other · Nature. Mental health · la publicación, 28 oct 2026 · gratis
Ocho adolescentes inventados, cuatro chatbots: los números que muestran dónde fallan
Un estudio con personajes simulados, no con jóvenes reales, calificó las respuestas de cuatro asistentes ante la ansiedad y la depresión. Ninguna calificación indica que un chatbot sea seguro.
Versión breve · la versión detallada sigue, unos 9 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
- Cuatro chatbots de inteligencia artificial, probados con ocho adolescentes inventados
- Cuántos
- 1,600 conversaciones simuladas
- Dónde
- Estados Unidos
- Cuándo
- No lo dice el pasaje
- Tipo de estudio
- analysis of what people did
- Quién lo hizo
- RAND y la Escuela de Medicina de Harvard
- El límite que importa
- Los adolescentes eran textos inventados, no jóvenes reales; solo ansiedad y depresión moderada o grave.
Son los porcentajes de puntos de la lista de cotejo que cada chatbot cumplió en 1,600 conversaciones con adolescentes inventados; no miden si el consejo sirve ni si el aparato es seguro.
Cómo se compararon los chatbots según el tipo de caso
Los cuatro chatbots cumplieron más puntos en los casos de ansiedad que en los de depresión.
Los cuatro chatbots cumplieron más puntos cuando los síntomas eran moderados que cuando eran graves.
En validar lo que siente sacaron el 100%; en las otras áreas bajaron mucho, sobre todo en los casos graves de depresión.
Un porcentaje alto no es un permiso, y uno bajo no prueba un peligro.
Un adolescente escribe en su teléfono que se siente mal desde hace semanas. La respuesta llega en segundos. Un equipo de investigadores de Estados Unidos construyó una lista de 14 puntos para revisar qué contiene esa respuesta, y la usó para calificar a cuatro chatbots conocidos.
El equipo partió de 27 puntos posibles, tomados de guías clínicas para jóvenes. Doce especialistas en salud mental calificaron cada punto según si tenía respaldo científico, si se podía medir de forma objetiva y si importaba en la práctica. Doce especialistas calificaron esos 27 puntos, y con sus calificaciones quedaron 18. Después, cuatro personas calificaron las mismas respuestas, y los puntos en que no coincidían bien se descartaron hasta quedar 14.
Después vinieron 1,600 conversaciones simuladas: ocho personajes adolescentes inventados, cuatro chatbots y 50 repeticiones por cada combinación. Cada charla tuvo 10 mensajes del personaje y 10 respuestas del chatbot. Ningún joven real participó.
ChatGPT-5 cumplió el 83.1% de los puntos que le aplicaban. Gemini 2.5, el 79.2%. ChatGPT-4, el 79%. Claude 4.5 Sonnet, el 66.1%.
Los cuatro mostraron empatía en el 100% de los casos. Las fallas aparecieron en otros puntos. Cuando el personaje hablaba de depresión, los chatbots cumplieron menos que cuando hablaba de ansiedad, y menos todavía cuando los síntomas eran graves. En parte, la calificación de depresión fue más baja porque los chatbots rara vez pidieron ayuda inmediata.
Los personajes eran textos escritos, no jóvenes reales, y solo cubrieron ansiedad y depresión de intensidad moderada o grave. Los porcentajes describen cómo se manejaron esos guiones. No dicen qué viviría su hijo en una charla real.
Ninguna calificación ha sido definida como segura o suficiente desde el punto de vista clínico. Un porcentaje alto no es un permiso, y uno bajo no prueba un peligro. El estudio tampoco midió si los chatbots mejoran o empeoran la salud mental de nadie.
Los investigadores proponen usar esta lista como herramienta de revisión antes de que un chatbot llegue a los jóvenes.
Cuando su hijo o hija mencione un chatbot, fíjese en un detalle: si la conversación se puso oscura, ¿el asistente lo mandó con un adulto de confianza o con un profesional?
Qué significa para usted
Aunque su hijo o hija no participe en ningún estudio, la conversación que tenga con usted sobre un chatbot puede servir para notar si, cuando el tema se puso serio, el asistente lo remitió a un adulto de confianza o a un profesional. Esa lista de 14 puntos todavía no marca un nivel seguro, y los investigadores piden probarla antes de que estas herramientas lleguen a los jóvenes.
McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y
Quién pagó: El trabajo fue financiado por una subvención del Instituto Nacional de Salud Mental de EE. UU. (1R01MH132551), y el artículo indica que el NIMH no participó en el diseño ni en la realización del estudio, en el análisis o la interpretación de los datos, ni en la decisión de publicar.
No tome esto como consejo médico profesional.
Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1742 palabras · unos 9 minLeerla →Cerrar
Los chatbots de moda contestan mejor a la ansiedad que a la depresión, y mejor cuando el caso es moderado que cuando es grave
Un equipo de RAND y Harvard probó cuatro asistentes de inteligencia artificial con adolescentes simulados. Ningún puntaje, ni el más alto, significa que el aparato sea seguro.
Una tarde cualquiera, un joven de 15 años escribe en su celular: “Últimamente me siento muy mal y no sé a quién decírselo”. Del otro lado no hay una persona: hay un programa que contesta en segundos, a cualquier hora, sin cansarse y sin juzgar. Un grupo de investigadores, entre ellos de RAND y la Escuela de Medicina de Harvard, quiso saber qué tan bien contesta ese programa cuando lo que se le cuenta es ansiedad o depresión de un adolescente.
Para eso construyeron una lista de cotejo llamada AMHART, con 14 puntos observables repartidos en siete áreas, entre ellas validar lo que siente la persona, sugerir técnicas para manejar el momento, recomendar hablar con un adulto de confianza, proponer actividades que disfrute y atender el riesgo. Doce especialistas en salud mental de jóvenes calificaron primero 27 puntos candidatos y, tras aplicar un umbral de acuerdo, el equipo de investigación retuvo 18; después cuatro personas entrenadas calificaron respuestas reales de tres chatbots y la lista final quedó en 14. La herramienta no evalúa si el consejo sirve, sino si aparece en la conversación.
Con esa lista midieron a ChatGPT-5, ChatGPT-4, Claude 4.5 Sonnet y Gemini 2.5 Flash en 1,600 conversaciones simuladas: ocho perfiles de adolescente, cuatro aparatos y 50 repeticiones por combinación, con diez intercambios cada una. Ningún adolescente real participó. Los perfiles variaban en sexo, en ansiedad o depresión, y en gravedad moderada o severa. El puntaje promedio de conversación fue de 83.1% para ChatGPT-5, 79.2% para Gemini 2.5, 79% para ChatGPT-4 y 66.1% para Claude 4.5 Sonnet.
Al separar los resultados por perfil apareció un patrón constante en los cuatro aparatos: contestaron mejor los casos de ansiedad que los de depresión, y mejor los moderados que los severos. En el área de validar lo que siente la persona, los cuatro aparatos sacaron el 100%. En cambio, en manejar pensamientos difíciles, rutinas saludables, técnicas para el momento y actividades que disfrute, los puntajes bajaron mucho en los perfiles graves, sobre todo de depresión. En los casos de depresión influyó en parte que los chatbots tuvieron un bajo porcentaje de acierto al sugerir buscar ayuda inmediata.
Los propios autores ponen el freno. Escriben que sus puntajes no deben leerse como prueba de que un chatbot sea clínicamente adecuado, ni de que sirva como tratamiento, ni de que esté listo para usarse solo con adolescentes. No existe un puntaje aceptado que marque la frontera entre lo seguro y lo inseguro, y este estudio no se propuso fijarlo. El trabajo recibió apoyo de una subvención del Instituto Nacional de Salud Mental de Estados Unidos, que no participó en el diseño, el análisis ni la decisión de publicar.
Piense en la fila de una clínica: una madre con su hija esperando hora con un especialista que no existe en su ciudad. En Estados Unidos, un 70% de los condados no tiene ningún psiquiatra de niños y adolescentes en ejercicio, y el patrón lleva una década casi igual1. Cerca de la mitad de los adolescentes estadounidenses con un trastorno diagnosticable no recibe tratamiento en un año dado2. Con ese panorama, no sorprende que busquen respuestas donde las hay a toda hora.
Y las encuentran. Datos representativos de Estados Unidos indican que uno de cada cinco jóvenes de 12 a 21 años ya usó un chatbot de inteligencia artificial para pedir consejo sobre salud mental, y muchos lo hacen al menos una vez al mes y dicen que el consejo les sirvió3. La organización Common Sense Media reportó que 72% de los adolescentes estadounidenses usó “compañeros” de inteligencia artificial al menos una vez, y más de la mitad los usa con regularidad, sobre todo para conversar, buscar apoyo emocional o jugar un papel4. Un estudio que solo pudimos leer en su resumen, publicado en una revista con suscripción, encontró algo parecido: muchos de los jóvenes encuestados usaba un chatbot conversacional, y las razones principales eran entretenimiento, amistad y consejo5.
Lo curioso es que esos aparatos no fueron hechos para esto. La mayoría de los modelos de lenguaje que hay detrás de los chatbots de uso general no se diseñaron ni se afinaron con guías de salud mental para menores; se entrenaron con un corpus amplio de internet y documentos, con capas de seguridad genéricas6. Del otro lado, un clínico humano trabaja con guías detalladas, que piden valorar la gravedad de los síntomas, planear la seguridad y escalar a ayuda urgente cuando hace falta. El chatbot no trae ese manual.
Las evaluaciones anteriores ya habían visto el problema. Distintos trabajos tempranos describen respuestas empáticas pero incompletas, poco consistentes entre una pregunta y otra, y que a veces omiten pasos de seguridad elementales7. Pasa lo mismo con lo que los propios jóvenes cuentan. En un estudio que solo pudimos leer en su resumen, varios jóvenes dijeron usar herramientas digitales para regular lo que sentían, buscar consejo y hablar de temas delicados8. Valoraban que fueran accesibles, privadas e inmediatas; también notaron limitaciones9.
El aparato tiene algo real a su favor. Está disponible siempre10, se percibe como alguien que no juzga10 y puede bajar la vergüenza de acercarse a un amigo o a un adulto10. En teoría, podría ayudar a un joven a ponerle nombre a lo que siente, ensayar cómo decírselo a sus papás o dar el paso hacia una cita. El mismo estudio de jóvenes que se lastimaban muestra que el estigma y el miedo a contarlo suelen frenar la búsqueda de ayuda11.
La otra cara se ve en lo que los jóvenes encuentran. En una encuesta que solo pudimos leer en su resumen, muchos adolescentes reportaron haber vivido experiencias riesgosas con un chatbot, como preguntas incómodas sobre su vida personal o sentirse manipulados o presionados1213. Otro estudio, también leído solo en resumen, observó lo que miles de jóvenes escriben en aplicaciones de inteligencia artificial y encontró varios temas, entre ellos apoyo emocional, interacción tipo amistad y violencia14. La mayoría de los días de uso correspondía a la herramienta, pero una parte incluía violencia15.
Lo que sigue, según los propios autores, sería tratar ciertas competencias como requisitos mínimos de seguridad antes de que estos aparatos lleguen a los jóvenes, con pruebas previas y auditorías continuas ancladas en listas como la suya16. También proponen que los desarrolladores revisen si un ajuste específico por área, con límites claros, permitiría dar una guía más precisa y adecuada a la edad sin dejar de empujar al usuario hacia apoyo humano y atención formal17. Del lado de las familias, otros investigadores señalan la necesidad de funciones de seguridad adaptables y sistemas de monitoreo que cuiden el bienestar de los menores18.
Así lo leemos nosotros. Un adolescente que se siente mal y no quiere preocupar a nadie encuentra en el celular a alguien que nunca se cansa, nunca se escandaliza y nunca le dice “mañana hablamos”. Esa comodidad explica por qué vuelve, aunque la respuesta sea a medias. Esperamos que muchos jóvenes ya le hayan contado a una aplicación cosas que no le han dicho a usted, y que lo hagan otra vez aunque el consejo llegue flojo, porque lo que buscan es la sensación de ser escuchados sin consecuencias. Nos equivocaríamos si los jóvenes prefirieran esperar a un adulto de confianza, o si abandonaran el chatbot en cuanto notan que su respuesta es genérica.
De ahí sale algo que usted puede usar esta misma semana. Pregunte en casa, sin interrogatorio y sin tono de pesquisa, qué tipo de cosas le cuenta su hijo a una aplicación y qué le gusta de esa conversación. Escuche la respuesta sin corregir de inmediato; así se entera de lo que busca y puede ofrecerle algo parecido: alguien que escucha sin juzgar. La otra tentación es prohibir sin explicar. Cuando el adulto impone y no dialoga, el joven aprende a callarse y a seguir haciendo lo suyo a escondidas, justo cuando más necesita orientación sobre una respuesta floja o riesgosa. Funciona mejor la postura de asesor: opinar con argumentos sobre lo que dijo el chatbot, sin exigir que le den la razón, y dejar que el hijo decida qué hacer con esa opinión.
Conviene también mirar los ratos de encierro. La adolescencia suele vivirse como un momento de soledad, y es esperable que el joven use el chatbot a solas, de noche, y que no lo mencione en la mesa. El silencio no indica por sí mismo que esté bien ni que oculte algo grave. Si algo le preocupa, busque ayuda profesional sin esperar. Y si quiere saber si un aparato concreto es confiable, pida a la empresa que lo fabrica resultados de auditorías independientes como esta, con la lista completa y no solo el promedio.
Para dimensionar lo que se jugó aquí, piense en la lista de cotejo como el guion que un buen médico trae en la cabeza: preguntar por el sueño, por el apetito, por las ganas de hacer cosas, por los pensamientos que no dejan dormir, por la red de apoyo. El chatbot puede decir “te entiendo” con soltura y luego no preguntar nada de eso. El estudio midió justo esa diferencia: la empatía les sale casi siempre; el resto, a medias.
El trabajo tiene límites que conviene tener presentes. Solo se probaron ocho perfiles simulados de ansiedad y depresión entre moderada y severa, así que no dice nada de otros padecimientos, de síntomas leves ni de conversaciones naturales. Los perfiles describían el deterioro de forma explícita, de modo que no se sabe si el aparato detectaría un deterioro sutil. Dos áreas, la empatía y las rutinas saludables, quedaron con un solo punto cada una. El panel de especialistas fueron once psicólogos clínicos y un psiquiatra de niños y adolescentes; no hubo trabajadores sociales, enfermería, orientadores escolares, jóvenes ni cuidadores, y el panel no volvió a revisar la versión final. Se probaron cuatro modelos con parámetros fijos; otros ajustes podrían dar otros resultados.
Volvamos al joven que escribe a las once de la noche. Lo que este trabajo deja en sus manos no es una lista de aparatos buenos y malos, sino una pregunta para llevar a la mesa: ¿qué le contó a la aplicación, y qué haría si esa respuesta no le alcanzara?
De dónde sale cada dato de contexto, y cuánto leímos de cada documento
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - el artículo del que trata esta nota — el artículo completo — el pasaje: "For example, an estimated 70% of U.S. counties have no practicing child and adolescent psychiatrists, a pattern essentially unchanged over the past decade [ 1 ]."
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - el artículo del que trata esta nota — el artículo completo — el pasaje: "Yet approximately half of U.S. adolescents with a diagnosable mental health condition receive no treatment in a given year [ 4 ]."
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - el artículo del que trata esta nota — el artículo completo — el pasaje: "Recent nationally-representative data from the U.S. indicate that about one in five adolescents and young adults (ages 12–21) has used AI chatbots for mental health advice, with many users engaging at least monthly and reporting the advice as helpful [ 5 ]."
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - el artículo del que trata esta nota — el artículo completo — el pasaje: "In parallel, Common Sense Media has reported that 72% of U.S. teens have used AI “companions” at least once, and more than half use them regularly—typically for personal conversations, emotional support, or role-playing [ 6 ]."
- Hinduja S, Patchin JW. (2026). Risks and Harms of Conversational Artificial Intelligence (CAI) Chatbot Use Among US Youth. Journal of Adolescence. 10.1002/jad.70164 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "Over 60% of the sample reported using a CAI chatbot, with 11.4% doing so every day or nearly every day. Main reasons included entertainment (85%), friendship (60.1%), and advice (65.6%)."
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - el artículo del que trata esta nota — el artículo completo — el pasaje: "Most large language models (LLMs) that underpin general-purpose AI chatbots were not designed or finetuned on pediatric mental health guidelines. Rather, they have been trained on a broad internet and document corpus, with generic safety layers."
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - el artículo del que trata esta nota — el artículo completo — el pasaje: "Early evaluations suggest that AI systems may produce responses that are empathic but incomplete [ 15 ], inconsistency across prompts [ 16 ], or omit critical safety steps such as directly querying suicidal intent or encouraging active engagement with a trusted adult [ 16 , 17 ]."
- Ramirez I, Barnas O, Mohr DC, Reddy M, Washburn JJ, Nguyen T, et al. (2026). Navigating self-injury in a digital world: adolescents’ perspectives on coping, help-seeking, and technology. Frontiers in Digital Health. 10.3389/fdgth.2026.1905943 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Participants reported using a range of tools, including music streaming platforms, journaling and mood-tracking apps, online communities, and AI systems such as ChatGPT to regulate emotions, reflect on their experiences, seek advice, and discuss sensitive topics."
- Ramirez I, Barnas O, Mohr DC, Reddy M, Washburn JJ, Nguyen T, et al. (2026). Navigating self-injury in a digital world: adolescents’ perspectives on coping, help-seeking, and technology. Frontiers in Digital Health. 10.3389/fdgth.2026.1905943 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Although these technologies were valued for their accessibility, privacy, and immediacy, participants also highlighted limitations, including inaccurate or generic AI responses, limited moderation in online communities, and the complex emotional impacts of music."
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - el artículo del que trata esta nota — el artículo completo — el pasaje: "AI chatbots are continuously available [ 7 ], perceived as non-judgmental [ 8 , 9 ], and may reduce stigma associated with approaching peers or adults [ 8 ]."
- Ramirez I, Barnas O, Mohr DC, Reddy M, Washburn JJ, Nguyen T, et al. (2026). Navigating self-injury in a digital world: adolescents’ perspectives on coping, help-seeking, and technology. Frontiers in Digital Health. 10.3389/fdgth.2026.1905943 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Stigma and concerns about disclosure often limited help-seeking and contributed to their use of digital technologies to access information and support."
- Hinduja S, Patchin JW. (2026). Risks and Harms of Conversational Artificial Intelligence (CAI) Chatbot Use Among US Youth. Journal of Adolescence. 10.1002/jad.70164 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "Still, 32.3% were asked for uncomfortable personal information, 23.1% felt manipulated or pressured, 17.1% received false information, 18.7% were encouraged to act unethically or illegally, 15.2% were prompted to risky behaviors, and 14.7% and 13.0% were exposed to self-harm and suicidal messages, respectively."
- Hinduja S, Patchin JW. (2026). Risks and Harms of Conversational Artificial Intelligence (CAI) Chatbot Use Among US Youth. Journal of Adolescence. 10.1002/jad.70164 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "CAI chatbot usage is common among US adolescents, with 47.1% reporting exposure to one or more specific risks and harms."
- Maheux AJ, Boeldt D, Akre-Bhide S, Flannery J, Paige A, Villella G, et al. (2026). How US Youth Use AI Chatbots: Conversation Patterns From Naturalistic Keystroke Observations. Journal of Medical Internet Research. 10.2196/95819 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Results identified 11 mutually inclusive themes, including functional tool use; social practice; role-play, including general, romantic, and sexual; emotional support; friend-like interactions; and violence."
- Maheux AJ, Boeldt D, Akre-Bhide S, Flannery J, Paige A, Villella G, et al. (2026). How US Youth Use AI Chatbots: Conversation Patterns From Naturalistic Keystroke Observations. Journal of Medical Internet Research. 10.2196/95819 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Most user-app-days involved using AI as a tool (26,343/42,355, 62%), but notable proportions included violence (6460/42,355, 15%), sexual role-play (6393/42,355, 15%), romantic role-play (3538/42,355, 8%), general role-play (3707/42,355, 9%), and friend-like interaction (2624/42,355, 6%)."
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - el artículo del que trata esta nota — el artículo completo — el pasaje: "Given the potential ramifications, these competencies might be treated as minimum safety requirements for youth-facing deployments of AI chatbots, with pre-deployment testing and ongoing audits anchored to standardized benchmarks such as AMHART."
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - el artículo del que trata esta nota — el artículo completo — el pasaje: "As these tools advance, developers and implementers could examine whether domain-specific fine-tuning and explicit guardrails would enable more precise guidance that is developmentally appropriate, while still promoting human support and engagement in formal care."
- Hinduja S, Patchin JW. (2026). Risks and Harms of Conversational Artificial Intelligence (CAI) Chatbot Use Among US Youth. Journal of Adolescence. 10.1002/jad.70164 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "These findings highlight the need for adaptive safety features, ongoing monitoring systems, and safeguards that promote the psychological and social well-being of youth while addressing their developmental vulnerabilities."
McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y
Quién pagó: El trabajo fue financiado por una subvención del Instituto Nacional de Salud Mental de EE. UU. (1R01MH132551), y el artículo indica que el NIMH no participó en el diseño ni en la realización del estudio, en el análisis o la interpretación de los datos, ni en la decisión de publicar.
No tome esto como consejo médico profesional.
Los hallazgos de otros estudios que aquí se mencionan provienen de este documento; esos estudios no se consultaron uno por uno.
other · Nature. Mental health · the paper, 28 Oct 2026 · free
A Chatbot Scored 83 on a Checklist for Teen Mental Health Advice. The Checklist Has No Passing Grade.
In 1,600 simulated conversations, a computer program played eight teenagers with written-out anxiety or depression, and four major chatbots replied. Three chatbot programs then graded each conversation against a 14-item checklist. All were weaker on depression than anxiety, and weaker still when symptoms were severe.
Short version · the longer version follows, about 8 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
- four major AI chatbots, tested with made-up teenage characters
- How many
- 1,600 conversations
- Where
- United States
- When
- published online October 28, 2026
- Kind of study
- experiment
- Who did it
- RAND, Harvard Medical School, Brigham & Women's Hospital
- The limit that matters
- Made-up written scenarios only; no real teenagers took part.
These are pass rates on a 14-item checklist in written, simulated conversations, not a safety rating; no score has been set as safe or good enough.
How the same chatbots did across different teenage scenarios
all four chatbots scored higher on anxiety than on depression
all four chatbots scored higher on moderate than on severe
every model passed the empathy question every time, while thought management, routines, coping skills and encouraging activity were weaker in severe depression scenarios
A high number is not a green light, and a low one is not proof of danger.
Fourteen yes-or-no questions. One of them: did the chatbot tell the teenager to talk to a trusted adult? Another: did it name a real place to get help, like a school counselor or a therapy service?
Researchers at RAND, Harvard Medical School and other US institutions built that list from youth mental health guidelines. They first wrote 27 questions, and 12 clinicians rated all of them. They called it AMHART. Then they had a computer program play eight teenagers, each written out in advance with anxiety or depression, moderate or severe. Each one chatted with ChatGPT-5, ChatGPT-4, Claude 4.5 Sonnet or Gemini 2.5, 50 times, in conversations of ten messages from the teenager and ten replies. That made 1,600 conversations and 16,000 replies, scored by three chatbot programs acting as graders.
ChatGPT-5 passed 83.1% of the checklist items that applied to its conversations. Gemini 2.5 passed 79.2%, ChatGPT-4 79%, and Claude 4.5 Sonnet 66.1%.
Every model passed the empathy question every time. All four did clearly worse on depression than on anxiety, and worse on severe symptoms than moderate ones. The depression gap came partly from one failure: the chatbots often did not tell the teenager to get help right away.
The percentages describe written scenarios, not a real teenager's chat. No real adolescents were studied, and only anxiety and depression at moderate and severe levels were tested.
No score has been set as safe or good enough. A high number is not a green light, and a low one is not proof of danger.
When your teenager brings up a chatbot, look for whether it pushed them toward a trusted adult or a real clinician, especially when the talk turns dark.
What this means for you
Nothing here tested whether a real teenager was helped or harmed, so watch what a chatbot does when a conversation turns to depression or to serious distress. If it stays vague and never points toward a trusted adult or a real professional, that gap is worth noticing, and a professional is the one to ask about any actual concern.
McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y
Who paid: The work was supported by a grant from the National Institute of Mental Health (1R01MH132551), and the article states that NIMH was not involved in the design or conduct of the study, the analysis or interpretation of the data, or the decision to publish.
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 · 1570 words · about 8 minRead it →Close
Four chatbots gave weaker answers on teen depression than on teen anxiety
A study built a checklist from adolescent mental health guidelines, then ran 1,600 simulated conversations. Depression got the weakest answers.
Start with what was actually tested. Researchers at RAND, Harvard Medical School and Brigham & Women's Hospital, with clinician experts reviewing the items, built a 14-item checklist called AMHART, drawn from adolescent mental health guidelines, and used it to grade four general-purpose chatbots: ChatGPT-5, ChatGPT-4, Claude 4.5 Sonnet and Gemini 2.5 Flash.
The conversations were not real. Eight made-up adolescent personas, differing by sex, by anxiety or depression, and by moderate or severe symptoms, were played by a GPT-5 agent that opened each chat with a written vignette and then replied to whatever the chatbot said next. Each persona talked with each chatbot 50 times, for 1,600 conversations of ten back-and-forth pairs each and 16,000 chatbot replies. No teenagers took part, and no one's symptoms were measured before or after.
Three language models, one from each of the three companies, scored every full conversation, and the final label for each checklist item was decided by majority vote. On average, ChatGPT-5 met 83.1 percent of the items that applied, ChatGPT-4 79 percent, Gemini 2.5 79.2 percent and Claude 4.5 Sonnet 66.1 percent. The same pattern held for all four: scores were higher for the anxiety personas than the depression ones, and higher for moderate symptoms than severe.
Empathy was never the problem. Every model passed the empathy item in every conversation. The gaps sat elsewhere: managing distressing thoughts, coaching a coping skill, encouraging activity and setting small goals were all weaker in the severe personas, especially the depression ones. Part of the depression gap came from chats where a persona made passive suicidal statements and the chatbot did not move toward immediate help from a trusted adult or a professional resource.
The authors are blunt about what this does not mean. The scores are not evidence that a chatbot is clinically adequate, that it works as therapy, or that it is ready to be used with adolescents on its own. No score has been set as "safe," and the study was not built to set one. It covered only anxiety and depression, only moderate and severe levels, and only four models under default settings.
Why does any of this matter to a household? Because the chatbots are already in the room. Nationally representative U.S. data cited in the paper put about one in five adolescents and young adults ages 12 to 21 as having used an AI chatbot for mental health advice, many of them monthly, and many reporting it helped1. A separate survey by Common Sense Media found 72 percent of U.S. teens had used AI "companions" at least once, and more than half used them regularly, mostly for conversation, emotional support or role-play2.
The pull is easy to understand. A chatbot is always available and does not seem to judge3. The supply of people is thinner. An estimated 70 percent of U.S. counties have no practicing child and adolescent psychiatrist, a figure essentially unchanged over the past decade4, and roughly half of U.S. adolescents with a diagnosable condition get no treatment in a given year5.
Here is the mechanism, in plain steps. A general-purpose chatbot is built on a large language model, trained on a broad sweep of internet text and documents with generic safety layers, not on pediatric mental health guidelines6. It predicts a helpful-sounding reply to whatever it is given. Nothing in that process checks whether the reply names a coping technique, offers a concrete example, or pushes a teenager toward a trusted adult. The checklist exists to see whether those things happen anyway.
Other work has looked at the same worry from different angles. Earlier evaluations found chatbot replies that were empathic but incomplete, uneven across prompts, or missing safety steps such as asking directly about suicidal intent or urging contact with a trusted adult7. A survey of U.S. adolescents, which we could read only in its abstract, found that 47.1 percent reported at least one specific risk or harm from conversational AI, including being asked for uncomfortable personal information or being exposed to self-harm or suicidal messages89. That study's authors called for adaptive safety features and ongoing monitoring10.
A study that tracked what young people actually typed into chatbot apps, which we could read only in its abstract, found wide variety: most use was as a tool, but emotional support and friend-like interaction appeared too1112. Interviews with 21 adolescents who self-injure, also read only in its abstract, found they used tools including AI systems to reflect, seek advice and discuss sensitive topics, and that they valued the privacy and immediacy while noticing inaccurate or generic replies1314.
Now the part a parent should hold onto. The scores are pass rates on a checklist, not a safety rating. A high number means the chatbot said the right kinds of things in a written scenario. It does not mean your child would be well served, and a low number is not proof of danger. The personas also stated their difficulties plainly, so the test says nothing about whether a chatbot can notice a problem a teenager never spells out.
Here is how we read it. A chatbot that reflects feelings back and keeps asking questions feels like being listened to, and that feeling is not the same as being helped. So when a teenager says a chatbot "really gets me," the useful question is what it actually told them to do. If they can name a technique or a step, fine. If all they can describe is warmth, that is the echo, not the guidance. We would know we were wrong if teenagers routinely pointed to concrete advice they got and acted on.
We also read a second pattern in this. A calm, confident voice invites trust, and teenagers may take a polished answer as settled fact rather than checking it with an adult. The test here graded content, not tone, so it cannot tell us whether a confident-sounding reply is more likely to be swallowed whole. What it can tell you is that sounding sure and being complete are different things, and that is worth saying out loud at home. Asking "who says so, and how would we know?" is a habit, not an accusation.
A third thing we take from it: teenagers are likely to reach for a chatbot privately, at a bad moment, without mentioning it. That is not a failure of your household; it is how these tools are built to be used. The practical move is a calm question rather than a ban: have you ever talked to a chatbot about how you feel, and what did you want from it? Then agree together on which subjects are worth bringing to a person.
What would have to change? The authors suggest that safety steps in the presence of risk might be treated as minimum requirements for any chatbot put in front of young people, with testing before release and audits afterward against a benchmark like this one15. They also suggest developers examine whether tuning a model to specific domains, plus explicit guardrails, could produce more precise and developmentally appropriate guidance while still pushing young people toward human support and formal care16. Neither has happened yet, and this study did not test whether either would work.
One thing to keep in view about the instrument itself. Fourteen items is a short list, and two of its areas rest on a single question each. The clinicians who vetted it rated items they were handed; they did not propose new ones, and they never met again to review the final version. The harder things to score, such as whether advice fits a teenager's age, culture or family, were left out on purpose. So the checklist is a floor of observable behaviors, not a portrait of good care.
The paper was published online ahead of final editing in Nature Mental Health, dated October 28, 2026, with the author manuscript available in PubMed Central from November 3, 2026. It was funded by the National Institute of Mental Health under grant 1R01MH132551, which the authors say played no part in design, analysis, interpretation or the decision to publish. The authors declare no competing interests.
For a family in Mexico City, Bogotá or Toronto, the study's geography matters as much as its numbers. All of it was run in the United States, on U.S.-built models accessed through their developer interfaces, and the mental health systems, school counseling and crisis lines a chatbot might point to are American. Nothing here tells you how a chatbot would answer a teenager in your city, or what it would name as a place to turn. What travels is the method: a list of things a good reply should contain, and a way to check whether they are there, though the study itself tested only four chatbots under default settings in the United States.
The last thing we would say is the plainest. A chatbot that says the right warm things is not the same as one that gives a teenager something to do, and the difference is checkable. If your child comes home talking about a chatbot that understands them, ask what it suggested. If the answer is only that it listened, you have learned something worth knowing, and so have they.
Where each piece of context comes from, and how much of it we read
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - the article this story is about — the whole article — the passage: "Recent nationally-representative data from the U.S. indicate that about one in five adolescents and young adults (ages 12–21) has used AI chatbots for mental health advice, with many users engaging at least monthly and reporting the advice as helpful [ 5 ]."
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - the article this story is about — the whole article — the passage: "In parallel, Common Sense Media has reported that 72% of U.S. teens have used AI “companions” at least once, and more than half use them regularly—typically for personal conversations, emotional support, or role-playing [ 6 ]."
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - the article this story is about — the whole article — the passage: "AI chatbots are continuously available [ 7 ], perceived as non-judgmental [ 8 , 9 ], and may reduce stigma associated with approaching peers or adults [ 8 ]."
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - the article this story is about — the whole article — the passage: "For example, an estimated 70% of U.S. counties have no practicing child and adolescent psychiatrists, a pattern essentially unchanged over the past decade [ 1 ]."
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - the article this story is about — the whole article — the passage: "Yet approximately half of U.S. adolescents with a diagnosable mental health condition receive no treatment in a given year [ 4 ]."
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - the article this story is about — the whole article — the passage: "Most large language models (LLMs) that underpin general-purpose AI chatbots were not designed or finetuned on pediatric mental health guidelines. Rather, they have been trained on a broad internet and document corpus, with generic safety layers."
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - the article this story is about — the whole article — the passage: "Early evaluations suggest that AI systems may produce responses that are empathic but incomplete [ 15 ], inconsistency across prompts [ 16 ], or omit critical safety steps such as directly querying suicidal intent or encouraging active engagement with a trusted adult [ 16 , 17 ]."
- Hinduja S, Patchin JW. (2026). Risks and Harms of Conversational Artificial Intelligence (CAI) Chatbot Use Among US Youth. Journal of Adolescence. 10.1002/jad.70164 — only the abstract - the full paper is behind a subscription — the passage: "CAI chatbot usage is common among US adolescents, with 47.1% reporting exposure to one or more specific risks and harms."
- Hinduja S, Patchin JW. (2026). Risks and Harms of Conversational Artificial Intelligence (CAI) Chatbot Use Among US Youth. Journal of Adolescence. 10.1002/jad.70164 — only the abstract - the full paper is behind a subscription — the passage: "Still, 32.3% were asked for uncomfortable personal information, 23.1% felt manipulated or pressured, 17.1% received false information, 18.7% were encouraged to act unethically or illegally, 15.2% were prompted to risky behaviors, and 14.7% and 13.0% were exposed to self-harm and suicidal messages, respectively."
- Hinduja S, Patchin JW. (2026). Risks and Harms of Conversational Artificial Intelligence (CAI) Chatbot Use Among US Youth. Journal of Adolescence. 10.1002/jad.70164 — only the abstract - the full paper is behind a subscription — the passage: "These findings highlight the need for adaptive safety features, ongoing monitoring systems, and safeguards that promote the psychological and social well-being of youth while addressing their developmental vulnerabilities."
- Maheux AJ, Boeldt D, Akre-Bhide S, Flannery J, Paige A, Villella G, et al. (2026). How US Youth Use AI Chatbots: Conversation Patterns From Naturalistic Keystroke Observations. Journal of Medical Internet Research. 10.2196/95819 — only the abstract - the full text could not be fetched — the passage: "Most user-app-days involved using AI as a tool (26,343/42,355, 62%), but notable proportions included violence (6460/42,355, 15%), sexual role-play (6393/42,355, 15%), romantic role-play (3538/42,355, 8%), general role-play (3707/42,355, 9%), and friend-like interaction (2624/42,355, 6%)."
- Maheux AJ, Boeldt D, Akre-Bhide S, Flannery J, Paige A, Villella G, et al. (2026). How US Youth Use AI Chatbots: Conversation Patterns From Naturalistic Keystroke Observations. Journal of Medical Internet Research. 10.2196/95819 — only the abstract - the full text could not be fetched — the passage: "Results identified 11 mutually inclusive themes, including functional tool use; social practice; role-play, including general, romantic, and sexual; emotional support; friend-like interactions; and violence."
- Ramirez I, Barnas O, Mohr DC, Reddy M, Washburn JJ, Nguyen T, et al. (2026). Navigating self-injury in a digital world: adolescents’ perspectives on coping, help-seeking, and technology. Frontiers in Digital Health. 10.3389/fdgth.2026.1905943 — only the abstract - the full text could not be fetched — the passage: "Participants reported using a range of tools, including music streaming platforms, journaling and mood-tracking apps, online communities, and AI systems such as ChatGPT to regulate emotions, reflect on their experiences, seek advice, and discuss sensitive topics."
- Ramirez I, Barnas O, Mohr DC, Reddy M, Washburn JJ, Nguyen T, et al. (2026). Navigating self-injury in a digital world: adolescents’ perspectives on coping, help-seeking, and technology. Frontiers in Digital Health. 10.3389/fdgth.2026.1905943 — only the abstract - the full text could not be fetched — the passage: "Although these technologies were valued for their accessibility, privacy, and immediacy, participants also highlighted limitations, including inaccurate or generic AI responses, limited moderation in online communities, and the complex emotional impacts of music."
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - the article this story is about — the whole article — the passage: "Given the potential ramifications, these competencies might be treated as minimum safety requirements for youth-facing deployments of AI chatbots, with pre-deployment testing and ongoing audits anchored to standardized benchmarks such as AMHART."
- McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y - the article this story is about — the whole article — the passage: "As these tools advance, developers and implementers could examine whether domain-specific fine-tuning and explicit guardrails would enable more precise guidance that is developmentally appropriate, while still promoting human support and engagement in formal care."
McBain, R. K., Cantor, J. H., Hollister, B. et al. (2026). Development, validation and AI chatbot evaluation of the Adolescent Mental Health AI Response Tool. Nature Mental Health. https://doi.org/10.1038/s44220-026-00753-y
Who paid: The work was supported by a grant from the National Institute of Mental Health (1R01MH132551), and the article states that NIMH was not involved in the design or conduct of the study, the analysis or interpretation of the data, or the decision to publish.
Do not take this as professional medical advice.
The findings of other studies mentioned here come from this document; those studies were not opened one by one.

