interviews · JMIR formative research · la publicación, 3 nov 2026 · gratis
Un programa lee tu texto y ordena tus síntomas de depresión; los terapeutas que lo probaron confiaron en él para casi todo, menos para el riesgo de suicidio
Seis terapeutas en la primera ronda y ocho en la segunda compararon su lectura de publicaciones de Reddit con la del programa. Es un estudio de talleres con pocos profesionales, hecho en Reino Unido y sobre textos en inglés, y no mide si alguien mejora.
Versión breve · la versión detallada sigue, unos 8 min
Pregunte al chat de weeklyAI
Pregúnteme sobre la IA en su vida: le señalo las notas que weeklyAI ha publicado sobre el tema, esta semana y antes.
Las conversaciones se guardan mientras exista weeklyAI, para mejorar la publicación. Se responde en el idioma en que usted escribe.
- El estudio, de un vistazo
- Quiénes
- Terapeutas que probaron un programa que clasifica síntomas de depresión
- Cuántos
- 6 en la primera ronda y 8 en la segunda
- Dónde
- Reino Unido
- Cuándo
- febrero de 2025 y agosto de 2025
- Tipo de estudio
- interviews
- Quién lo hizo
- Universidad de Lancaster, Reino Unido
- El límite que importa
- Solo 6 y luego 8 terapeutas, y todos los textos vienen de Reddit
Lo que el programa clasificaba como estilo de vida y lo que los terapeutas leían como señal de depresión
Eso llevó a ampliar de 5 a 10 categorías en la segunda versión
Si le quitas el riesgo, creo que es bastante confiable
"Si le quitas el riesgo, creo que es bastante confiable", dijo una de las terapeutas sobre la segunda versión. Lo demás del programa le parecía claro.
Investigadores de la Universidad de Lancaster, en Reino Unido, construyeron DepressionSymp, un programa basado en el modelo de lenguaje GPT. Recibe un texto escrito por una persona y lo separa en categorías de síntomas de depresión. Lo mostraron a terapeutas en dos rondas de talleres en línea.
En la primera ronda participaron seis profesionales, de 25 a 35 años y con 2 a 10 años de experiencia. En la segunda, ocho, de 27 a 47 años y con 3 a 13 años de experiencia. Cada uno leyó cinco publicaciones tomadas de Reddit y anotó los síntomas a mano. Después comparó su lectura con la del programa.
La segunda versión, con diez categorías en lugar de cinco, coincidió con el juicio de los terapeutas entre el 80% y el 95% de las veces, según el cálculo de ellos mismos. Valoraron los resúmenes en viñetas porque ahorran tiempo y ayudan a planear la intervención. Los autores reportaron puntajes de coincidencia de texto altos para las dos versiones. Esos puntajes miden palabras compartidas y no pueden detectar bien un cambio de sentido, como una negación.
El programa lee sólo palabras. No oye el tono ni la velocidad al hablar, no ve la apariencia ni el lenguaje del cuerpo, y no capta lo que la persona calla. Los terapeutas también advirtieron que ordenar casos de forma automática, sin una persona que revise, podría mandar a un cliente de alto riesgo con alguien sin la preparación adecuada.
Los mismos autores dicen que falta una validación independiente con más participantes y una evaluación formal de seguridad. Nada de esto muestra cómo trataría una ley, un regulador o un tribunal de su país una herramienta así, ni qué podría usted exigir si alguien la usara con sus palabras.
Si algún servicio cerca de usted empieza a pasar por un programa como éste lo que usted escribe sobre su ánimo, la pregunta que vale la pena hacer es quién lo revisa antes de que influya en su atención, y si usted puede ver y corregir lo que el programa dice de usted.
Qué significa para usted
Lo que este estudio muestra es que unos terapeutas del Reino Unido encontraron claro el resumen de síntomas, y nada más. Si en su consulta o en un servicio cerca de usted empiezan a usar un programa así con lo que usted escribe, pida ver ese resumen antes de que influya en su atención y pregunte quién lo revisa. Usted puede corregir lo que el programa dice de usted.
Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422
Quién pagó: Los autores declararon que no recibieron apoyo financiero para este trabajo, y el artículo no menciona ningún financiador, proveedor de equipos ni de software con influencia alguna.
No tome esto como consejo médico profesional.
Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1640 palabras · unos 8 minLeerla →Cerrar
Una herramienta de IA lee lo que usted escribe sobre su ánimo. Los terapeutas que la probaron pidieron que una persona lo revise antes de usarlo
Seis profesionales de la salud mental compararon sus propias lecturas con las de un modelo de lenguaje, y después otros ocho revisaron una versión revisada. El punto más delicado, dijeron, es el riesgo: ahí la decisión debe seguir siendo humana.
Imagine la escena: una persona escribe en un foro que ya no sale de casa, que no le alcanza el dinero, que su vida cambió de golpe. Un programa lee ese texto y devuelve una lista ordenada: síntomas, riesgo, tiempo, estilo de vida, otras condiciones médicas. Eso es DepressionSymp 1.0, un modelo construido sobre GPT-3.5 y afinado con textos públicos de un foro de consejería, que clasificaba en cinco categorías tomadas de los criterios del manual diagnóstico DSM-5 y de un diccionario psicolingüístico llamado LIWC-221. Seis terapeutas lo revisaron en talleres en línea en febrero de 2025. Después vino la versión 2.0, con diez categorías, revisada por otros ocho profesionales en agosto de 20252.
Lo que encontraron los autores, de la Universidad de Lancaster, en el Reino Unido, fue que los profesionales valoraron esas listas estructuradas porque les bajaban la carga mental y les ayudaban a planear la intervención3. Pero hubo un desacuerdo revelador en la primera ronda: cosas como el aislamiento social, las dificultades económicas y los cambios grandes de vida el modelo las ponía en "estilo de vida", y los terapeutas las leían como señales indirectas de depresión. Eso llevó a ampliar las categorías a diez, que la segunda ronda describió como más claras y más útiles para la evaluación previa4.
Sobre el riesgo, los dos grupos coincidieron en algo: es la parte más delicada de todo el asunto, junto con la falta de información no verbal en un texto escrito, las diferencias culturales y de idioma al expresar el malestar, y el peligro de que la gente confíe en el sistema sin supervisión humana suficiente5. En la segunda ronda, varios profesionales calcularon que la coincidencia entre su lectura y la del modelo andaba entre el 80% y el 95%6. Algunos notaron que el sistema detectó síntomas que ellos habían pasado por alto, y al revisarlos estuvieron de acuerdo en que eran válidos7. Los terapeutas también valoraron que el sistema se pegue a las palabras del cliente en vez de suponer cosas8, y vieron un uso práctico: un resumen automático antes de la primera sesión acortaría la entrevista inicial y dejaría más tiempo para el trabajo terapéutico9.
Los límites importan tanto como los hallazgos. Fueron seis profesionales y luego ocho, y los propios autores dicen que eso limita cuánto se puede generalizar10. Todos los textos venían de Reddit, con un lenguaje informal que puede no parecerse al de una consulta, y no se sabe si quienes escribieron tenían el inglés como lengua materna11. La anotación de los 939 mensajes que quedaron tras descartar los irrelevantes la hizo la investigadora principal, no clínicos titulados. Y las medidas de coincidencia que usaron comparan palabras, no significados: un texto que diga lo contrario de lo que el paciente quiso decir podría puntuar alto. Los autores piden validación independiente a mayor escala y una evaluación formal de seguridad12.
Piense en lo que hay hoy en el mercado de la salud mental digital. Según el resumen de una revisión de aplicaciones de inteligencia artificial para adolescentes, lo único disponible de ese trabajo porque el texto completo está detrás de una suscripción, de 95 estudios revisados, 83 reportaban algún desempeño predictivo o diagnóstico, pero solo uno medía efectividad a nivel de síntomas13. Ningún estudio primario ofrecía una evaluación prospectiva dedicada a la seguridad de la IA14, y ninguno evaluó el costo de llevarla a la práctica rutinaria15. El grueso de esa investigación se quedó en el desarrollo técnico16. Esa misma revisión propone evaluar estas herramientas como sistemas supervisados, con una persona en el circuito y garantías explícitas de consentimiento, privacidad y equidad17.
Hay otra manera de acercarse, que no pasa por leer el texto del paciente sino por conversar con él. Según el resumen de un estudio hecho en Hong Kong, lo único disponible de él, 511 personas atendidas en una consulta psiquiátrica respondieron sobre sus necesidades, y de ahí salió un prototipo de agente conversacional en cantonés, pensado como apoyo adicional al tratamiento habitual1819. Dos tercios de esos pacientes, el 66.2%, ya usaban chatbots genéricos para hablar de lo que sentían20. Lo que más pedían era empatía local, un trato sin juicios y disponibilidad a toda hora21. Y en Nigeria, según el resumen de otro estudio, lo único disponible, 37 adultos que habían usado chatbots de salud mental durante al menos cuatro semanas describieron algo parecido a una ambivalencia: valoraban la disponibilidad permanente, pero extrañaban reciprocidad humana genuina2223.
Lo curioso es que ese estudio nigeriano encontró algo que toca de cerca a cualquiera que viva donde la atención escasea: los participantes adaptaban marcos terapéuticos occidentales a sus propias formas de nombrar el sufrimiento, mezclando la conversación con la IA con el consejo familiar y las prácticas espirituales24. Y el uso prolongado generaba una dependencia frágil, con ansiedad por si la aplicación desaparecía o por una filtración de datos, agravada porque no había un camino claro hacia un humano25. Los autores de ese trabajo concluyen que, en países de ingreso medio y bajo, hace falta diseño culturalmente adaptado, límites claros sobre qué hace el algoritmo y protocolos obligatorios de paso a una persona26.
Así lo lee weeklyAI: cuando una institución pone una máquina a clasificar en lugar de una persona, la máquina suele presentarse como limpia, pareja y sin cansancio, y así se le delegan también las decisiones delicadas que antes exigían criterio. En un servicio de salud mental que adopte algo así, es probable que lo presenten como más objetivo y menos cargado de prejuicios que su personal, y que la revisión humana se vuelva un trámite rápido, sobre todo con presión de tiempo o de costo. Esta lectura se equivocaría si los protocolos del servicio exigen que una persona revise y pueda corregir cada resumen antes de que influya en la atención, con tiempo pagado para hacerlo y registro de quién revisó qué.
Hay un segundo punto, y es el que más le toca al paciente. Quien recibe una decisión tomada por un sistema opaco rara vez puede entender por qué se tomó ni discutirla. Aunque los terapeutas del estudio digan que entienden las categorías, es probable que la persona cuyo texto fue leído no reciba ninguna explicación de por qué se le clasificó de cierta manera, ni sepa a quién reclamar. Eso cambiaría si el servicio le entrega a esa persona una explicación comprensible de qué se detectó en su texto, con qué palabras, y una vía concreta para corregirlo ante alguien identificable. Los propios autores proponen algo en esa dirección: que el cliente pueda revisar, aceptar o disputar el resumen antes de que alimente una decisión27, y que el consentimiento aclare que el sistema solo lee lo escrito, que algunas señales se infieren de frases indirectas y que todo pasa por un profesional28.
Y una tercera cosa que weeklyAI espera ver, sin que este estudio lo haya probado: un sistema que solo lee texto escrito probablemente capte mejor a quien escribe largo, ordenado y en el idioma dominante, y pase por alto a quien se expresa con rodeos, con humor, en otra lengua o con pocas palabras. Los terapeutas del estudio lo señalaron: las palabras cargan implicaciones distintas según el origen de cada quien, y vieron como responsabilidad propia seguir atentos a esos matices29. La lectura se equivocaría si el servicio prueba la herramienta con gente de distintos idiomas, niveles de escritura y formas de hablar, y muestra que la clasificación se sostiene igual.
¿Y qué haría falta para que algo así llegue a una consulta? Los autores lo tienen claro: un mecanismo con el profesional en el circuito, que revise y corrija las evaluaciones preliminares, marque los casos de alto riesgo y dé retroalimentación que mejore el sistema con el tiempo30. También señalan que la confiabilidad del modelo choca con la honestidad del paciente: la gente puede omitir o minimizar lo que siente, y el sistema tendría que anticipar esa ambigüedad en vez de suponer que todo se contó31. Y agregan algo que suena a sentido común pero que pocas veces se escribe: los datos de un paciente así de sensibles necesitan anonimizarse, acceso restringido y registro de quién los consultó32.
La pregunta que queda para quien vive en México, en Colombia o en cualquier ciudad de este continente es concreta, y no la responde este estudio. Si su clínica, su universidad o su servicio de consejería usa una herramienta así para leer lo que usted escribe sobre su ánimo, usted puede preguntar quién revisa ese resumen y en cuánto tiempo, y pedir que quede asentado en su expediente que fue una persona quien tomó la decisión. Puede pedir también que le muestren qué partes de su texto se usaron para clasificarlo, y que su propia versión quede junto al resumen automático. Y si escribe poco o en otra lengua, decirlo en voz alta en la consulta y pedir que lo que no quedó escrito se anote también. Nada de esto sustituye la atención: es lo que se puede exigir alrededor de ella.
Vuelva a la persona que escribió en el foro que ya no sale de casa. Lo que el sistema devolvió fue una lista. Lo que los terapeutas pidieron fue que, antes de que esa lista pese en la vida de alguien, haya una persona que la lea, la entienda y responda por ella. La próxima vez que firme un consentimiento para que un programa lea lo que escribe, pregunte: ¿quién va a revisar esto, y con cuánto tiempo para hacerlo?
De dónde sale cada dato de contexto, y cuánto leímos de cada documento
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "In stage 1, we developed DepressionSymp 1.0, a fine-tuned GPT-3.5 model trained on the CounselChat dataset, identifying 5 symptom categories based on the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition ( DSM-5 ) criteria and the Linguistic Inquiry and Word Count (LIWC-22) psycholinguistic framework"
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "In stage 2, this feedback informed DepressionSymp 2.0, expanded to 10 symptom categories and reviewed by 8 further practitioners in a second round of workshops."
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Practitioners valued the structured symptom outputs for reducing cognitive workload and supporting intervention planning."
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "In stage 1, indirect indicators such as social withdrawal, financial difficulties, and major life changes were often classified as lifestyle factors rather than symptoms; this informed the expansion to 10 categories in stage 2, described as clearer and more useful for preassessment."
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Risk categorization was identified as the most ethically sensitive aspect across both stages, alongside broader concerns including the lack of nonverbal information in text-based assessments, cultural and linguistic differences in symptom expression, and the potential for automation bias without sufficient human oversight."
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "DepressionSymp 2.0 achieved strong alignment with practitioners’ assessments, with many (P1-P8) estimating similarity rates around 80%‐95%"
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Several practitioners (P2, P3, and P4) also noted that DepressionSymp identified symptoms they had initially missed, and agreed on review that these were valid"
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Practitioners appreciated that the system categorized inputs directly from the words provided by clients rather than making assumptions (P7 and P5), viewing this as a safeguard against bias that kept outputs close to clients’ lived experiences"
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "AI-generated summaries before meeting a client would shorten initial sessions, support note-taking and early formulation, and reduce repeated questioning, freeing more time for therapeutic engagement"
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The study also involved a small number of mental health practitioners (6 in workshop 1 and 8 in workshop 2) due to practical constraints, limiting generalizability"
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "all posts come from Reddit, which may introduce bias in language style, such as informal expressions and vocabulary differing from clinical or other platform norms, and it is unclear whether writers are native English speakers"
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Future work should prioritize larger-scale independent validation, formal safety evaluation, and culturally diverse dataset development."
- Dhamayanti M, Susaldi S, Hanifa F, Agustini Y, Emilia O, Hardinsyah H, et al. (2026). Artificial intelligence-enabled mental health applications for adolescents: a scoping review of safety, effectiveness and clinical implementation. BMJ Open. 10.1136/bmjopen-2026-125586 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "Predictive or diagnostic performance was reported in 83 studies (87.4%), but symptom-level effectiveness in only 1 (1.1%)."
- Dhamayanti M, Susaldi S, Hanifa F, Agustini Y, Emilia O, Hardinsyah H, et al. (2026). Artificial intelligence-enabled mental health applications for adolescents: a scoping review of safety, effectiveness and clinical implementation. BMJ Open. 10.1136/bmjopen-2026-125586 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "No primary study provided a dedicated prospective AI-safety evaluation."
- Dhamayanti M, Susaldi S, Hanifa F, Agustini Y, Emilia O, Hardinsyah H, et al. (2026). Artificial intelligence-enabled mental health applications for adolescents: a scoping review of safety, effectiveness and clinical implementation. BMJ Open. 10.1136/bmjopen-2026-125586 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "No study evaluated sustained routine-care scale-up or implementation cost as a principal outcome."
- Dhamayanti M, Susaldi S, Hanifa F, Agustini Y, Emilia O, Hardinsyah H, et al. (2026). Artificial intelligence-enabled mental health applications for adolescents: a scoping review of safety, effectiveness and clinical implementation. BMJ Open. 10.1136/bmjopen-2026-125586 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "Within the retrieved evidence base, adolescent mental-health AI research was concentrated in technical development, with limited prospective evidence of clinical benefit, safety and routine implementation."
- Dhamayanti M, Susaldi S, Hanifa F, Agustini Y, Emilia O, Hardinsyah H, et al. (2026). Artificial intelligence-enabled mental health applications for adolescents: a scoping review of safety, effectiveness and clinical implementation. BMJ Open. 10.1136/bmjopen-2026-125586 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "Future applications should be evaluated as supervised, human-in-the-loop systems with explicit safeguards for consent, privacy, fairness"
- Lo HKY, Yu MSY, Sze PKS, Wong RWK, Ku MCT, Cheng YYH, et al. (2026). From Needs Assessment to Pre-Trial Specification: Participatory Co-Design of a Generative AI Chatbot (CARES-MDD) for Persons Living with Depression in Hong Kong Psychiatric Care. Healthcare. 10.3390/healthcare14203547 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Phase 1 comprised a needs assessment of 511 psychiatric outpatient attendees in Hong Kong."
- Lo HKY, Yu MSY, Sze PKS, Wong RWK, Ku MCT, Cheng YYH, et al. (2026). From Needs Assessment to Pre-Trial Specification: Participatory Co-Design of a Generative AI Chatbot (CARES-MDD) for Persons Living with Depression in Hong Kong Psychiatric Care. Healthcare. 10.3390/healthcare14203547 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "This process informed the development of CARES-MDD, a Hong Kong Cantonese, CBT-informed generative AI conversational-agent prototype intended to provide adjunctive support alongside usual psychiatric care."
- Lo HKY, Yu MSY, Sze PKS, Wong RWK, Ku MCT, Cheng YYH, et al. (2026). From Needs Assessment to Pre-Trial Specification: Participatory Co-Design of a Generative AI Chatbot (CARES-MDD) for Persons Living with Depression in Hong Kong Psychiatric Care. Healthcare. 10.3390/healthcare14203547 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "In Phase 1, 66.2% reported using generic AI chatbots for emotional concerns; privacy, anticipated negative evaluation, and concern about burdening others were recurring themes in the qualitative data."
- Lo HKY, Yu MSY, Sze PKS, Wong RWK, Ku MCT, Cheng YYH, et al. (2026). From Needs Assessment to Pre-Trial Specification: Participatory Co-Design of a Generative AI Chatbot (CARES-MDD) for Persons Living with Depression in Hong Kong Psychiatric Care. Healthcare. 10.3390/healthcare14203547 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Localized empathy, non-judgmental, and 24/7 accessibility were the main themes of a desired mental health generative AI chatbot."
- Isaiah US, Adili-George CF, Naibo AT, George I, Useh EM, Effiong JE. (2026). Lived experiences of relationships to generative AI mental health chatbots among adults. Discover Psychology. https://doi.org/10.1007/s44202-026-01001-1 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Interpretative Phenomenological Analysis with 37 adults (18–45 years) who engaged AI mental health chatbots for ≥ 4 weeks."
- Isaiah US, Adili-George CF, Naibo AT, George I, Useh EM, Effiong JE. (2026). Lived experiences of relationships to generative AI mental health chatbots among adults. Discover Psychology. https://doi.org/10.1007/s44202-026-01001-1 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Relational Ambivalence ; participants valued 24/7 availability and non-judgmental space yet struggled with perceived lack of genuine human reciprocity"
- Isaiah US, Adili-George CF, Naibo AT, George I, Useh EM, Effiong JE. (2026). Lived experiences of relationships to generative AI mental health chatbots among adults. Discover Psychology. https://doi.org/10.1007/s44202-026-01001-1 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Cultural Navigation ; users actively adapted Western therapeutic frameworks to local idioms of distress, blending AI interactions with family counsel and spiritual practices"
- Isaiah US, Adili-George CF, Naibo AT, George I, Useh EM, Effiong JE. (2026). Lived experiences of relationships to generative AI mental health chatbots among adults. Discover Psychology. https://doi.org/10.1007/s44202-026-01001-1 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Fragile Dependency ; prolonged use generated attachment anxiety about potential app discontinuation or data breaches, compounded by absence of human escalation pathways."
- Isaiah US, Adili-George CF, Naibo AT, George I, Useh EM, Effiong JE. (2026). Lived experiences of relationships to generative AI mental health chatbots among adults. Discover Psychology. https://doi.org/10.1007/s44202-026-01001-1 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Findings suggest ethical deployment in LMICs requires culturally adaptive design, transparent algorithmic boundaries, and mandatory human handoff protocols to prevent dependency in unsupported contexts."
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "client review and correction loops, where clients can review, agree with, or dispute symptom summaries before automated outputs inform decisions, may help address this (P6)"
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Based on our findings, consent procedures [ 127 ] should clarify that the system identifies only the client’s written language, that some symptom cues may be interpreted from indirect or implied expressions, and that all outputs are reviewed by practitioners before being used in decision-making"
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Language and culture also interact (P6, P7, and P8), with words carrying different implications depending on background; while the model can categorize information consistently, practitioners saw it as their responsibility to remain sensitive to these nuances for each client."
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "We recommend embedding a human-in-the-loop mechanism in DepressionSymp [ 120 ], where practitioners review and correct preliminary LLM-generated assessments, flag high-risk cases, and provide feedback that iteratively improves reliability [ 121 ]."
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Model reliability is also constrained by honesty, disclosure, and stigma, as clients may omit or minimize sensitive information; systems should therefore anticipate ambiguity rather than assume complete disclosure (P4)."
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Our findings highlight the need for strong data governance to ensure client privacy, security, and responsible use of sensitive information [ 128 ]: anonymizing client inputs, controlling access so only authorized practitioners can view outputs, and keeping access records can safeguard confidentiality"
Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422
Quién pagó: Los autores declararon que no recibieron apoyo financiero para este trabajo, y el artículo no menciona ningún financiador, proveedor de equipos ni de software con influencia alguna.
No tome esto como consejo médico profesional.
interviews · JMIR formative research · the paper, 3 Nov 2026 · free
A Tool Sorts Their Words by Symptom. The People Who Read People Say It Misses Things.
In two rounds of workshops with 14 mental health practitioners, a GPT-based system matched much of their reading of Reddit posts, but it could not hear a voice or see a face.
Short version · the longer version follows, about 8 min
Ask the weeklyAI chat
Ask me about AI in your life: I point you to the stories weeklyAI has published on it, this week and before.
Conversations are saved for as long as weeklyAI exists, to improve the publication. Answers come in the language you write in.
- The study at a glance
- Who
- mental health practitioners reviewing a language model's sorting of written posts
- How many
- 6 in workshop 1 and 8 in workshop 2
- Where
- Lancaster University, United Kingdom
- When
- February 2025 and August 2025
- Kind of study
- interviews
- Who did it
- Lancaster University, School of Computing and Communication
- The limit that matters
- Only 6 and then 8 practitioners; posts came from Reddit, not clinics or patients.
The two versions of the tool, as practitioners read them
1.0 sorted into 5 categories; 2.0, built on practitioner feedback, sorted into 10, described as clearer and more useful for preassessment
In the second round, many practitioners estimated the match at around 80%‐95%
Trust held for the symptoms. It stopped at risk.
"If you took risk, I think it's pretty trustworthy." One mental health practitioner was talking about what it would take to rely on DepressionSymp, a language model that reads a person's written post and sorts it into categories of depression symptoms. Trust held for the symptoms. It stopped at risk.
A lead researcher at Lancaster University, in the United Kingdom, labelled the posts and built the tool. Two co-authors helped with the method, the software, the analysis and the supervision. The tool went to practitioners in two rounds of workshops, first in February 2025, then in August 2025. Six practitioners joined the first round: two men and four women, aged 25 to 35. Eight joined the second: five women and three men, aged 27 to 47. Each person read five posts and wrote down the symptoms they saw, then compared their notes with what the model produced.
The posts came from a public Reddit dataset on depression, not from clinics or from patients. The first version of the system used five categories: symptoms, risk, time, lifestyle, and medical conditions. Practitioners read social withdrawal, money trouble and sudden life changes as indirect signs of depression, not as mere lifestyle. The second version, built with that feedback, used ten categories. Practitioners in that round estimated that its output matched their own judgment about 80% to 95% of the time. Tell that figure for what it is: their own estimate, from reading posts, not a count of correct calls on patients.
The model reads only words. Practitioners said it cannot hear tone or see how a person looks, and that someone can be "very well put together and still be actively suicidal." One warned that a fully automated sort could send a client at high risk to an untrained practitioner. The authors say larger independent checks and a formal safety review are work still to come.
Nothing in this study tells you what the law or a regulator in your country says about a service that runs your writing through such a tool, or what you could demand if it did.
So keep the question: if a clinic near you starts sorting what you write with a tool like this, who reads the output before it shapes your care, and can you see it and correct it?
What this means for you
When a service near you starts sorting what you write, the people in these workshops said the tool caught symptoms well but held back on risk, and that a person who looked fine could still be in danger. Ask who reads the output before it shapes your care, and whether you can see it and correct it.
Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422
Who paid: The authors declared that no financial support was received for this work, and the article names no funder, equipment lender or software provider with any say.
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 · 1522 words · about 8 minRead it →Close
# An AI Read Their Posts and Sorted Their Symptoms. The Therapists Wanted a Say in What Happens Next.
A co-design study with 14 mental health practitioners found the tool's summaries matched their own reading of client posts — and that risk was the part they trusted least.
Consider what a therapist receives before a first session: a client's written post, a few hundred words about mood, sleep, money, family. In this study, a language model read that post and returned a tidy list — in the first version, symptoms, risk, time, lifestyle and medical conditions — and a practitioner, in an online workshop, compared it with her own reading. That comparison is the whole of what was measured.
The system is called DepressionSymp, built at Lancaster University by Abeer Alotaibi, Corina Sas and Hossein Rahmani. Version 1.0 was a fine-tuned GPT-3.5 model trained on posts from CounselChat, a public Reddit dataset, and it sorted text into five categories drawn from DSM-5 depression criteria and a psycholinguistic word-counting framework called LIWC-221. Version 2.0, built after practitioner feedback, used few-shot prompting instead and sorted into ten categories2. No financial support was declared, and the authors declared no conflicts.
The workshops were small and deliberate. Six practitioners joined the first round in February 2025 — two men and four women, aged 25 to 35, with two to ten years of experience. Eight joined the second in August 2025 — five women and three men, aged 27 to 47, with three to 13 years. Each reviewed five posts, comparing their own manual identification of symptoms with the model's output. In the second round, practitioners estimated the model's similarity to their own reading at around 80 to 95 percent3.
What did they value? The structured output itself — a bulleted summary that reduced the work of scanning a long post and gave a starting point for planning an intervention4. They liked that the system used the client's own words rather than paraphrasing them, which they saw as a guard against bias5. Some found symptoms they had missed, and agreed on review that the model was right6. And they saw a use before the session even begins: a summary that shortens the first meeting and frees time for actual therapy7.
What did they not trust? Risk. Across both rounds, risk categorization was named the most ethically sensitive part of the system, alongside the plain fact that text carries no tone, no posture, no silence — and the worry that a busy clinic might lean on the output without checking it8. One practitioner put the safeguarding concern sharply: fully automated allocation could send a high-risk client to an undertrained therapist9.
The first round had already shown the model's blind spot. Social withdrawal, money trouble, a sudden life change — these were filed as "lifestyle" when practitioners read them as indirect signs of depression10. That gap drove the expansion to ten categories, which the second round of practitioners described as clearer and more useful for preassessment10. The authors are candid about the study's shape: six and then eight practitioners limits what can be generalized11, and the participants were mostly younger and digitally at ease, which limits it further. The posts came from Reddit, with informal language and no certainty that writers were native English speakers12. The lead researcher annotated the training data rather than qualified mental health practitioners. And the scoring method measures word overlap, which cannot reliably catch a negation — a reversed clinical meaning could still score well.
This is one study in a crowded field, and the crowd is mostly building, not testing. According to the summary of a review of 95 studies of AI-enabled mental health applications for adolescents — the only part of it available; the full paper is behind a subscription — structured-data machine learning was the dominant approach, at 42 of 95 studies, and risk prediction was the leading function, at 45 of 9513. Predictive or diagnostic performance was reported in 83 studies, but symptom-level effectiveness in just one14. No primary study provided a dedicated prospective safety evaluation15. The review's conclusion is blunt: the field is concentrated in technical development, with limited prospective evidence of clinical benefit, safety and routine implementation16.
Set beside that, the Lancaster work is unusual in one respect: it asked the people who would use the tool what they thought, twice, and changed the tool in between. According to the summary of a study from Hong Kong — again, only the abstract was available — a co-design process with 10 people with lived experience of major depression and eight clinical professionals informed a Cantonese-language conversational agent intended as add-on support alongside usual psychiatric care1718. That study's authors note that evidence remains limited on how needs assessment and co-design can inform AI chatbot trials19.
A third study, from Nigeria, looked at what happens when people use AI chatbots for emotional support in a system that serves less than 10 percent of those in need20. According to its abstract — the full text could not be fetched — 37 adults who had used such chatbots for at least four weeks described valuing the round-the-clock, non-judgmental space while struggling with the absence of genuine human reciprocity2122. Some described a "fragile dependency," an attachment anxiety about the app being discontinued or the data breached, sharpened by the lack of any route to a human23. The authors' recommendation is plain: culturally adaptive design, transparent limits, and mandatory handoff protocols to a person24.
The trade-off in the Lancaster design is visible once you name it. What is gained: a shorter, clearer summary that a therapist can scan, and categories detailed enough to point toward an intervention47. What is given up: the model reads only what is written, and practitioners were explicit that meaning lives between the words8. A client who jokes about severe trauma, or who never uses the word "lonely," can pass through the system intact and unread. The authors' own remedy is a person in the loop — a practitioner who reviews and corrects each preliminary assessment, flags high-risk cases and feeds the correction back25. They add that clients should be able to review, agree with or dispute a summary before it informs any decision26, and that consent should say plainly that the system reads only written language and that a practitioner reviews every output27.
Here is weeklyAI's reading. A person scored by a system they cannot see tends to trust the individual worker in front of them more — not because that worker is unbiased, but because a worker can be talked to, corrected and held to account in the room. Expect clients in such a service to ask who read the summary before it touched their care. When a clinic tells you it runs an AI tool over what you write, ask who reviews the output, and whether you can see the summary and dispute it. This reading would be wrong if clients turned out indifferent to human review, or preferred the machine's judgment when the two disagree.
A second pattern is worth watching for. A flag from a text-reading tool can change what happens next: more questions, closer watching, earlier referral. That added attention can then look like proof the flag was right, even when the first reading was wrong. Ask whether being flagged changes what happens to you, and whether anyone records that the flag came from a tool rather than from a clinician's own judgment.
And a third, quieter one. A practitioner handed a tidy bullet-point summary before a session may lean on it without tracing each item back to the client's own sentences — especially in a busy service. The tool's mistakes then travel into care unchallenged. Ask your practitioner which of your sentences produced each item, and ask that any item you say is wrong be noted.
The authors themselves point to what would have to happen before any of this reaches a clinic: larger independent validation, formal safety evaluation, and datasets that reflect more cultures and more languages28. Their own data governance proposal — anonymized inputs, access limited to authorized practitioners, records of who looked29 — is the kind of thing a regulator, not a research team, would have to enforce.
None of this measured whether a patient got better. No outcomes were tracked, no comparison was run against care without the tool, and no safety trial was conducted. What the study shows is narrower and still useful: fourteen practitioners, reading real posts, found a language model's structured summary close to their own reading and helpful as a starting point — and drew a firm line at risk, where they wanted a human to decide.
So the question for a reader is not whether the machine can spot depression in a paragraph. It is who is accountable when it gets the paragraph wrong. The next time a counselling service mentions an AI tool, the sentence to listen for is not "it's accurate" but "a named practitioner reviews it, and you can see what it said about you." That sentence is the one this study argues for — and the one no rule yet requires.
Where each piece of context comes from, and how much of it we read
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - the article this story is about — the whole article — the passage: "In stage 1, we developed DepressionSymp 1.0, a fine-tuned GPT-3.5 model trained on the CounselChat dataset, identifying 5 symptom categories based on the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition ( DSM-5 ) criteria and the Linguistic Inquiry and Word Count (LIWC-22) psycholinguistic framework"
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - the article this story is about — the whole article — the passage: "In stage 2, this feedback informed DepressionSymp 2.0, expanded to 10 symptom categories and reviewed by 8 further practitioners in a second round of workshops."
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - the article this story is about — the whole article — the passage: "DepressionSymp 2.0 achieved strong alignment with practitioners’ assessments, with many (P1-P8) estimating similarity rates around 80%‐95%"
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - the article this story is about — the whole article — the passage: "Practitioners valued the structured symptom outputs for reducing cognitive workload and supporting intervention planning."
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - the article this story is about — the whole article — the passage: "Practitioners appreciated that the system categorized inputs directly from the words provided by clients rather than making assumptions (P7 and P5), viewing this as a safeguard against bias that kept outputs close to clients’ lived experiences"
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - the article this story is about — the whole article — the passage: "Several practitioners (P2, P3, and P4) also noted that DepressionSymp identified symptoms they had initially missed, and agreed on review that these were valid"
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - the article this story is about — the whole article — the passage: "AI-generated summaries before meeting a client would shorten initial sessions, support note-taking and early formulation, and reduce repeated questioning, freeing more time for therapeutic engagement"
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - the article this story is about — the whole article — the passage: "Risk categorization was identified as the most ethically sensitive aspect across both stages, alongside broader concerns including the lack of nonverbal information in text-based assessments, cultural and linguistic differences in symptom expression, and the potential for automation bias without sufficient human oversight."
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - the article this story is about — the whole article — the passage: "Automation without human oversight was seen as a safeguarding risk, as fully automated allocation could route high-risk clients to undertrained practitioners (P4)."
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - the article this story is about — the whole article — the passage: "In stage 1, indirect indicators such as social withdrawal, financial difficulties, and major life changes were often classified as lifestyle factors rather than symptoms; this informed the expansion to 10 categories in stage 2, described as clearer and more useful for preassessment."
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - the article this story is about — the whole article — the passage: "The study also involved a small number of mental health practitioners (6 in workshop 1 and 8 in workshop 2) due to practical constraints, limiting generalizability"
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - the article this story is about — the whole article — the passage: "all posts come from Reddit, which may introduce bias in language style, such as informal expressions and vocabulary differing from clinical or other platform norms, and it is unclear whether writers are native English speakers"
- Dhamayanti M, Susaldi S, Hanifa F, Agustini Y, Emilia O, Hardinsyah H, et al. (2026). Artificial intelligence-enabled mental health applications for adolescents: a scoping review of safety, effectiveness and clinical implementation. BMJ Open. 10.1136/bmjopen-2026-125586 — only the abstract - the full paper is behind a subscription — the passage: "Structured-data machine learning was the dominant modality (42/95, 44.2%), and risk prediction or prognosis was the leading clinical function (45/95, 47.4%)."
- Dhamayanti M, Susaldi S, Hanifa F, Agustini Y, Emilia O, Hardinsyah H, et al. (2026). Artificial intelligence-enabled mental health applications for adolescents: a scoping review of safety, effectiveness and clinical implementation. BMJ Open. 10.1136/bmjopen-2026-125586 — only the abstract - the full paper is behind a subscription — the passage: "Predictive or diagnostic performance was reported in 83 studies (87.4%), but symptom-level effectiveness in only 1 (1.1%)."
- Dhamayanti M, Susaldi S, Hanifa F, Agustini Y, Emilia O, Hardinsyah H, et al. (2026). Artificial intelligence-enabled mental health applications for adolescents: a scoping review of safety, effectiveness and clinical implementation. BMJ Open. 10.1136/bmjopen-2026-125586 — only the abstract - the full paper is behind a subscription — the passage: "No primary study provided a dedicated prospective AI-safety evaluation."
- Dhamayanti M, Susaldi S, Hanifa F, Agustini Y, Emilia O, Hardinsyah H, et al. (2026). Artificial intelligence-enabled mental health applications for adolescents: a scoping review of safety, effectiveness and clinical implementation. BMJ Open. 10.1136/bmjopen-2026-125586 — only the abstract - the full paper is behind a subscription — the passage: "Within the retrieved evidence base, adolescent mental-health AI research was concentrated in technical development, with limited prospective evidence of clinical benefit, safety and routine implementation."
- Lo HKY, Yu MSY, Sze PKS, Wong RWK, Ku MCT, Cheng YYH, et al. (2026). From Needs Assessment to Pre-Trial Specification: Participatory Co-Design of a Generative AI Chatbot (CARES-MDD) for Persons Living with Depression in Hong Kong Psychiatric Care. Healthcare. 10.3390/healthcare14203547 — only the abstract - the full text could not be fetched — the passage: "Phase 2 involved a participatory co-design and test-run experiences with 10 lived-experience experts with MDD and eight multidisciplinary clinical professionals."
- Lo HKY, Yu MSY, Sze PKS, Wong RWK, Ku MCT, Cheng YYH, et al. (2026). From Needs Assessment to Pre-Trial Specification: Participatory Co-Design of a Generative AI Chatbot (CARES-MDD) for Persons Living with Depression in Hong Kong Psychiatric Care. Healthcare. 10.3390/healthcare14203547 — only the abstract - the full text could not be fetched — the passage: "This process informed the development of CARES-MDD, a Hong Kong Cantonese, CBT-informed generative AI conversational-agent prototype intended to provide adjunctive support alongside usual psychiatric care."
- Lo HKY, Yu MSY, Sze PKS, Wong RWK, Ku MCT, Cheng YYH, et al. (2026). From Needs Assessment to Pre-Trial Specification: Participatory Co-Design of a Generative AI Chatbot (CARES-MDD) for Persons Living with Depression in Hong Kong Psychiatric Care. Healthcare. 10.3390/healthcare14203547 — only the abstract - the full text could not be fetched — the passage: "Evidence remains limited on how needs assessment and co-participatory design can inform AI chatbot trials for MDD."
- Isaiah US, Adili-George CF, Naibo AT, George I, Useh EM, Effiong JE. (2026). Lived experiences of relationships to generative AI mental health chatbots among adults. Discover Psychology. https://doi.org/10.1007/s44202-026-01001-1 — only the abstract - the full text could not be fetched — the passage: "Nigeria’s mental health system serves less than 10% of those in need, with severe provider shortages and deep-seated stigma."
- Isaiah US, Adili-George CF, Naibo AT, George I, Useh EM, Effiong JE. (2026). Lived experiences of relationships to generative AI mental health chatbots among adults. Discover Psychology. https://doi.org/10.1007/s44202-026-01001-1 — only the abstract - the full text could not be fetched — the passage: "Interpretative Phenomenological Analysis with 37 adults (18–45 years) who engaged AI mental health chatbots for ≥ 4 weeks."
- Isaiah US, Adili-George CF, Naibo AT, George I, Useh EM, Effiong JE. (2026). Lived experiences of relationships to generative AI mental health chatbots among adults. Discover Psychology. https://doi.org/10.1007/s44202-026-01001-1 — only the abstract - the full text could not be fetched — the passage: "Relational Ambivalence ; participants valued 24/7 availability and non-judgmental space yet struggled with perceived lack of genuine human reciprocity"
- Isaiah US, Adili-George CF, Naibo AT, George I, Useh EM, Effiong JE. (2026). Lived experiences of relationships to generative AI mental health chatbots among adults. Discover Psychology. https://doi.org/10.1007/s44202-026-01001-1 — only the abstract - the full text could not be fetched — the passage: "Fragile Dependency ; prolonged use generated attachment anxiety about potential app discontinuation or data breaches, compounded by absence of human escalation pathways."
- Isaiah US, Adili-George CF, Naibo AT, George I, Useh EM, Effiong JE. (2026). Lived experiences of relationships to generative AI mental health chatbots among adults. Discover Psychology. https://doi.org/10.1007/s44202-026-01001-1 — only the abstract - the full text could not be fetched — the passage: "Findings suggest ethical deployment in LMICs requires culturally adaptive design, transparent algorithmic boundaries, and mandatory human handoff protocols to prevent dependency in unsupported contexts."
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - the article this story is about — the whole article — the passage: "We recommend embedding a human-in-the-loop mechanism in DepressionSymp [ 120 ], where practitioners review and correct preliminary LLM-generated assessments, flag high-risk cases, and provide feedback that iteratively improves reliability [ 121 ]."
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - the article this story is about — the whole article — the passage: "client review and correction loops, where clients can review, agree with, or dispute symptom summaries before automated outputs inform decisions, may help address this (P6)"
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - the article this story is about — the whole article — the passage: "Based on our findings, consent procedures [ 127 ] should clarify that the system identifies only the client’s written language, that some symptom cues may be interpreted from indirect or implied expressions, and that all outputs are reviewed by practitioners before being used in decision-making"
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - the article this story is about — the whole article — the passage: "Future work should prioritize larger-scale independent validation, formal safety evaluation, and culturally diverse dataset development."
- Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422 - the article this story is about — the whole article — the passage: "Our findings highlight the need for strong data governance to ensure client privacy, security, and responsible use of sensitive information [ 128 ]: anonymizing client inputs, controlling access so only authorized practitioners can view outputs, and keeping access records can safeguard confidentiality"
Alotaibi, A., Sas, C., Rahmani, H. (2026). A Large Language Model–Based System for Depression Symptom Identification: Iterative Qualitative Co-Design Studies. JMIR Formative Research. https://doi.org/10.2196/95422
Who paid: The authors declared that no financial support was received for this work, and the article names no funder, equipment lender or software provider with any say.
Do not take this as professional medical advice.

