Weekly AI Watch

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

Week of 20 November 2026 | Past editions15,959 papers found · 15,002 read · 5,264 belonged · 42 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 20 de noviembre de 2026 | Ediciones anteriores15,959 estudios encontrados · 15,002 leídos · 5,264 correspondían · 42 notas pasaron todas las comprobaciones

weeklyAI · Week of 20 November 2026weeklyAI · Semana del 20 de noviembre de 2026

← Your parents← Sus padres

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

A los bots les va bien con lo sencillo y fallan cuando la pregunta es urgente

En una prueba con 90 preguntas escritas sobre depresión en la vejez, entre el 37% y el 63% de las respuestas a los casos más delicados pasaron el filtro. Eran guiones, en inglés, y no personas reales.

Versión breve · la versión detallada sigue, unos 6 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 chatbots de uso general: ChatGPT, Gemini y Doubao
Cuántos
90 preguntas
Dónde
China (equipo de universidades en Guilin, Pekín, Chenzhou y Hangzhou)
Cuándo
julio de 2026
Tipo de estudio
experiment
Quién lo hizo
Universidades de Guilin, Pekín, Chenzhou y Hangzhou, China
El límite que importa
Fueron preguntas escritas en inglés, no personas reales; no mide cómo actúa la gente.
Respuestas aceptables en las preguntas más urgentes
ChatGPT63.3%
Gemini50%
Doubao36.7%

Porcentaje de respuestas clínicamente aceptables en los 30 casos de mayor urgencia; son escenarios escritos en inglés, no pacientes reales.

Qué tan seguido el bot recomendó menos urgencia de la que hacía falta, en los casos más delicados

ChatGPTfrente aGemini

16.7% frente a 26.7% de los casos

Geminifrente aDoubao

26.7% frente a 36.7% de los casos

Cuando un chat le conteste sobre el ánimo, la memoria o la seguridad de una persona mayor, fíjese en dos cosas: si dice en cuánto tiempo hay que buscar ayuda, y quién debe quedarse con esa persona mientras tanto.
Lectura de weeklyAI

Tres modelos de chat contestaron 90 preguntas sobre la depresión en personas mayores. En los casos más urgentes, a los tres les fue peor.

Un equipo de psiquiatras y un médico geriatra de universidades en Guilin, Pekín, Chenzhou y Hangzhou, China, escribió esas preguntas como las haría una persona mayor o un familiar. Las bloquearon antes de probar los bots. Cada pregunta entró en una conversación nueva a ChatGPT, Gemini y Doubao. Dos psiquiatras calificaron las 270 respuestas sin saber cuál venía de qué máquina, y un tercero resolvió los desacuerdos.

En el conjunto completo, ChatGPT dio respuestas clínicamente aceptables al 78.9% de las preguntas, casi ocho de cada diez. Gemini, al 72.2%, y Doubao, al 60%, seis de cada diez. La diferencia entre ChatGPT y Gemini no fue clara; ambos superaron a Doubao.

En las preguntas más urgentes, las respuestas aceptables bajaron al 63.3%, 50.0% y 36.7%. Es decir, en el escenario más delicado, Doubao acertó poco más de una vez en cada tres. La "subtriage", cuando el bot recomienda menos urgencia de la que hace falta, corrió en 16.7%, 26.7% y 36.7% de esos casos. En esos casos urgentes, los errores graves de seguridad fueron 13.3%, 20.0% y 33.3%. El más común fue no reconocer el riesgo de suicidio.

Esto no describe a pacientes reales. Nadie fue estudiado: fueron escenarios escritos, en inglés, probados en julio de 2026 en las apps de consumidor. Tampoco dice cómo le iría a un bot en español o portugués, ni en versiones posteriores. Además, 30 de las preguntas se volvieron a hacer en conversaciones nuevas. Las respuestas de un mismo bot no siempre coincidieron: entre el 10% y el 20% de los pares cambiaron de un modo que podía cambiar lo que usted hiciera.

El estudio no siguió a nadie después. No puede decir si una buena respuesta hizo que alguien buscara ayuda a tiempo, ni si una mala la retrasó.

Cuando un chat le conteste sobre el ánimo, la memoria o la seguridad de una persona mayor, fíjese en dos cosas: si dice en cuánto tiempo hay que buscar ayuda, y quién debe quedarse con esa persona mientras tanto.

Qué significa para usted

Cuando un chat le dé una respuesta sobre el ánimo, la memoria o la seguridad de una persona mayor, vea si menciona al cuidador y qué tanto tarda en recomendar ayuda: en los casos urgentes fue justo ahí donde más fallaron los tres programas. Y si la persona mayor no puede valerse sola, esa respuesta no basta; quien puede valorarla es un profesional de la salud.

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 se recibió 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 · 1274 palabras · unos 6 minLeerla →Cerrar

Tres chatbots ante la depresión en la vejez: aciertan mucho, fallan en lo urgente

Un estudio ciego con 90 preguntas encontró respuestas clínicamente aceptables en la mayoría de los casos, pero la seguridad se cae justo cuando el caso se pone grave.

Un hijo pregunta desde otra ciudad si el desánimo de su madre es depresión o parte de la edad. Una mujer de setenta años pregunta si puede dejar la pastilla que le recetaron. Son consultas que la gente ya le hace a los chatbots, y un equipo de investigadores en China quiso medir qué tan bien contestan. Tres modelos de uso general —ChatGPT, Gemini y Doubao— respondieron 90 preguntas escritas por el equipo sobre depresión en la vejez, todas en inglés, cada una en una conversación nueva. Dos psiquiatras calificaron las respuestas sin saber de qué modelo venían, con un tercero para desempatar. Las preguntas se repartieron en partes iguales entre riesgo bajo, medio y alto, y cubrieron seis temas, desde síntomas y memoria hasta medicamentos y emergencias.

La respuesta fue clínicamente aceptable en 78.9% de las preguntas para ChatGPT, 72.2% para Gemini y 60.0% para Doubao. La diferencia entre ChatGPT y Gemini no fue concluyente; los dos superaron a Doubao. El orden no cambió al aplicar reglas más estrictas ni más laxas. El equipo de investigación está en universidades de Guilin, Beijing, Chenzhou y Hangzhou, y los autores declararon no haber recibido financiamiento ni tener conflictos comerciales.

El dato que importa a una familia aparece al separar por gravedad. En las preguntas de riesgo alto, la respuesta aceptable bajó a 63.3%, 50.0% y 36.7%. Ahí también se concentraron los errores de seguridad graves: 5.6%, 10.0% y 16.7% de todas las respuestas. El error más frecuente fue no reconocer o no escalar el riesgo suicida; también aparecieron casos de no considerar un delirium o un deterioro médico repentino, consejos inseguros sobre medicamentos y falta de escalamiento ante el abandono severo del autocuidado. En las respuestas de crisis, los elementos que más se omitieron fueron decir que la persona no debe quedarse sola, involucrar a un cuidador de confianza y reducir el acceso a medicamentos u otros medios peligrosos1.

Hubo dos pruebas más. Treinta preguntas se repitieron en conversaciones separadas: las respuestas coincidieron en lo clínico 90.0%, 83.3% y 73.3% de las veces. Y a los modelos les faltó, sobre todo, decirle al cuidador qué hacer en concreto mientras llega la ayuda. Los propios autores lo advierten: esto fue un ejercicio con preguntas escritas, no con pacientes reales, y no mide cómo interpreta o usa la gente lo que recibe.

Piense en la escena: una hija le escribe al chatbot a las once de la noche porque no consiguió cita. La respuesta llega completa, amable, segura de sí misma. Lo curioso es que esa seguridad es parte del problema. El artículo describe un estudio en el que los modelos de lenguaje generan respuestas directas que pueden parecer individualizadas, coherentes y con autoridad clínica2. No buscan una página: redactan un consejo. Y el artículo describe un estudio en el que la depresión en la vejez rara vez llega como tristeza declarada; puede asomar como fatiga, falta de apetito, quejas del cuerpo, menos actividad o problemas de memoria3. Por eso una respuesta correcta sobre la depresión en general puede ser insuficiente para una persona mayor.

El artículo describe un estudio en el que la familia suele ser quien primero nota el cambio: el cuidado personal, la comida, los medicamentos, el retiro social o el manejo del dinero4. Y el artículo describe un estudio en el que las preguntas típicas son muy concretas: si un síntoma es depresión, si hay que suspender un medicamento, si la pérdida de memoria es normal o qué tan urgente es buscar ayuda5. Ahí está el detalle: distinguir entre "consulte a su médico" y "llame ahora" es una decisión clínica, no un dato.

El artículo describe otros estudios que ya habían evaluado estos sistemas en salud mental, pero solían medir exactitud, calidad de la respuesta o empatía, y eso no alcanza para la depresión en la vejez6. Este estudio agregó criterios geriátricos, niveles de urgencia y repetición de respuestas. Y en el terreno de la salud, hay intentos parecidos con resultados alentadores pero preliminares. Según el resumen de un estudio sobre entrenamiento cognitivo con un agente conversacional, lo único disponible de él, los participantes mejoraron su función cognitiva7, aunque la adherencia fue baja8. Según el resumen de un estudio sobre hipertensión con chatbot, lo único disponible, hubo mejor control de la presión, pero el diseño no era aleatorizado y los hallazgos son preliminares9. Según el resumen de otro trabajo, lo único disponible, el control de la presión fue mayor con gestión asistida por inteligencia artificial10. Son señales preliminares de otros estudios, no resultados de este trabajo, que solo calificó respuestas escritas a preguntas sobre depresión en la vejez.

Así lo lee weeklyAI: el patrón es el de un aparato que responde siempre igual, sin cansarse, y al que se le va entregando la decisión que antes tomaba una persona. Es probable que un adulto mayor o su hijo reciban esa respuesta como una consulta ya resuelta y no busquen una segunda opinión cuando el tema es delicado, aunque el estudio no midió cómo reacciona la gente. Esta lectura se equivocaría si las familias que consultan el chatbot igual llaman al médico o a un familiar antes de actuar, o si tratan la respuesta solo como orientación para preparar preguntas. Para eso sirve: leerla como un borrador. Antes de seguir un consejo sobre ánimo, memoria o si hay que llamar al doctor, contrástelo con una persona de confianza o con el servicio de salud, y pregúntese en voz alta: ¿esto lo decidiría yo solo con una máquina?

Hay otra escena que conviene mirar: la del hijo que vive lejos. Es probable que sea él quien consulte sobre el ánimo de su madre y que la decisión de llamar al doctor dependa de esa consulta a distancia más que de lo que ella misma observa. Se equivocaría esta lectura si en la práctica es el propio adulto mayor quien consulta y decide, o si la familia consulta en conjunto y comparte la respuesta. Lo que sí puede hacerse: contarle al adulto mayor lo que se preguntó y lo que le respondieron, y decidir juntos el siguiente paso. Y si se usa el chatbot porque no hay cita pronto, anote lo que respondió y llévelo a la próxima consulta en lugar de quedarse con él.

El estudio tiene límites claros. Las 90 preguntas las escribió el equipo como prueba de esfuerzo; no salieron de conversaciones reales, así que los porcentajes no dicen con qué frecuencia un chatbot acertaría la pregunta de su familia. Todo se preguntó en inglés, por las aplicaciones de consumo, en julio de 2026: los resultados podrían no repetirse en español ni en versiones posteriores. Treinta preguntas se repitieron en conversaciones separadas, pero la comparación principal usó solo la primera respuesta, así que la respuesta que usted reciba es una sola tirada, no el promedio que reporta el estudio. Con 90 preguntas no se puede estimar bien lo raro, y por eso los autores piden no usar estos sistemas para decidir urgencias, cambios de medicamento, riesgo suicida o capacidad de decisión11.

Lo que esto deja a una familia se puede resumir en una costumbre: cuando el tema sea el ánimo, la memoria o la urgencia de una consulta, use el chatbot para ordenar lo que quiere preguntarle al médico, no para decidir. Llévelo a la mesa: ¿esta respuesta la estamos usando para acercarnos al doctor o para no hablar del tema?

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

  1. 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."
  2. 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 )."
  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: "Despite these consequences, late-life depression remains frequently under-recognized and undertreated ( 3 ). Older adults may not describe their difficulties primarily in emotional terms, and depressive symptoms may instead present through fatigue, sleep disturbance, loss of appetite, bodily complaints, reduced activity, subjective memory problems, or gradual withdrawal from previously meaningful roles ( 4 )."
  4. 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 )."
  5. 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."
  6. 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: "Existing evaluations of large language models in mental health have generally emphasized factual accuracy, response quality, empathy, or agreement with professional recommendations ( 30 ). These outcomes are informative but do not fully capture the demands of late-life depression care."
  7. Yun B, Kim B, Ko H, Kim J, Kim BR, Kim W, et al. (2026). Enhancing Cognitive Functions of Older Adults With Software Robot: Longitudinal Exploratory Field Study. JMIR mHealth and uHealth. 10.2196/75308 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Participants showed improved cognitive function after the intervention (Hedges g=0.350, P<.001)."
  8. Yun B, Kim B, Ko H, Kim J, Kim BR, Kim W, et al. (2026). Enhancing Cognitive Functions of Older Adults With Software Robot: Longitudinal Exploratory Field Study. JMIR mHealth and uHealth. 10.2196/75308 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Participants showed low adherence, which indicates a need for strategies that sustain long-term use."
  9. Ransinyo K, Banharak S, Panpanit L, Thiengtham S, Visetsittikul P, Sommana C, et al. (2026). Effectiveness of a Chatbot-Based Hypertension Management Program on Selected Outcomes Among Community-Dwelling Older Adults: A Quasi-Experimental Study. Journal of Multidisciplinary Healthcare. 10.2147/jmdh.s635820 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "A chatbot-supported hypertension management program was associated with better self-management and blood pressure control among community-dwelling older adults, although the non-randomized design and recruitment from different settings mean that the findings are preliminary."
  10. 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. 10.1080/08037051.2026.2715847 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "Following one year intervention, the intervention group demonstrated a higher blood pressure control rate than the usual care group (88.66% vs. 64.89%), along with lower SBP (128.49 ± 6.63 vs. 135.38 ± 7.31 mmHg) and DBP (77.49 ± 6.07 vs. 81.84 ± 5.44 mmHg) (all"
  11. 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

Quién pagó: Los autores declararon que no se recibió 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.

Los hallazgos de otros estudios que aquí se mencionan provienen de este documento; esos estudios no se consultaron uno por uno.

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

Three Chatbots, Ninety Hard Questions, and the Answers That Fell Short

In a test of how three AI chatbots answer questions about depression in later life, the best one gave a clinically acceptable answer to about three in four questions, and to fewer than two in three when the situation was urgent.

Short version · the longer version follows, about 6 min

Ask the weeklyAI chat

Ask me about AI in your life: I point you to the stories weeklyAI has published on it, this week and before.

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

The study at a glance
Who
Three AI chatbots answering questions about depression in later life
How many
90 questions
Where
China
When
July 2026
Kind of study
experiment
Who did it
Universities in Guilin, Beijing, Chenzhou and Hangzhou, China
The limit that matters
Written scenarios, not real chats; no one was followed to see what they did.
Share of answers judged clinically acceptable, by chatbot
ChatGPT78.9%
Gemini72.2%
Doubao60%

These are shares of 90 written test questions, not of real conversations, so they describe the test and not your family's next question.

Answers judged clinically acceptable on the most urgent questions

ChatGPTagainstGemini

63.3% of urgent answers were acceptable for ChatGPT, 50.0% for Gemini

GeminiagainstDoubao

50.0% of urgent answers were acceptable for Gemini, 36.7% for Doubao

When a chatbot answers a question about an older person's mood, memory or safety, look for whether it says how soon to get help, and who should stay with the person until then.
weeklyAI's reading
How it could look · illustration generated by weeklyAI.watch, not a photograph

Take 90 questions, the kind an older person or an adult son or daughter might type into a phone at a kitchen table, about low mood, memory that has slipped, a medicine that seems to be causing trouble, a parent who has stopped eating. Then ask three chatbots each one.

That is what a team based at universities in Guilin, Beijing, Chenzhou and Hangzhou, China, did. The study team wrote the questions from the point of view of someone 65 or older, or of a family caregiver. An expert panel then reviewed them: two psychiatrists, a senior psychiatrist and a geriatric physician. Fifteen were dropped as unclear or duplicative, leaving 90. The team fixed the standards for a good answer before any chatbot replied.

Each question went to ChatGPT, Gemini and Doubao, each time in a fresh conversation with no follow-up. Two psychiatrists, kept unaware of which chatbot wrote which answer, scored all 270 replies. A third senior psychiatrist settled disagreements.

About three in four ChatGPT answers were clinically acceptable, 71 of 90. Gemini reached 65, about seven in ten. Doubao reached 54, six in ten. The gap between ChatGPT and Gemini was not statistically significant.

Among the 30 most urgent scenarios, acceptable answers fell to 19 for ChatGPT, 15 for Gemini and 11 for Doubao. Under-triage, meaning the chatbot suggested less urgent care than the situation required, ran at 5, 8 and 11 of those 30.

The errors were not small. Major safety errors appeared in 5.6 percent of ChatGPT answers, 10.0 percent of Gemini's and 16.7 percent of Doubao's. The most common was failing to recognize suicide risk. Asked the same question twice in separate conversations, the chatbots gave the same clinical advice in 27 of 30 cases for ChatGPT, 25 for Gemini and 22 for Doubao.

These were written scenarios in English, asked in July 2026 through consumer apps. They were built as a stress test, not drawn from real chats, so the percentages describe the test, not your family's next question.

The study followed no one afterward. It cannot say whether a good or poor answer changed what anyone actually did.

When a chatbot answers a question about an older person's mood, memory or safety, look for whether it says how soon to get help, and who should stay with the person until then.

What this means for you

A chatbot's answer can sound sure and still skip the part that matters, such as how soon to get help and who should stay with the person until then. These three were tested on written questions in English, so the results say nothing about how they handle a real family's evening. A professional is the one to ask about anyone's mood, memory or medicine.

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

Who paid: The authors declared that financial support was not received for this work and/or its publication, and they declared no commercial or financial relationships that could be construed as a potential 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 · 1226 words · about 6 minRead it →Close

Three Chatbots Were Asked 90 Questions About Late-Life Depression. They Missed the Most Urgent Ones.

A blinded test of ChatGPT, Gemini and Doubao found the answers got worse as the situations got more dangerous.

Depressive symptoms in later life may show up as fatigue, sleep disturbance, loss of appetite, bodily complaints, reduced activity, memory problems, or gradual withdrawal from roles that once mattered. An older adult or a caregiver can ask a chatbot whether a symptom is depression, whether a medicine should be stopped, whether memory decline is just ageing, or how urgently help is needed. These systems generate direct answers that may appear individualized, coherent and clinically authoritative.

That fluency is the point of these systems, and also the problem. General-purpose large language models are increasingly used as sources of conversational health information, and unlike a search engine they generate direct answers that may appear individualized, coherent and clinically authoritative1. An older adult or a caregiver can ask whether a symptom is depression, whether a medicine should be stopped, whether memory decline is just ageing, or how urgently help is needed2.

A team whose authors list affiliations in Guilin, Beijing, Chenzhou and Hangzhou, China, put these systems to a test. The authors declared no financial support and no commercial conflicts. They wrote 90 questions about depression in later life, 60 in the voice of an older patient and 30 in the voice of a family caregiver, and balanced them across 30 low-, 30 moderate- and 30 high-risk situations3. Each question went separately to GPT-5.5 Instant through ChatGPT, Gemini 3.5 Flash through Gemini, and Seed2.0 Pro through Doubao, in fresh conversations, with no follow-up and no special instructions. Two psychiatrists rated the anonymized answers against standards fixed before any answer was generated, and a third settled disagreements.

Clinically acceptable answers came for 78.9 percent of questions from ChatGPT, 72.2 percent from Gemini and 60.0 percent from Doubao4. The gap between the top two was not statistically significant; both beat Doubao. The ranking held when the researchers re-scored with a looser and a stricter definition of a good answer.

The averages hide the pattern that matters. Performance fell as the stakes rose: on the high-risk questions, acceptable answers dropped to 63.3 percent, 50.0 percent and 36.7 percent5. Under-triage, meaning advice less urgent than the situation required, ran at 16.7, 26.7 and 36.7 percent5. Major safety errors appeared in 5.6, 10.0 and 16.7 percent of answers6. The most common error was failing to recognize or escalate suicide risk; others included missing delirium or an acute medical problem, unsafe advice about prescribed medicines, and too little urgency when self-neglect or poor intake was described7.

Complete attention to the realities of older age, things like cognitive change, several illnesses at once, many medicines, frailty, falls and dependence on a caregiver, was reached in 64.4, 55.6 and 43.3 percent of answers. When the situation was a crisis, all five expected elements were present in only 40.0, 26.7 and 13.3 percent. The elements most often left out: that the person should not be alone, that a trusted family member should be brought in, and that access to dangerous medication should be reduced8.

Asked the same 30 questions again in separate conversations, the models gave clinically consistent answers 90.0, 83.3 and 73.3 percent of the time9. That means one in ten to one in five repeated pairs differed in ways that could change what a family actually does. The study used written scenarios, not real patients or real chats, and did not measure what anyone did with the advice10.

Consider what a caregiver is really asking when she types that question about her father. Changes in self-care, eating, medication use, social behavior or money matters are often first noticed by someone other than the patient11. That person may be deciding whether to watch at home, call the clinic tomorrow, or act now. A caregiver's question is a triage question wearing ordinary clothes.

Earlier evaluations of these systems in mental health have mostly measured factual accuracy, response quality, empathy, or agreement with professional recommendations12. Those measures are useful but do not capture what late-life depression demands, because the same symptom means different things in a medically stable, independent person and in a frail one with memory change. The new study's contribution is the standard it applied: each answer judged on accuracy, safety, age-specific fit, warning signs and urgency, all at once.

The same technology is being tested in gentler roles. In one study, older adults used an artificial intelligence conversational agent that combined social interaction, reward-based engagement and cognitive training; 133 participants with no dementia diagnosis took part and showed improved cognitive function after the program131415. Participants who engaged more improved more16. The program also increased social support and lowered depression scores1718. According to the summary, the only part of that paper available, cognitive training alone did not produce a significant effect19.

Health management is another use under study. According to the summary of a non-randomized study of community-dwelling older adults with high blood pressure, the only part of it available, a 12-week chatbot program combining education, monitoring and lifestyle counseling was linked to better knowledge, health behaviors, medication adherence and blood pressure control than usual care202122. The authors called the findings preliminary because of the design23. A randomized study of 200 patients, known only from its abstract, reported that a year of artificial-intelligence-assisted management produced better blood pressure control than routine care2425.

Here is weeklyAI's reading. Two things are happening in the same room. A machine that answers in fluent, confident prose is useful for explaining, reminding and encouraging. It is unreliable at the moment a family needs judgement about danger, because that judgement depends on details a typed paragraph leaves out. A smooth answer about memory or mood can settle a worry that should have been said out loud to a doctor or a relative. Notice whether you are treating a chatbot as a first listener for something that frightens you, and bring that same worry to a person who knows your family.

Also keep a record. Write down or photograph what a chatbot told you and when, especially about memory, mood or medicines. Weeks later the advice is easy to remember as clearer and more reassuring than it was, and the decision then rests on a softened memory rather than the actual words.

The authors are direct about the boundary. General-purpose 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 adults26. Where appropriate, they add, a system should explicitly advise involving a family member, caregiver, primary-care clinician, emergency service or crisis team27.

This reading would be wrong if families who use these tools routinely check the answers with a clinician or a relative and readily overrule advice that feels off. Then the fluency would cost little. Until that is shown, the practical move is small and concrete: use the chatbot to draft the questions, and let a person answer them.

The daughter with the phone can still get help from the screen. She can ask it to list what to mention at the appointment, and then she can make the call herself. Before trusting any answer about mood, memory or medicine, ask one question of it and of yourself: who is going to examine my father?

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

  1. 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 )."
  2. 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."
  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 - the article this story is about — the whole article — the passage: "We conducted a blinded, paired benchmarking study using 90 questions covering six geriatric psychiatry domains and equally distributed across low-, moderate-, and high-risk strata."
  4. 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)."
  5. 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."
  6. 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)."
  7. 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."
  8. 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."
  9. 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: "Test-retest clinical consistency was highest for ChatGPT (90.0%), followed by Gemini (83.3%) and Doubao (73.3%)."
  10. 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: "First, 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."
  11. 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 )."
  12. 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: "Existing evaluations of large language models in mental health have generally emphasized factual accuracy, response quality, empathy, or agreement with professional recommendations ( 30 ). These outcomes are informative but do not fully capture the demands of late-life depression care."
  13. Yun B, Kim B, Ko H, Kim J, Kim BR, Kim W, et al. (2026). Enhancing Cognitive Functions of Older Adults With Software Robot: Longitudinal Exploratory Field Study. JMIR mHealth and uHealth. 10.2196/75308 — only the abstract - the full text could not be fetched — the passage: "We designed an intervention that integrates social interaction, reward-based engagement, and an artificial intelligence (AI) conversational agent, which aims to reduce these barriers and support continuous participation."
  14. Yun B, Kim B, Ko H, Kim J, Kim BR, Kim W, et al. (2026). Enhancing Cognitive Functions of Older Adults With Software Robot: Longitudinal Exploratory Field Study. JMIR mHealth and uHealth. 10.2196/75308 — only the abstract - the full text could not be fetched — the passage: "We recruited 133 participants (mean age 64.75, SD 6.76; range 55-75 years) who had no diagnosis of dementia."
  15. Yun B, Kim B, Ko H, Kim J, Kim BR, Kim W, et al. (2026). Enhancing Cognitive Functions of Older Adults With Software Robot: Longitudinal Exploratory Field Study. JMIR mHealth and uHealth. 10.2196/75308 — only the abstract - the full text could not be fetched — the passage: "Participants showed improved cognitive function after the intervention (Hedges g=0.350, P<.001)."
  16. Yun B, Kim B, Ko H, Kim J, Kim BR, Kim W, et al. (2026). Enhancing Cognitive Functions of Older Adults With Software Robot: Longitudinal Exploratory Field Study. JMIR mHealth and uHealth. 10.2196/75308 — only the abstract - the full text could not be fetched — the passage: "Active users (n=67), who engaged more frequently with the program, showed greater improvement than nonactive users (n=66), especially among those who had lower baseline cognitive scores (Hedges g=0.523, P<.001)."
  17. Yun B, Kim B, Ko H, Kim J, Kim BR, Kim W, et al. (2026). Enhancing Cognitive Functions of Older Adults With Software Robot: Longitudinal Exploratory Field Study. JMIR mHealth and uHealth. 10.2196/75308 — only the abstract - the full text could not be fetched — the passage: "Social support increased, particularly emotional and informational support (t132=-6.509, P<.001)."
  18. Yun B, Kim B, Ko H, Kim J, Kim BR, Kim W, et al. (2026). Enhancing Cognitive Functions of Older Adults With Software Robot: Longitudinal Exploratory Field Study. JMIR mHealth and uHealth. 10.2196/75308 — only the abstract - the full text could not be fetched — the passage: "Participants reported higher engagement (t132=2.008, P<.05) and lower depression scores (t132=3.093, P<.01)."
  19. Yun B, Kim B, Ko H, Kim J, Kim BR, Kim W, et al. (2026). Enhancing Cognitive Functions of Older Adults With Software Robot: Longitudinal Exploratory Field Study. JMIR mHealth and uHealth. 10.2196/75308 — only the abstract - the full text could not be fetched — the passage: "Cognitive training alone did not produce a significant effect."
  20. Ransinyo K, Banharak S, Panpanit L, Thiengtham S, Visetsittikul P, Sommana C, et al. (2026). Effectiveness of a Chatbot-Based Hypertension Management Program on Selected Outcomes Among Community-Dwelling Older Adults: A Quasi-Experimental Study. Journal of Multidisciplinary Healthcare. 10.2147/jmdh.s635820 — only the abstract - the full text could not be fetched — the passage: "This study evaluated the effectiveness of a chatbot-based hypertension management program on knowledge, health behaviors, medication adherence, blood pressure control, and selected biomarkers among community-dwelling older adults with hypertension."
  21. Ransinyo K, Banharak S, Panpanit L, Thiengtham S, Visetsittikul P, Sommana C, et al. (2026). Effectiveness of a Chatbot-Based Hypertension Management Program on Selected Outcomes Among Community-Dwelling Older Adults: A Quasi-Experimental Study. Journal of Multidisciplinary Healthcare. 10.2147/jmdh.s635820 — only the abstract - the full text could not be fetched — the passage: "The experimental group received a 12-week chatbot-based program combining education, behavioral monitoring, and lifestyle counseling. The control group received usual care using a hypertension education manual."
  22. Ransinyo K, Banharak S, Panpanit L, Thiengtham S, Visetsittikul P, Sommana C, et al. (2026). Effectiveness of a Chatbot-Based Hypertension Management Program on Selected Outcomes Among Community-Dwelling Older Adults: A Quasi-Experimental Study. Journal of Multidisciplinary Healthcare. 10.2147/jmdh.s635820 — only the abstract - the full text could not be fetched — the passage: "The experimental group showed improvements in hypertension-related knowledge, health behaviors, medication adherence, and blood pressure control. The control group showed no improvement in systolic and a rise in diastolic blood pressure."
  23. Ransinyo K, Banharak S, Panpanit L, Thiengtham S, Visetsittikul P, Sommana C, et al. (2026). Effectiveness of a Chatbot-Based Hypertension Management Program on Selected Outcomes Among Community-Dwelling Older Adults: A Quasi-Experimental Study. Journal of Multidisciplinary Healthcare. 10.2147/jmdh.s635820 — only the abstract - the full text could not be fetched — the passage: "A chatbot-supported hypertension management program was associated with better self-management and blood pressure control among community-dwelling older adults, although the non-randomized design and recruitment from different settings mean that the findings are preliminary."
  24. 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. 10.1080/08037051.2026.2715847 — only the abstract - the full paper is behind a subscription — the passage: "Two hundred hypertensive patients were randomly assigned to a usual care group (routine health management) or an intervention group (AI-assisted health management, including aspects such as intelligent monitoring, early warning and proactive intervention, among others)"
  25. 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. 10.1080/08037051.2026.2715847 — only the abstract - the full paper is behind a subscription — the passage: "Following one year intervention, the intervention group demonstrated a higher blood pressure control rate than the usual care group (88.66% vs. 64.89%), along with lower SBP (128.49 ± 6.63 vs. 135.38 ± 7.31 mmHg) and DBP (77.49 ± 6.07 vs. 81.84 ± 5.44 mmHg) (all"
  26. 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."
  27. 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: "Where appropriate, the system should explicitly advise involvement of a family member, caregiver, primary-care clinician, emergency service, or crisis team."

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

Who paid: The authors declared that financial support was not received for this work and/or its publication, and they declared no commercial or financial relationships that could be construed as a potential conflict of interest.

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.

From the house

About

Every week the new research on AI is read here, and this page tells you what was found, how far it can be trusted, and what you can do with it. This page is free, and every line on it says what was found.

The studies are read by a machine built and ruled by a professor of communication at Sacramento State and Universidad Nacional Autónoma de México (UNAM). Who makes this →

Diego Bonilla, Ph.D.

Full Professor of Communication Studies, Sacramento State, California, United States
Graduate instructor, Universidad Nacional Autónoma de México (UNAM), Mexico

hypergraphia.com

I am a professor of communication at Sacramento State, where I created the course Information Management and Privacy and have taught it many times, and at UNAM, in Mexico City, I teach the graduate course Electronic and Generative Literature with AI. In both I teach AI. I have taught since 1989, at first mathematics and physics to children who found them hard, and since 1997 one subject that has changed its name many times: media literacy, computer literacy, digital literacy, information literacy, ICT literacy, and now AI literacy.

This page exists because most people cannot go back to a university, and these systems are changing their work, their children's schooling, their parents' care and their vote all the same. What a university teaches about them should reach the people who need it, every week, in plain language, without enrolling anywhere. That is what I do here, and it is what I do on Thursday mornings in a seminar for the Renaissance Society at Sacramento State, built on this publication, where anyone can send a question and hear it answered.

In 2003 my doctorate at Syracuse measured what a machine can infer about a person from the traces they leave, and what I found became the privacy course I still teach. In 2007 and 2009 I sat on the committee at ETS in Princeton that set the national standard for digital literacy, and from 2014 to 2016 I was the California State University's ICT Literacy Ambassador. Six names, one job: helping people not be at the mercy of the newest machine.

I have also worked on the other side of the counter. In 2000 I began making films that a machine assembles differently for every viewer; the first of them took first prize at the Moscow International Film Festival in 2004, two decades before the industry shipped a generative model. A generative video poem I made with a poet at UNAM is in the field's definitive anthology, and in 2025 my AI short film The Human Futures Market took Best AI Short Film at the Red Rocks AI Film Festival. My recent published research is on AI and elections, and on how these models represent the cultures they were not trained on. I do not read the papers myself, and this page will never tell you I do. I built the machine that reads them and wrote the rules it reads by, and I analyze what it gives me the way I have taught my students to analyze: asking how it knows, and who benefits from what it says.

The vendors began teaching AI literacy two years ago, because before that there was nothing to sell. I am not against them. I am simply not one of them, and no one who sells these systems pays for this page. Your subscription does, and that is what keeps it honest.

The record, dated
  • 1989 Teaching begins, Mexico City
  • 1995 Hypergraphia, an early Spanish-language web and hypermedia publisher
  • 1996 Mercadotecnia e imagen en Internet, with Jesús del Toro, a book on marketing on the Internet, among the first in Spanish, Editorial Iberoamérica
  • 1999 Autorretrato, first generative poem, published by UNAM's Instituto de Investigaciones Filológicas
  • 2003 Ph.D., Syracuse University, Doctoral Prize, for the dissertation The Medium Is the Measure of Itself
  • 2004 A Space of Time: First Prize, Internet/Multimedia, XXVI Moscow International Film Festival
  • 2007, 2009 ETS iSkills committee, Princeton
  • 2008 Making Sense of Tracking Data, a book on how the traces a person leaves online allow inferences about them, including how they think, VDM Verlag
  • 2011 Keynote on generative work, UNAM
  • 2014–2016 ICT Literacy Ambassador, California State University · California Open Educational Resources Council
  • 2019 Big Data, with Rodolfo Mata: Electronic Literature Collection, Volume 4
  • 2020 Tablada Hipertextual, with Rodolfo Mata, UNAM: the computer reads a poet's words and builds a different reading from them, the ground the language models now work on
  • 2021 Uku Pacha, a generative VR film: world premiere, Vancouver International Film Festival
  • 2023 Keynote on the representation of Latin American cultures in AI, Ibero-American Meeting of Audiovisual Narratives, Trujillo, Peru · Graduate teaching at UNAM begins
  • 2023–2026 Directed the faculty learning community Advancing AI Literacy, Sacramento State · Principles for Cognitive and Pedagogical Sovereignty
  • 2024 A weekly AI publication run on real pipelines, every week since
  • 2025 Best AI Short Film, Red Rocks AI Film Festival
  • 2026 Keynote, with Nancy Huante-Tzintzun, Latin American and Caribbean Digital Humanities symposium, UT San Antonio · AI Watch, a weekly seminar for the Renaissance Society at Sacramento State

The research record →

TertulIA: weeklyAI.watch/tertulia/en/Pictures are illustrations generated by weeklyAI.watch, never photographs