A Chatbot Answers at Once. The Question Is Whether It Says How Soon, and Who Should Be There.

US · CanadaAmérica Latina

Weekly AI Watch

Every week, what changed and what it means for you. No hurry, no alarm.

Week of 30 October 2026EN | ESPast editions15,230 papers found · 14,455 read · 4,301 belonged · 46 stories passed every check

Observatorio de la
Inteligencia Artificial

Cada semana, qué cambió y qué significa para usted. Sin prisa y sin alarma.

Semana del 30 de octubre de 2026EN | ESEdiciones anteriores15,230 estudios encontrados · 14,455 leídos · 4,301 correspondían · 46 notas pasaron todas las comprobaciones

weeklyAI · Week of 30 October 2026weeklyAI · Semana del 30 de octubre de 2026

← Your parents← Sus padres

experiment · Frontiers in psychiatry · la publicación, 11 sep 2026 · gratis

Los chatbots contestan con soltura. En los casos de más riesgo, fallan más.

En una prueba con 90 preguntas escritas por investigadores en inglés, los tres modelos dieron respuestas aceptables en la mayoría de los casos, pero en las preguntas de mayor riesgo bajaron mucho.

Versión breve · la versión detallada sigue, unos 5 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: tres programas de conversación (ChatGPT, Gemini y Doubao) ante preguntas sobre depresión en la vejez
Cuántos: 90 preguntas
Dónde: China (equipo de investigadores)
Cuándo: julio de 2026
Tipo de estudio: experiment
Quién lo hizo: investigadores en China
El límite que importa: Las preguntas las escribieron los investigadores como prueba exigente, no las hicieron personas reales, y todo se preguntó en inglés.

Respuestas aceptables de cada programa en las preguntas de mayor riesgo

ChatGPT63.3%

Gemini50%

Doubao36.7%

Son las respuestas aceptables de cada programa entre las 30 preguntas de mayor riesgo de una prueba escrita por investigadores, no de consultas reales.

Respuestas aceptables en todas las preguntas frente a las de mayor riesgo

ChatGPT en todas las preguntasfrente aChatGPT en las de mayor riesgo

bajó de 78.9% a 63.3%

Gemini en todas las preguntasfrente aGemini en las de mayor riesgo

bajó de 72.2% a 50.0%

Doubao en todas las preguntasfrente aDoubao en las de mayor riesgo

bajó de 60.0% a 36.7%

La respuesta suena completa y serena, y aun así no dice en cuánto tiempo hay que buscar ayuda, quién debe quedarse acompañando ni qué hacer mientras se consigue la cita.
Lectura de weeklyAI

Ver · 0:52

Un hijo o una hija escribe en el teléfono: mi mamá dejó de comer y se le olvidan las cosas desde la semana pasada. El chatbot contesta en segundos, con frases tranquilas y bien ordenadas. La pregunta es si esa respuesta sirve.

Un equipo de investigadores escribió 90 preguntas sobre la depresión en la vejez. Sesenta las planteó como si las hiciera una persona mayor; treinta, como si las hiciera un familiar. Se las pasó a ChatGPT, Gemini y Doubao en julio de 2026 y dos psiquiatras calificaron las respuestas sin saber cuál venía de qué modelo.

ChatGPT dio respuestas aceptables al 78.9% de las preguntas, Gemini al 72.2% y Doubao al 60%. La diferencia entre ChatGPT y Gemini no fue concluyente; los dos superaron a Doubao.

El patrón cambia cuando el caso es grave. De las 30 preguntas de mayor riesgo, las respuestas aceptables fueron 63.3% en ChatGPT, 50% en Gemini y 36.7% en Doubao. Además, esos modelos mandaron a la persona a una atención menos urgente de la que hacía falta en 16.7%, 26.7% y 36.7% de esos casos.

Los errores graves de seguridad aparecieron en 5.6% de las respuestas de ChatGPT, 10% de Gemini y 16.7% de Doubao. Entre ellos estuvo no reconocer el riesgo de suicidio o no considerar que una confusión repentina podía tener una causa médica.

Esas cifras no anuncian lo que le pasará en su casa. Las preguntas las escribieron los investigadores como prueba exigente, no las hicieron personas reales. Se preguntaron en inglés, a una versión de cada aplicación. Treinta de ellas se preguntaron otra vez, en una conversación nueva, para ver si la respuesta cambiaba. Lo que su propia aplicación responda hoy, o en español, puede ser distinto.

Lo que el estudio sí muestra es la forma del fallo. La respuesta suena completa y serena, y aun así no dice en cuánto tiempo hay que buscar ayuda, quién debe quedarse acompañando ni qué hacer mientras se consigue la cita.

La próxima vez que consulte uno de estos programas, fíjese en si la respuesta dice cuándo y con quién, y si trata un cambio repentino en la memoria o en el apetito como motivo para llamar a un médico.

Qué significa para usted

Cuando su familiar le conteste, note si la respuesta dice en cuánto tiempo buscar ayuda y quién debe acompañar a la persona mientras llega la cita. Un cambio repentino en la memoria o en el apetito merece que consulte a un profesional, y eso el programa no lo decide por usted.

Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736

Quién pagó: Los autores declararon que no recibieron apoyo financiero para este trabajo ni para su publicación, y declararon no tener relaciones comerciales o financieras que pudieran constituir un conflicto de interés.

No tome esto como consejo médico profesional.

Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1032 palabras · unos 5 minLeerla →Cerrar

Tres chatbots respondieron preguntas sobre depresión en la vejez: acertaron en la mayoría, pero fallaron más cuando el caso era grave

Un estudio ciego con 90 preguntas encontró respuestas aceptables en el 78.9%, 72.2% y 60.0% de los casos según el modelo, con menos aciertos en los escenarios de mayor riesgo.

Imagine a una hija que escribe en su teléfono: "Mi mamá de 78 años dejó de comer y dice que ya no quiere seguir". La respuesta llega en segundos, con párrafos ordenados y tono amable. Parece escrita para ella.

Ese es el punto de partida del estudio que leemos esta semana. Un equipo de investigadores en China puso a prueba tres modelos de lenguaje de uso público —los programas de conversación que responden en ChatGPT, Gemini y Doubao— con 90 preguntas sobre depresión en personas mayores. Sesenta estaban escritas desde la perspectiva del propio adulto mayor y treinta desde la de un familiar o cuidador. Dos psiquiatras calificaron las respuestas sin saber cuál modelo las había escrito, y un tercero resolvió los desacuerdos importantes. Los autores declararon no haber recibido financiamiento y no tener conflictos comerciales.

Los tres modelos conversan, pero no "saben" medicina: generan respuestas directas que pueden parecer personalizadas, coherentes y con autoridad clínica, y por eso una respuesta puede sonar segura sin ser completa. A los modelos no se les pidió seguir ninguna guía clínica ni citar fuentes; se les hizo la pregunta tal cual, como la haría cualquier persona, y se calificó lo que contestaron¹.

Los resultados: ChatGPT dio respuestas clínicamente aceptables al 78.9% de las preguntas, Gemini al 72.2% y Doubao al 60.0%². La diferencia entre los dos primeros no fue concluyente; ambos superaron a Doubao. Y algo más importante: el desempeño bajó conforme subía la gravedad del caso³. En las preguntas de alto riesgo, los aciertos cayeron a 63.3%, 50.0% y 36.7%, y la recomendación de buscar atención resultó menos urgente de lo que hacía falta en 16.7%, 26.7% y 36.7% de las respuestas⁴.

Los errores graves de seguridad aparecieron en 5.6%, 10.0% y 16.7% de las respuestas⁵. El más común fue no reconocer o no escalar el riesgo suicida; también hubo fallas al no considerar delirium o un deterioro médico agudo, consejos inseguros sobre medicamentos recetados y poca urgencia ante el abandono severo del autocuidado o la mala alimentación⁶. En las respuestas de crisis, lo que más se omitió fue decir que la persona no debe quedarse sola e involucrar a un familiar de confianza⁷. La depresión en la vejez se asocia a un mayor uso de servicios de salud, a peores resultados de las enfermedades crónicas y a una mayor mortalidad.

La consistencia también varió: al repetir 30 preguntas en conversaciones nuevas, ChatGPT coincidió consigo mismo en el 90.0% de los casos, Gemini en 83.3% y Doubao en 73.3%⁸.

Conviene ver el tamaño del asunto. Los modelos de lenguaje ya se usan como fuente de información de salud⁹, y las preguntas típicas son justo las de esta prueba: si un síntoma es depresión, si hay que suspender un medicamento, si la pérdida de memoria es parte del envejecimiento o qué tan pronto hace falta ayuda profesional¹⁰. En muchas casas, quien nota primero los cambios es un hijo o una hija: el abandono del autocuidado, la comida, los medicamentos, el comportamiento o el manejo del dinero¹¹. Y a ese familiar le toca decidir si algo se puede vigilar en casa, si requiere revisión pronta o si es una emergencia¹².

No es la primera vez que se mide algo así. Según el resumen de un estudio sobre información para prevenir caídas —pudimos leer solo el resumen, el texto completo está detrás de una suscripción—, las respuestas de ChatGPT resultaron aceptables en general, aunque la facilidad de lectura variaba y los propios autores pidieron revisión experta antes de usarlas con pacientes¹³¹⁴¹⁵. Otro trabajo, también leído solo en resumen, probó un acompañamiento digital en 200 pacientes con hipertensión y reportó mejor control de la presión en el grupo con la herramienta que en el de atención habitual¹⁶¹⁷. Y una encuesta a 316 adultos de 55 años o más, de la que igualmente leímos solo el resumen, encontró que la ansiedad ante la tecnología era el freno principal para usar servicios digitales de salud¹⁸¹⁹.

Aquí está el límite que usted necesita: las 90 preguntas las escribieron los investigadores como prueba de esfuerzo, no se recogieron de personas mayores reales, y nadie midió cómo interpreta o usa un paciente esas respuestas²⁰. Todo se preguntó en inglés, una sola vez por pregunta, en conversaciones nuevas y sin repreguntar²¹. Lo que su propia aplicación conteste hoy, o en español, puede ser distinto. Las respuestas las calificaron dos psiquiatras y un tercero resolvió los desacuerdos.

Así lo leemos nosotros. En muchas familias ya no hay a quién preguntar dentro de casa sobre ánimo, memoria o medicinas, y entonces un chatbot ocupa ese lugar sin que sepamos quién lo entrenó ni qué intereses tiene. Es probable que eso siga creciendo. Sabríamos que nos equivocamos si las familias consultaran primero a alguien cercano y usaran la máquina solo como complemento.

Lo segundo que vemos: cuando una respuesta automática suena a veredicto, la decisión se vuelve técnica y nadie responde por ella. Usted puede cambiar eso con una pregunta simple: quién revisó esto y con qué criterios.

Y lo tercero: quien acompaña no decide por el otro. Si la herramienta le dice a un hijo qué hacer con su madre sin preguntarle a ella qué quiere, la máquina está sustituyendo, no acompañando. Antes de actuar, pregúntele a la persona mayor qué desea hacer; si hay duda sobre su capacidad de decidir, eso se consulta con un profesional.

Los propios autores lo dicen: estos sistemas podrían apoyar tareas educativas de bajo riesgo, pero no deben guiar por su cuenta una urgencia, un cambio de medicamento ni el manejo del riesgo suicida, y la supervisión clínica debe seguir siendo central²²²³.

Volvamos al teléfono de la hija. La respuesta que reciba puede ser útil para ordenar sus dudas y preparar la consulta, nada más. Antes de confiar en ella, vale una pregunta: ¿a quién de mi familia o de mi comunidad le pasó algo parecido, y qué hizo?

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

Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The prompt was designed to approximate an ordinary patient-facing interaction while maintaining a consistent communicative context across platforms. The models were not instructed to follow a named guideline, cite external sources, adopt a predetermined response structure, prioritize crisis recognition, or disclose internal reasoning."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Clinically acceptable responses were generated for 78.9% of questions by ChatGPT, 72.2% by Gemini, and 60.0% by Doubao (overall P<0.001)."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Performance declined substantially with increasing clinical risk."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Among high-risk questions, clinically acceptable response rates were 63.3%, 50.0%, and 36.7%, while under-triage occurred in 16.7%, 26.7%, and 36.7% of responses, respectively."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Major safety errors occurred in 5.6% of ChatGPT responses, 10.0% of Gemini responses, and 16.7% of Doubao responses (raw P = 0.015; FDR-adjusted q=0.023)."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The most common error involved inadequate recognition or escalation of suicide risk. Other errors included failure to consider delirium or acute medical deterioration, unsafe recommendations concerning prescribed medication, and insufficient escalation in the presence of severe self-neglect, poor oral intake, or potentially impaired decision-making capacity."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Advice that the older adult should not remain alone, explicit involvement of a trusted caregiver, and reduction of access to potentially lethal medication or other means of harm were the elements most frequently omitted."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Test-retest clinical consistency was highest for ChatGPT (90.0%), followed by Gemini (83.3%) and Doubao (73.3%)."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - el artículo del que trata esta nota — el artículo completo — el pasaje: "General-purpose large language models are increasingly available as sources of conversational health information ( 16 ). Unlike conventional search tools, these systems generate direct responses that may appear individualized, coherent, and clinically authoritative ( 17 )."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - el artículo del que trata esta nota — el artículo completo — el pasaje: "An older adult or caregiver can ask whether a symptom represents depression, whether a medicine should be stopped, whether memory decline is part of normal ageing, or how urgently professional help is required."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Family members and other informal caregivers often play a central role in recognizing late-life depression and facilitating access to care ( 13 ). Changes in self-care, eating, medication use, social behavior, or financial arrangements may first be noticed by someone other than the patient ( 14 )."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Caregivers may also need to decide whether a change can be monitored at home, requires an expedited clinical review, or represents an immediate psychiatric or medical emergency."
Arı M, İlçin N, Yağcıoğlu H, Akkar İ. (2026). Quality, readability, and patient safety of ChatGPT-generated responses to fall-related questions in older adults: a multidisciplinary evaluation. Revista da Associação Médica Brasileira. 10.1590/1806-9282.20260388 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "This study aimed to evaluate the quality, readability, and patient safety implications of ChatGPT-generated responses to common questions about fall risk and home safety in older adults."
Arı M, İlçin N, Yağcıoğlu H, Akkar İ. (2026). Quality, readability, and patient safety of ChatGPT-generated responses to fall-related questions in older adults: a multidisciplinary evaluation. Revista da Associação Médica Brasileira. 10.1590/1806-9282.20260388 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Three responses were rated as "excellent," while seven responses were rated as "satisfactory requiring minimal clarification." No response received a rating corresponding to "moderately satisfactory" or "unsatisfactory.""
Arı M, İlçin N, Yağcıoğlu H, Akkar İ. (2026). Quality, readability, and patient safety of ChatGPT-generated responses to fall-related questions in older adults: a multidisciplinary evaluation. Revista da Associação Médica Brasileira. 10.1590/1806-9282.20260388 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "While ChatGPT provided generally acceptable clinical information, variability in readability and expert ratings raises patient safety concerns. AI-generated health content should be reviewed and tailored to older adults' health literacy needs before clinical use."
Yang X, Huang X, Cui Y, Zhao K. (2026). Impact of artificial intelligence-assisted health management on blood pressure and self-management efficacy in older adults with hypertension. Blood Pressure. https://doi.org/10.1080/08037051.2026.2715847 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "the intervention group demonstrated a higher blood pressure control rate than the usual care group (88.66% vs. 64.89%)"
Yang X, Huang X, Cui Y, Zhao K. (2026). Impact of artificial intelligence-assisted health management on blood pressure and self-management efficacy in older adults with hypertension. Blood Pressure. https://doi.org/10.1080/08037051.2026.2715847 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Two hundred hypertensive patients were randomly assigned to a usual care group (routine health management) or an intervention group"
Alzghaibi H. (2026). Technology anxiety and digital health literacy as structural determinants of mHealth engagement among older adults: a SEM study in Saudi Arabia. Frontiers in Public Health. 10.3389/fpubh.2026.1872211 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "A cross-sectional survey was conducted with 316 community-dwelling adults aged 55 years and above across Saudi Arabia."
Alzghaibi H. (2026). Technology anxiety and digital health literacy as structural determinants of mHealth engagement among older adults: a SEM study in Saudi Arabia. Frontiers in Public Health. 10.3389/fpubh.2026.1872211 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Technology anxiety was the dominant inhibitor, directly suppressing both engagement intention"
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The benchmark consisted of standardized scenarios rather than questions entered prospectively by older adults or caregivers during actual healthcare decision-making. The study therefore assessed model responses under controlled conditions and did not measure how users would interpret, trust, or act on the advice."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - el artículo del que trata esta nota — el artículo completo — el pasaje: "All questions were presented in English. This improved standardization across models but may have influenced the performance of systems optimized for other languages or cultural contexts."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - el artículo del que trata esta nota — el artículo completo — el pasaje: "General-purpose large language models may support selected low-risk educational tasks but should not independently guide emergency triage, medication changes, suicide-risk management, or other safety-critical decisions in older adults."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Until these capabilities are demonstrated consistently across model versions, languages, and real-world settings, clinician oversight should remain central to the use of large language models in late-life depression care."

No tome esto como consejo médico profesional.

experiment · Frontiers in psychiatry · the paper, 11 Sep 2026 · free

A Chatbot Answers at Once. The Question Is Whether It Says How Soon, and Who Should Be There.

In a test of 90 written questions about depression in later life, the three chatbots gave acceptable answers 79, 72 and 60 percent of the time, and did worse on the riskiest ones.

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: Three consumer chatbots answering questions about depression in later life
How many: 90 questions
Where: China (authors' institutions); questions in English
When: July 2026
Kind of study: experiment
Who did it: Researchers at institutions in Guilin, Beijing, Chenzhou and Hangzhou, China
The limit that matters: The questions were written by researchers as a stress test, not asked by real families

Share of answers that met the researchers' standards, by chatbot

ChatGPT78.9%

Gemini72.2%

Doubao60%

These are shares of 90 researcher-written questions about depression in later life, not of questions real families asked.

Answers to the riskiest questions compared with answers to all questions

ChatGPT, all questionsagainstChatGPT, riskiest questions

Acceptable answers fell from 78.9 percent to 63.3 percent

Gemini, all questionsagainstGemini, riskiest questions

Acceptable answers fell from 72.2 percent to 50.0 percent

Doubao, all questionsagainstDoubao, riskiest questions

Acceptable answers fell from 60.0 percent to 36.7 percent

What the study suggests is that the answer on the screen is a good place to start a conversation and a poor place to end one.
weeklyAI's reading

Watch · 0:31

You type a question into a chatbot: a father who has stopped eating much, a mother who seems confused since a new pill, a low mood that will not lift. A fluent answer comes back in seconds.

Researchers wrote 90 questions like these about depression in later life. Sixty were written from the older person's side and 30 from a family member's, and each one went to ChatGPT, Gemini and Doubao in July 2026. Two psychiatrists scored the answers without knowing which chatbot wrote them, against standards fixed before the answers arrived.

ChatGPT's answers met those standards 78.9 percent of the time, Gemini's 72.2 percent and Doubao's 60.0 percent. The gap between ChatGPT and Gemini was not statistically significant.

The pattern mattered more than the ranking. On the 30 high-risk questions, acceptable answers fell to 63.3, 50.0 and 36.7 percent. The chatbots sent people to less urgent care than the question called for in 16.7, 26.7 and 36.7 percent of those cases. For emergencies, only 40.0, 26.7 and 13.3 percent of answers carried all five steps the researchers required, such as saying the person should not be left alone.

These were the researchers' own questions, built as a stress test. They did not come from real families, so the percentages do not tell you how often a chatbot would fail you at home. Each app was tested in two rounds in July 2026, in English, on one version. All the questions were in English, so the Doubao results say nothing about its answers in Chinese or in other languages. The researchers asked 30 of the questions a second time, in a fresh conversation, to see if the answers stayed the same. They did not ask follow-up questions or push back, so the study cannot say what happens when someone does.

A sudden change in memory or eating is one place to look. An answer that treats it as ordinary ageing, or that never names a time frame or a person to call, is the one to set aside. Ask the clinician instead.

What this means for you

Look for the time frame and the person to call. An answer that leaves out how soon to act, or that treats a sudden change in memory or eating as ordinary ageing, is the one to set aside. The study used written questions, not real families, so it cannot tell you how often a chatbot would fail you at home. For anything that worries you, the clinician is the one to ask.

Who paid: The authors declared that no financial support was received for this work or its publication, and they declared no commercial or financial relationships that could be a conflict of interest.

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 · 1395 words · about 7 minRead it →Close

A Chatbot Answered 90 Questions About Late-Life Depression. It Missed the Urgent Ones Most Often.

In a blinded test, three leading chatbots gave acceptable answers to most questions — and the quality fell as the danger rose.

Picture a kitchen table at ten at night. A daughter has just watched her father turn down dinner for the third time this week and forget where he put his keys. She types the question into her phone. The answer comes back in seconds, fluent and kind and sure of itself.

That is the scene the study was built around. General-purpose large language models are increasingly available as sources of conversational health information, and unlike a search engine they answer directly, in a voice that can seem individualized and clinically authoritative¹. An older adult or a caregiver can ask whether a symptom is depression, whether a medicine should be stopped, whether memory loss is normal ageing, or how soon a doctor is needed².

The authors are based at institutions in Guilin, Beijing, Chenzhou and Hangzhou, in China, and the study appears in Frontiers in Psychiatry in 2026. They wrote 90 questions about late-life depression: 60 in the voice of an older patient, 30 in the voice of a family member or caregiver. Each question went separately to three consumer chatbots — GPT-5.5 Instant through ChatGPT, Gemini 3.5 Flash through Gemini, and Seed2.0 Pro through Doubao — in fresh conversations, in English, in July 2026. Two psychiatrists, blind to which machine had answered, scored every reply against standards written in advance, with a third senior psychiatrist settling disagreements. The authors declared no financial support and no commercial conflicts of interest.

Clinically acceptable answers came back for 78.9 percent of questions from ChatGPT, 72.2 percent from Gemini and 60.0 percent from Doubao³. ChatGPT and Gemini were not shown to differ meaningfully from each other; both did better than Doubao. Performance declined substantially as clinical risk rose⁴. On the 30 highest-risk questions — the ones describing something that may need emergency care — acceptable answers fell to 63.3, 50.0 and 36.7 percent, and under-triage, meaning advice less urgent than the situation called for, occurred in 16.7, 26.7 and 36.7 percent of responses⁵.

Major safety errors were found in 5.6 percent of ChatGPT replies, 10.0 percent of Gemini replies and 16.7 percent of Doubao replies⁶. The most common was failure to recognize or escalate suicide risk; others were missing delirium or acute medical deterioration, unsafe advice about prescribed medication, and too little escalation when there was severe self-neglect, poor eating or doubtful decision-making capacity⁷. In crisis answers, the elements most often left out were advice that the older person should not be left alone, explicit involvement of a trusted caregiver, and reducing access to potentially lethal medication or other means of harm⁸.

Consider what a caregiver is actually asking in that moment: whether a change can be watched at home, needs a rushed appointment, or is an emergency now⁹. Family members are usually the ones who notice the changes first — in self-care, eating, medication use, social behavior or money¹⁰. So the question is not abstract. It is the ordinary work of looking after someone.

Using a chatbot is also easy to picture, and the study made it deliberately plain. Each question was sent as an ordinary person would send it, with no instruction to follow a guideline, cite sources, use a set structure or put crisis first¹¹. That is the everyday use case, and it is also the reason the results matter: no one had tuned the machine to be careful.

Two limits shape what the numbers can mean. The 90 questions were written by the researchers as a deliberate stress test, not collected from real older people or caregivers, and the study did not measure how anyone would understand, trust or act on an answer¹². And every question was in English, which may have affected a system built for other languages or cultures¹³. The authors also note, as a further limitation, that the study had little power to catch rare safety events, small differences between machines, or patterns within each of its six topic areas of 15 questions.

This belongs to a wider family of work on whether machines can help older people manage their health. One summary of a study of 200 people with high blood pressure, which we could read only in its abstract, found that an AI-assisted management program produced a higher rate of blood pressure control than usual care, 88.66 percent against 64.89 percent¹⁴, along with better diet, exercise and smoking measures¹⁵. Another summary, also abstract only, looked at ChatGPT's answers on preventing falls at home and rated seven of ten as satisfactory or excellent, while five were written harder than patient materials usually are¹⁶¹⁷. A third, abstract only, surveyed 316 adults aged 55 and over in Saudi Arabia and found technology anxiety to be the main thing holding people back from digital health services¹⁸¹⁹.

Read side by side, those summaries and this study point in different directions, and that is worth holding. The blood pressure program was designed and supervised as a program; the fall-prevention review found the information decent but the reading level uneven; the Saudi survey found the barrier was often the person's unease with the tool itself¹⁹. What the new study adds is a test of the tool as it sits in an app store today, with no program around it.

Here is how we read it. A machine that answers warmly and seems to listen invites a kind of trust that a printed leaflet never did. That is our expectation, not a finding: readers and their adult children may take a fluent answer about low mood or memory as if a clinician had weighed in, and may not think to check it. We would know we are wrong if families told us they routinely read such answers skeptically and only used them to prepare questions. Until then, the useful habit is to notice what the answer did not ask about — falls, confusion, other medicines, who is nearby — and carry those gaps to the appointment.

We also read a second thing into the setup. A conversation held in a chat window can be kept, and a new chat can start clean. We cannot say what any particular app stores. What we can say is that starting fresh and saying the key facts out loud, rather than assuming the machine remembers them, costs nothing.

One more, about honesty at the table. When a system is watching for warning signs, people sometimes learn to say what sounds acceptable and to leave out what might count against them. If that happens here, the calmest account may be the least complete one — to the machine and to the doctor alike. Saying the difficult things plainly, and encouraging a parent to do the same, is the counterweight.

What would have to happen next is already written into the study's own conclusion. The authors say these models may support selected low-risk educational tasks but should not independently guide emergency triage, medication changes, suicide-risk management or other safety-critical decisions in older adults²⁰. They add that until the abilities are shown consistently across model versions, languages and real-world settings, clinician oversight should remain central²¹. Note the phrase "model versions": each chatbot was tested on one version, through one public interface, and the authors warn that results may not carry to other versions or to specialist setups.

So what does this leave you with, practically? Not a verdict on any app, and not a reason to avoid one. A short list of questions to ask of any health answer that arrives on a screen: Did it ask about falls or confusion? Did it mention the other medicines? Did it say who should be with the person, and how soon? If the answer is vague on any of those, it is a draft, not a decision. Bring the draft to a clinician.

The daughter at the kitchen table still has her phone in her hand. What the study suggests is that the answer on the screen is a good place to start a conversation and a poor place to end one.

Next time a chatbot answers a health question for you or a parent, ask it: what should make us go to a doctor today rather than next week?

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

Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - the article this story is about — the whole article — the passage: "General-purpose large language models are increasingly available as sources of conversational health information ( 16 ). Unlike conventional search tools, these systems generate direct responses that may appear individualized, coherent, and clinically authoritative ( 17 )."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - the article this story is about — the whole article — the passage: "An older adult or caregiver can ask whether a symptom represents depression, whether a medicine should be stopped, whether memory decline is part of normal ageing, or how urgently professional help is required."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - the article this story is about — the whole article — the passage: "Clinically acceptable responses were generated for 78.9% of questions by ChatGPT, 72.2% by Gemini, and 60.0% by Doubao (overall P<0.001)."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - the article this story is about — the whole article — the passage: "Performance declined substantially with increasing clinical risk."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - the article this story is about — the whole article — the passage: "Among high-risk questions, clinically acceptable response rates were 63.3%, 50.0%, and 36.7%, while under-triage occurred in 16.7%, 26.7%, and 36.7% of responses, respectively."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - the article this story is about — the whole article — the passage: "Major safety errors occurred in 5.6% of ChatGPT responses, 10.0% of Gemini responses, and 16.7% of Doubao responses (raw P = 0.015; FDR-adjusted q=0.023)."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - the article this story is about — the whole article — the passage: "The most common error involved inadequate recognition or escalation of suicide risk. Other errors included failure to consider delirium or acute medical deterioration, unsafe recommendations concerning prescribed medication, and insufficient escalation in the presence of severe self-neglect, poor oral intake, or potentially impaired decision-making capacity."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - the article this story is about — the whole article — the passage: "Advice that the older adult should not remain alone, explicit involvement of a trusted caregiver, and reduction of access to potentially lethal medication or other means of harm were the elements most frequently omitted."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - the article this story is about — the whole article — the passage: "Caregivers may also need to decide whether a change can be monitored at home, requires an expedited clinical review, or represents an immediate psychiatric or medical emergency."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - the article this story is about — the whole article — the passage: "Family members and other informal caregivers often play a central role in recognizing late-life depression and facilitating access to care ( 13 ). Changes in self-care, eating, medication use, social behavior, or financial arrangements may first be noticed by someone other than the patient ( 14 )."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - the article this story is about — the whole article — the passage: "The prompt was designed to approximate an ordinary patient-facing interaction while maintaining a consistent communicative context across platforms. The models were not instructed to follow a named guideline, cite external sources, adopt a predetermined response structure, prioritize crisis recognition, or disclose internal reasoning."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - the article this story is about — the whole article — the passage: "The benchmark consisted of standardized scenarios rather than questions entered prospectively by older adults or caregivers during actual healthcare decision-making. The study therefore assessed model responses under controlled conditions and did not measure how users would interpret, trust, or act on the advice."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - the article this story is about — the whole article — the passage: "All questions were presented in English. This improved standardization across models but may have influenced the performance of systems optimized for other languages or cultural contexts."
Yang X, Huang X, Cui Y, Zhao K. (2026). Impact of artificial intelligence-assisted health management on blood pressure and self-management efficacy in older adults with hypertension. Blood Pressure. https://doi.org/10.1080/08037051.2026.2715847 — only the abstract - the full text could not be fetched — the passage: "the intervention group demonstrated a higher blood pressure control rate than the usual care group (88.66% vs. 64.89%)"
Yang X, Huang X, Cui Y, Zhao K. (2026). Impact of artificial intelligence-assisted health management on blood pressure and self-management efficacy in older adults with hypertension. Blood Pressure. https://doi.org/10.1080/08037051.2026.2715847 — only the abstract - the full text could not be fetched — the passage: "superior improvements in smoking behavior change, alcohol consumption behavior change, low-salt and low-fat diet adherence, and regular exercise attainment"
Arı M, İlçin N, Yağcıoğlu H, Akkar İ. (2026). Quality, readability, and patient safety of ChatGPT-generated responses to fall-related questions in older adults: a multidisciplinary evaluation. Revista da Associação Médica Brasileira. 10.1590/1806-9282.20260388 — only the abstract - the full text could not be fetched — the passage: "Three responses were rated as "excellent," while seven responses were rated as "satisfactory requiring minimal clarification." No response received a rating corresponding to "moderately satisfactory" or "unsatisfactory.""
Arı M, İlçin N, Yağcıoğlu H, Akkar İ. (2026). Quality, readability, and patient safety of ChatGPT-generated responses to fall-related questions in older adults: a multidisciplinary evaluation. Revista da Associação Médica Brasileira. 10.1590/1806-9282.20260388 — only the abstract - the full text could not be fetched — the passage: "The mean Flesch-Kincaid Grade Level was 8.4 (range 4.3-11.9). Five responses exceeded the readability levels commonly recommended for patient education materials."
Alzghaibi H. (2026). Technology anxiety and digital health literacy as structural determinants of mHealth engagement among older adults: a SEM study in Saudi Arabia. Frontiers in Public Health. 10.3389/fpubh.2026.1872211 — only the abstract - the full text could not be fetched — the passage: "A cross-sectional survey was conducted with 316 community-dwelling adults aged 55 years and above across Saudi Arabia."
Alzghaibi H. (2026). Technology anxiety and digital health literacy as structural determinants of mHealth engagement among older adults: a SEM study in Saudi Arabia. Frontiers in Public Health. 10.3389/fpubh.2026.1872211 — only the abstract - the full text could not be fetched — the passage: "Technology anxiety was the dominant inhibitor, directly suppressing both engagement intention"
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - the article this story is about — the whole article — the passage: "General-purpose large language models may support selected low-risk educational tasks but should not independently guide emergency triage, medication changes, suicide-risk management, or other safety-critical decisions in older adults."
Xiao, W., Zhang, H., Chen, X. et al. (2026). Large language models for late-life depression: a blinded benchmark of clinical safety, geriatric appropriateness, and triage. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1956736 - the article this story is about — the whole article — the passage: "Until these capabilities are demonstrated consistently across model versions, languages, and real-world settings, clinician oversight should remain central to the use of large language models in late-life depression care."

Do not take this as professional medical advice.

The free weekly drip by email

One email a week: the free story of every section, with its picture, and the headlines of the rest.

No payment, nothing to cancel. Unsubscribe from any issue.

El goteo semanal gratis por correo

Un correo por semana: la nota gratuita de cada sección, con su imagen, y los titulares de las demás.

Sin pago y sin nada que cancelar. Puede darse de baja desde cualquier envío.