreters take time to reach, and the right medical resources go out later6.
What the technology actually is, in the field today: most centers still rely on a plain telephone language line. The article describes an industry survey in which 90 percent of responding telecommunicators reported access to one, 24 percent had text-to-9-1-1 transcription, 12 percent had AI-based voice transcription, and 5 percent had AI-based live audio transcription; 5 percent reported no translation service at all7. Some places go further. The article describes a report in which, in 2025, a federal telecommunications agency reported that AI cut non-emergency calls by 36 percent in Monterey County, California, and 40 percent in Jefferson County, Colorado8. The article describes a report in which, in March 2026, the San Diego County Sheriff's Office launched Hyper, an AI voice system that answers and routes non-emergency calls, though it is not used for emergency medical calls there9. Note what those figures measure: calls diverted, not care delivered.
Four of the eight themes were hopeful. Dispatchers saw a use for AI in separating emergency from non-emergency calls — one imagined it absorbing Fourth of July fireworks reports and freeing the phones. They saw a use in confirming where a caller is, especially inside large buildings where current location software falters. They saw a use in identifying a caller's language faster. And they saw a use in cutting the wait for an interpreter, a wait one described as "at least 20 minutes and calls to multiple people" before an address came together.
The other themes were cautions, and they carry the weight. A dispatcher asked: "My biggest concern is that if we are using this technology who is liable if the technology fails"10. Another drew a comparison that stays with you: with a human interpreter who errs, "we can follow up with the answering service and file a complaint"; with AI, "I don't know what the follow up would be"11. Others worried about noise, poor audio, accents and hysterical callers. "Many 9-1-1 calls have less than ideal audio quality with a lot of background noise," one wrote, "and this artifact could easily make it harder for an AI model to reliably translate"12.
Here are the limits, stated plainly. This is a survey of opinions, not a test of any system: nothing here shows that any AI tool translates or locates callers better than a person. The three centers sit in relatively urban, linguistically diverse parts of California, so the findings may not carry to other call centers13. Between about 62 and 68 percent of those invited answered each open question, so some views went unrecorded. Callers, interpreters, paramedics and community members were not asked at all.
The authors' own recommendations follow from the answers. Dispatchers should keep final authority over classifying calls, giving pre-arrival instructions, and sending resources14. Responsibility for AI errors should be settled among vendors, call centers and dispatchers in contracts signed before deployment, not argued over afterward15. Call centers that take federal money have language duties toward people with limited English, but the relevant civil-rights law says nothing specific about AI interpretation16. And the study cites two warnings from elsewhere: a Danish randomized trial in which machine-learning help did not improve dispatchers' recognition of cardiac arrest despite the software catching more cases, and a study finding racial bias in automated speech recognition. We could read only the article's account of those two; both are cautions about implementation, not results of this survey.
Here is how we read it. The pattern in these answers is not fear of machines. It is a request for a chain of responsibility: someone to check the machine, someone to answer when it errs, and a way for the person on the phone to be heard. Expect the harder problem to be the uncommon language. Picture a caller speaking Tigrinya, as one dispatcher described, and a system that guesses wrong with perfect confidence while the dispatcher, unable to check the source, acts on it. You would know we are wrong if call centers published error rates for their language tools, language by language, and those rates held up for the rarest ones. Until then, if you ever call emergency services in a language other than English, ask the dispatcher to read the key details back to you, and ask which language service is being used.
There is a second pattern worth naming. When an automated system misjudges a call — as a non-emergency, as the wrong language, as an uncooperative caller — the people affected often have no obvious route to contest it. A dispatcher can complain about a human interpreter's agency. Nobody in these answers knew whom to complain to about software. If a system mishandles your call, note the time and the number, and ask the call center or your county how to file a complaint. Ask, too, whether the vendor's performance is reported by language and by accent.
One more comparison, from a different field, because it shows the shape of the problem: in a paper on courts in Russia and China (we could read only the summary), the author finds digital tools now woven through nearly every stage of judicial proceedings, including artificial intelligence used to support judges' decisions17. Two countries pursuing the same goal of accessible justice chose different paths — Russia adding individual digital services, China building one unified platform1819. The lesson for a call center is modest: the tool is the easy part; the rules around it are the work.
The phone rings again in that California call center. The caller cannot say, in English, what is wrong. What the dispatchers in this study ask for is not a robot on the line. It is a faster bridge to a person, and a person who stays answerable when the bridge fails.
So the next time you hear a public agency announce an AI helper for its emergency line, the question to put to your city or county is a short one: when the system gets it wrong, who answers for it — and how would I file a complaint?
Where each piece of context comes from, and how much of it we read
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - the article this story is about — the whole article — the passage: "9-1-1 telecommunicators generally viewed human-centered AI as a tool to supplement, rather than replace, human expertise in EMD."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - the article this story is about — the whole article — the passage: "I think AI can help as a tool, but it should be up to people to make the ultimate decisions or be able to override AI, assuming those people have proper training."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - the article this story is about — the whole article — the passage: "Each year, an estimated 240 million 9-1-1 calls are received across 5806 9-1-1 call centers nationwide."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - the article this story is about — the whole article — the passage: "As of 2024, 27.1 million people in the United States had LEP."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - the article this story is about — the whole article — the passage: "Among individuals with LEP, the most common languages were Spanish, Chinese, including Mandarin and Cantonese, and Vietnamese."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - the article this story is about — the whole article — the passage: "Language discordance between callers and 9-1-1 telecommunicators is associated with longer 9-1-1 calls, delays in connecting with interpreters, and delays in dispatching the appropriate EMS resources."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - the article this story is about — the whole article — the passage: "Among respondents, 90% reported access to a telephone language line, 24% had text-to-9-1-1 transcription, 12% had AI-based voice transcription, 5% had AI-based live audio transcription, 5% reported no language translation services, and 2% indicated another language service."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - the article this story is about — the whole article — the passage: "In 2025, the National Telecommunications and Information Administration (NTIA) published a report on the current use of AI in 9-1-1 emergency call centers, noting that AI reduced non-emergency calls by 36% in Monterey County, California, and 40% in Jefferson County, Colorado."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - the article this story is about — the whole article — the passage: "As of March 2026, for example, San Diego County Sheriff's Office launched Hyper, an AI voice processing system to autonomously answer and route non-emergency calls, although this technology is not currently used for emergency medical calls in this county."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - the article this story is about — the whole article — the passage: "My biggest concern is that if we are using this technology who is liable if the technology fails."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - the article this story is about — the whole article — the passage: "I am not comfortable with AI interpreting when I don't know exactly what is being said…If an interpreter does something wrong, we can follow up with the answering service and file a complaint. The idea that AI could make a mistake makes me nervous and I don't know what the follow up would be."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - the article this story is about — the whole article — the passage: "Many 9-1-1 calls have less than ideal audio quality with a lot of background noise…and this artifact could easily make it harder for an AI model to reliably translate."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - the article this story is about — the whole article — the passage: "This study was based on survey responses from 9-1-1 telecommunicators employed in call centers in Santa Clara, San Mateo, and San Diego Counties, which may limit the generalizability of the findings to other other 9-1-1 call centers."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - the article this story is about — the whole article — the passage: "Telecommunicators should retain final authority over 9-1-1 call processing, including call classification, pre-arrival instructions, and resource assignment and dispatch."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - the article this story is about — the whole article — the passage: "Third, participants asked about clear assignment of liability: responsibility should be predefined among vendors, call centers, and 9-1-1 telecommunicators in pre-deployment contracts and state authorizations rather than after an adverse event."
- Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262 - the article this story is about — the whole article — the passage: "Furthermore, 9-1-1 call centers receiving federal financial assistance carry Title VI obligations toward persons with LEP, yet Title VI does not comment specifically on AI-mediated interpretation."
- Rozhkova MA. (2026). Intellectual Property Litigation in Russia and China: Digital Technologies for Specialized Courts. Digital Law Journal. https://doi.org/10.38044/2686-9136-2026-7-3 — only the abstract - the full text could not be fetched — the passage: "The paper demonstrates that digital technologies are now integrated into virtually every stage of judicial proceedings, including electronic document management, the submission and examination of electronic evidence, remote access to case files, automated case allocation, electronic judicial notifications, remote court hearings, and the use of artificial intelligence to support judicial decision-making."
- Rozhkova MA. (2026). Intellectual Property Litigation in Russia and China: Digital Technologies for Specialized Courts. Digital Law Journal. https://doi.org/10.38044/2686-9136-2026-7-3 — only the abstract - the full text could not be fetched — the passage: "A comparison of the Russian and Chinese experiences shows that, although both countries pursue the common objective of improving the accessibility and effectiveness of judicial protection, they have adopted different models for the digitalization of the judiciary."
- Rozhkova MA. (2026). Intellectual Property Litigation in Russia and China: Digital Technologies for Specialized Courts. Digital Law Journal. https://doi.org/10.38044/2686-9136-2026-7-3 — only the abstract - the full text could not be fetched — the passage: "The Russian approach is primarily based on the implementation of individual digital services, whereas the Chinese model emphasizes the gradual development of a unified digital platform integrating different stages of judicial proceedings."
Melgoza, E., Torres, D. X., Hsiang, E. et al. (2026). 9-1-1, what is your emergency? Human-centered artificial intelligence, language access and health policy. Health Affairs Scholar. https://doi.org/10.1093/haschl/qxag262
Who paid: The study was funded by the CARESTAR Foundation, which the authors say had no role in study design, data collection, analysis, interpretation, manuscript preparation, or the decision to submit.
Do not take this as professional medical advice.
The findings of other studies mentioned here are known to us through this document, which is the one we read; we did not open each of those studies.