experiment · Technology in society · la publicación, 18 mar 2026 · gratis
Los niños confundieron lo hecho por una inteligencia artificial con lo hecho por una persona
En una prueba con 37 niños de 6 a 10 años, en promedio no lograron distinguir textos e imágenes, y quienes más usaban aparatos en casa tuvieron peores resultados. Eso no prueba que las pantallas causen el problema.
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
- Niños de 6 a 10 años, con un grupo de adultos como referencia
- Cuántos
- 37 niños y 49 adultos
- Dónde
- Estados Unidos (una ciudad pequeña del este del país)
- Cuándo
- No lo dice el estudio
- Tipo de estudio
- experiment
- Quién lo hizo
- Departamento de Psicología y Neurociencia, Universidad de Temple
- El límite que importa
- Fueron 37 niños de una sola ciudad y de familias con ingresos altos; puede no repetirse en otros lugares.
Son los aciertos promedio de 37 niños de 6 a 10 años al decir si cada material lo hizo una persona o una máquina; la mitad (50%) sería puro azar, y el estudio usó versiones anteriores de los programas.
Aciertos de niños y de adultos en la misma prueba
Los adultos acertaron más que los niños; en promedio, los niños quedaron en la mitad o por debajo.
Aun cuando se les dijo que la mitad del material venía del SmartBot, una y otra vez lo tomaron por obra de una persona.
Un teléfono reproduce un cuento. Una tableta muestra una foto. Detrás de ambos pudo haber una inteligencia artificial, y nada en la pantalla lo advierte. Así ocurre ya en aplicaciones, juegos y aparatos de casa que los niños usan a diario.
Para medir qué tan bien lo notan, investigadores de la Universidad de Temple, en Estados Unidos, contaron a 37 niños de 6 a 10 años la historia de un "SmartBot", una inteligencia artificial que escribe textos y crea imágenes. En esa historia también aparecía una maestra llamada Ms. Shelby, que escribe textos y crea imágenes. Luego les pidieron decidir quién había hecho cada texto y cada imagen de objetos, arte y caras.
En conjunto, los niños acertaron el 49% de las veces, la mitad. Una moneda al aire habría hecho lo mismo. Con las caras les fue peor: 43%. Con los textos, 47%. Sólo con los objetos superaron la mitad, con 54%.
Los investigadores también preguntaron a los papás cuánto usaban sus hijos los aparatos en casa. Los niños con más uso tendieron a acertar menos. Conviene decirlo con cuidado: es una asociación observada en un solo grupo, y no muestra que las pantallas causen el problema. Son 37 niños de una ciudad pequeña del este de Estados Unidos, en su mayoría de familias con ingresos altos, y el resultado puede no repetirse donde usted vive.
Hay otro límite que importa hoy. Las imágenes y los textos se hicieron con versiones anteriores de estas herramientas, como ChatGPT 4.0 y Gemini 1.0. Las de ahora producen resultados distintos, y el estudio no dice qué tan bien reconocería su hijo lo que esas herramientas crean.
Lo que sí quedó claro es hacia dónde se inclinan los niños. Aun cuando se les dijo que la mitad del material venía del SmartBot, una y otra vez lo tomaron por obra de una persona. Esa inclinación es la que conviene vigilar.
La próxima vez que su hijo le muestre un cuento o una foto en la pantalla, pregúntele cómo sabe que la hizo una persona. La respuesta dice más que el aparato.
Qué significa para usted
Lo que este estudio sugiere es una pista para la conversación diaria, no una alarma. Cuando su hijo le muestre un texto o una foto, vale la pena que usted mismo se detenga a mirar si parece hecho por una persona, porque a estos niños les costó notar la diferencia. Y si le preocupa el tiempo de pantalla de su hijo, recuerde que aquí solo se vio una relación entre más uso y peores resultados, no una causa.
Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303
Quién pagó: El artículo indica que el trabajo recibió apoyo de una beca de posgrado de la National Science Foundation (subvención n.º 2038235) y del NIH R01 HD098097 (investigador principal: Chein); no dice si los financiadores tuvieron alguna influencia ni menciona equipos o programas informáticos prestados para el estudio.
Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1627 palabras · unos 8 minLeerla →Cerrar
Los niños de 6 a 10 años que no distinguen lo que hizo una persona de lo que hizo una máquina
Un estudio con 37 niños en Estados Unidos encontró que, en promedio, acertaron a la mitad de las veces. Los que más usaban aparatos en casa fueron los que peor acertaron.
Imagine la escena: un niño de siete años frente a la pantalla de una computadora, escuchando la voz de un programa que le cuenta que un robot llamado "SmartBot" puede escribir cuentos, inventar fotos y hacer dibujos. Después ve cómo ese robot, en vivo, escribe un cuento sobre el espacio y produce la imagen de un perro. Luego le muestran una foto de un perro de verdad y un texto corto sobre la luna, y le preguntan quién lo hizo: el SmartBot o una maestra llamada "Ms. Shelby". El niño tiene que elegir.
Eso hicieron 37 niños de entre 6 y 10 años, con una edad promedio de casi 8 años. Los investigadores les dijeron de antemano que la mitad del material venía de la maestra y la mitad del robot. Los niños practicaron cuatro rondas antes de empezar. Aun así, en promedio acertaron el 49 por ciento de las veces: la misma probabilidad que tiene una moneda de caer de un lado o del otro. Los adultos que hicieron la misma prueba acertaron el 59 por ciento. La diferencia entre ambos grupos fue clara y consistente1.
El estudio lo hicieron investigadores del Departamento de Psicología y Neurociencia de la Universidad de Temple, en Filadelfia. Lo pagaron una beca de posgrado de la National Science Foundation y un subsidio de los Institutos Nacionales de Salud de Estados Unidos. Los autores declaran que no tienen conflictos de interés. El trabajo se hizo por videollamada, en sesiones de aproximadamente una hora, y los niños recibieron una tarjeta de regalo de 10 dólares.
El patrón se repitió por tipo de material. Con objetos acertaron el 54 por ciento; con obras de arte, el 51; con textos, el 47; y con caras, apenas el 432. Es decir: los niños identificaron mejor una foto de un parque o de un helado que el rostro de una persona. Cuando se equivocaban, casi siempre se equivocaban en la misma dirección: tomaban por humano lo que había hecho la máquina3.
Los niños dieron sus propias razones, y ahí está lo interesante. Muchos hablaron de si la imagen "se veía real" o si tenía detalles. Uno dijo que una imagen era de la maestra "porque hay muchos detalles". Otro mencionó, como pista general, que a veces las imágenes generadas "se ven borrosas en el fondo". Buscaban las pistas que cualquiera buscaría: textura, luz, sombras, coherencia. El problema es que esas son justamente las pistas que la inteligencia artificial ya sabe imitar.
Casi todos los niños, el 94.6 por ciento, dijeron que el SmartBot no estaba vivo4. Sin embargo, alrededor de la mitad le atribuyó capacidades humanas: que podía tomar decisiones, pensar por sí mismo o elegir hacer cosas. No es que creyeran que el robot era una persona. Es que le reconocían una mente.
El dato que más pesa para una familia es otro. Los investigadores preguntaron a los papás cuánto usaba cada niño los aparatos de la casa: televisión, tableta, computadora, celular, videojuegos, bocinas inteligentes. Mientras más tiempo reportado de uso, peor fue el desempeño en la prueba5. La relación fue negativa y significativa6. El estudio no puede decir que una cosa cause la otra. Solo muestra que van juntas.
Hay que ponerle límites a la lectura de este estudio. Fueron 37 niños de una sola ciudad pequeña del este de Estados Unidos, en su mayoría de familias con ingresos altos; los propios autores advierten que los resultados podrían no repetirse en otros lugares o con otras condiciones de vida7. Las imágenes y los textos se generaron con versiones anteriores de los programas (ChatGPT 4.0, Gemini 1.0), así que las herramientas de hoy podrían ser más difíciles o más fáciles de distinguir. Además, a los niños se les dijo desde el principio que la mitad del material era artificial, algo que en la vida real casi nunca se avisa. Los padres, por su parte, dijeron de forma unánime que sus hijos no habían usado herramientas de inteligencia artificial, lo que sugiere que los adultos subestiman a qué están expuestos los niños.
Piense en una bocina inteligente en la sala de una casa. Según una encuesta citada por los autores, el 88 por ciento de los niños tiene acceso a una o es dueña de una, y el 45 por ciento de los niños de 6 a 9 años le habla todos los días8. Otra encuesta encontró que la mitad de los padres de niños de 3 a 12 años dice que su hijo ya usó herramientas como ChatGPT, para buscar información, pedir consejo o inventar cuentos y dibujos9. No hace falta imaginar un futuro lejano: la mayoría de los niños ya convive con esto.
Vale la pena entender qué hace exactamente esta tecnología, porque de ahí viene la dificultad. Los programas de inteligencia artificial generativa están diseñados, por definición, para producir contenido que imite lo que haría una persona10. No fallan cuando el resultado parece humano: funcionan cuando el resultado parece humano. Un texto escrito por una máquina busca sonar como lo escribiría alguien; una foto generada busca tener la luz, la piel y las imperfecciones de una foto real. La pregunta "¿esto lo hizo una persona?" se vuelve, entonces, una pregunta contra el diseño mismo de la herramienta.
¿Y qué se sabía antes de este estudio? Bastante sobre adultos, poco sobre niños. Varios trabajos previos habían encontrado que los adultos aciertan apenas un poco por arriba del azar al distinguir textos e imágenes generados por máquinas11. Es decir: el problema no es solo de los niños. Los adultos también dudan, también se equivocan, también confían en pistas que ya no sirven. Lo que faltaba era medir esa misma capacidad en la infancia, y ahí el estudio de Temple abre terreno: según sus autores, ningún trabajo publicado había examinado directamente si los niños pueden distinguir lo humano de lo artificial12.
Así lo lee weeklyAI. Hay una tentación muy natural en las casas: creer que un niño que maneja la tableta con soltura entiende lo que ve en ella. El estudio sugiere lo contrario, o al menos pone esa creencia en duda. La soltura es destreza con el aparato; la comprensión es otra cosa, y se entrena de otra manera. Si esto se sostiene, en las familias donde nadie comenta lo que aparece en pantalla, donde nadie dice en voz alta "esto no sé si es real", el niño probablemente no desarrolle el hábito de preguntárselo. Esta lectura se equivocaría si resulta que los niños que crecen entre pantallas, sin ninguna guía, acaban distinguiendo mejor que los demás. Los datos de este estudio no apuntan en esa dirección, pero tampoco lo descartan del todo.
Hay otra cosa que weeklyAI espera encontrar en las casas, y que el estudio no midió: el momento en que un niño se siente engañado. Un cuento puede estar bien escrito y una foto puede ser hermosa, y aun así, cuando el niño descubre que no salió de una persona, algo cambia en su forma de mirar. La calidad del contenido no es lo único que importa; también importa quién lo firma. Esta lectura se equivocaría si a los niños les diera igual el origen de lo que leen y ven. Usted puede comprobarlo en su propia casa, con una pregunta sencilla: "¿Quién crees que hizo esto?".
¿Qué se puede hacer con todo esto? Los propios autores son cautelosos. Señalan que enseñar a los niños a detectar lo artificial puede volverse cada vez menos confiable conforme avanza la tecnología, y que quizá convenga poner el acento en habilidades más amplias: evaluar la calidad de la información, buscar confirmación en otras fuentes, reconocer que la intuición y la confianza en uno mismo no bastan13. También mencionan que exigir a las plataformas que etiqueten el contenido generado por inteligencia artificial, o que mantengan una distinción clara entre lo que hace una máquina y lo que hace una persona, podría reducir el riesgo de que los niños se dejen llevar por contenido que no es humano14.
Conviene no exagerar la alarma. El estudio midió una sola cosa: si un niño puede decir de dónde salió un texto o una imagen. No midió si los niños creen las mentiras que encuentran, ni si son más manipulables que antes, ni si el uso de pantallas les hace bien o mal en general. Tampoco encontró que la edad importara: dentro del rango de 6 a 10 años, los mayores no acertaron más que los pequeños15. Y no evaluó ninguna intervención: no hay aquí evidencia de que un taller o una clase mejore la capacidad de distinguir.
El cierre vuelve a la escena del principio: el niño frente a la pantalla, con dos opciones y el dedo en el aire. Lo que usted puede llevarse de este estudio no es una alarma, sino una costumbre. Antes de pedirle a su hijo que desconfíe de lo que ve, revise usted mismo, en voz alta y junto a él, algo que le haya llegado y no sepa si es real. La pregunta útil no es solo "¿será falso?", sino "¿de cuándo es, quién lo hizo, qué le quitaron alrededor?". Esa pregunta se puede practicar con cualquier foto, cualquier video, cualquier cadena que llegue por WhatsApp. No requiere saber de tecnología. Requiere detenerse un momento y decirlo en voz alta. Y la próxima vez que su hijo le muestre algo en la pantalla, en lugar de preguntarle si le gusta, pregúntele: ¿quién crees que lo hizo, y cómo lo sabrías?
De dónde sale cada dato de contexto, y cuánto leímos de cada documento
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - el artículo del que trata esta nota — el artículo completo — el pasaje: "On average, children performed at or below chance in discriminating human-created from AI-generated content across modalities, and their overall accuracy was significantly below that of adults."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Average accuracy was highest for objects (M = 54%, SD = 19%), then art (M = 51%, SD = 18%), texts (M = 47%, SD = 19%), and faces (M = 43%, SD = 18%)."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - el artículo del que trata esta nota — el artículo completo — el pasaje: "A one-tailed t -test revealed that participants were significantly below chance in their overall discernment accuracy (t(36) = −2.12, p = .02), and systematically over-endorsed stimuli as human rather than AI (t(36) = −2.15, p = .02)."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Most children said the SmartBot was not alive (94.6%). However, around half of children said the SmartBot could make decisions (48.6%), think for itself (48.6%), and choose to do things (51.4%)."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Children’s individual discernment varied and was significantly and negatively associated with parent reports of children’s at-home technology use."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Parent-report technology use was significantly negatively correlated with detection accuracy, R = −0.52, p < .001, 95% confidence interval [−0.73, −0.24]."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Further, our participant pool was recruited solely from a small city in the eastern United States, with high average annual income and caregiver education levels. While such results are still informative, they may not generalize to children in different geographic or socioeconomic contexts with different exposure to, experience with, or beliefs about AI."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Children’s access to AI is also becoming increasingly widespread, with one survey showing that eighty-eight percent of children have access to, or own, a smart speaker, and forty-five percent of 6-to-9-year-olds speak to their smart speaker daily ( Bickham et al., 2024 )."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Surveys also show that children as young as 3 are using generative AI tools like ChatGPT—about half of parents of 3- to 12-year-olds report that their child has used generative AI, including for seeking information or advice and for creative activities such as storytelling or art ( Bickham et al., 2024 )."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Generative AI platforms are designed to produce content that mimics human-created content, raising questions about how children at different developmental stages will understand these systems’ capabilities and limits, and in turn how children evaluate and trust their content."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Empirical work shows that adults consistently perform near or slightly above chance when discerning AI-generated texts and images ( Chein et al., 2024 ; Dugan et al., 2023 ; Miller et al., 2023 ) and that this discernment ability may be linked to factors such as fluid intelligence and time spent on digital devices ( Chein et al., 2024 )."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - el artículo del que trata esta nota — el artículo completo — el pasaje: "However, no study has investigated this ability in children, or whether individual characteristics make children better able to detect AI-generated content."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Rather than focusing primarily on teaching children to detect AI-generated content—a skill that may become increasingly unreliable as AI advances—AI literacy interventions may need to emphasize broader critical thinking strategies, including how to evaluate information quality, seek verification from multiple sources, and recognize that confidence or intuition may not suffice for assessment of AI content."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - el artículo del que trata esta nota — el artículo completo — el pasaje: "AI watermarking and transparency, where platforms are required to label AI-generated content as such or maintain clear distinctions between AI’s capabilities and human-like traits ( Collyer-Hoar & Rubegni, 2025 ; Kurian, 2025 ), may reduce the risk that children are misled by non-human generated content."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Age, scores on the Raven’s Colored Progressive Matrices (RCPM), and scores on the Peabody Picture Vocabulary Test (PPVT) were not significantly correlated with detection accuracy (see Fig. 4 )."
Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303
Quién pagó: El artículo indica que el trabajo recibió apoyo de una beca de posgrado de la National Science Foundation (subvención n.º 2038235) y del NIH R01 HD098097 (investigador principal: Chein); no dice si los financiadores tuvieron alguna influencia ni menciona equipos o programas informáticos prestados para el estudio.
Los hallazgos de otros estudios que aquí se mencionan provienen de este documento; esos estudios no se consultaron uno por uno.
experiment · Technology in society · the paper, 18 Mar 2026 · free
Your Child May Not Know What a Computer Made
In a study of 37 children, they did no better than guessing which texts and pictures came from AI, and children who used more devices at home did worse. The study cannot say one caused the other.
Short version · the longer version follows, about 7 min
Ask the weeklyAI chat
Ask me about AI in your life: I point you to the stories weeklyAI has published on it, this week and before.
Conversations are saved for as long as weeklyAI exists, to improve the publication. Answers come in the language you write in.
- The study at a glance
- Who
- children ages 6 to 10, plus a separate group of adults for comparison
- How many
- 37 children and 49 adults
- Where
- a small city in the eastern United States
- When
- not stated
- Kind of study
- experiment
- Who did it
- Temple University
- The limit that matters
- One small city, mostly higher-income families; may not hold elsewhere.
Children compared with adults on the same sorting test
Adults were right more often: 59 percent versus 49 percent for children.
The children were right 49 percent of the time.
Your child meets machine-made words and pictures every day, in apps, in games, in a smart speaker on the kitchen counter, often with nothing on them saying where they came from. A study from Temple University asked whether children can tell.
The researchers told 37 children, ages 6 to 10, about a "SmartBot" that writes text and makes images. They also told them about a human teacher, Ms. Shelby. The children watched the SmartBot produce a story and a picture on the spot. Then they sorted texts and photos as the SmartBot's or Ms. Shelby's.
The children were right 49 percent of the time. They did worse than guessing on faces and on texts: 43 percent and 47 percent. Adults who took the same test scored 59 percent, which is better and still far from sure.
The children were not confused about the machine itself. Most, 95 percent, said the SmartBot is not alive. Separately, the children leaned toward calling content a person's work whether it was or not.
Parents reported how often their children used screens, tablets and voice assistants. Children with more use did worse on the sorting. This is a pattern in one group; the study cannot say that screen time caused it, and cutting it is not a proven fix.
The sample came from one small city in the eastern United States, mostly from families earning more than $100,000 a year. Your child and your community may be different. The test material also came from older AI tools, ChatGPT 4.0 and Gemini 1.0, so the study cannot tell you how your child would do with what tools make today. And within these ages, 6 to 10, it found no sign of skill growing with age.
What it does show is a pull to assume a person made what a child sees. Next time a child shows you a photo or reads you a story, look at how sure they are that someone real made it.
What this means for you
For now, watch how readily a child hands over a photo or a story as someone's real work, and say plainly that a machine can make one too. Whether more time on devices makes this harder, or only travels with it, the study cannot say.
Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303
Who paid: The article says the work was supported by a National Science Foundation Graduate Research Fellowship (Grant No. 2038235) and NIH R01 HD098097 (PI: Chein); it does not say whether the funders had any say, and it names no equipment or software lent to the study.
The longer versionThe passages copied from the paper, the pictures, and every source with how much of it we read · 1387 words · about 7 minRead it →Close
Children Couldn’t Tell AI From People. The More Screen Time, the Worse They Did.
A study of 37 children found they were no better than a coin toss at spotting AI-written text and AI-made pictures, and more device use went with worse scores.
The test looked simple. A child saw a short text or a picture and chose between two answers: a human teacher made this, or a “SmartBot” did. The researchers told the children in advance that half the items came from each source. The children still averaged 49 percent — a hair under a coin toss1. Adults who took the same test with the same items averaged 59 percent1.
The children completed the study remotely over Zoom, and were recruited through a developmental psychology laboratory on a university campus in a small city in the eastern United States. The work was supported by a National Science Foundation Graduate Research Fellowship and an NIH grant. It ran 37 children ages 6 to 10, tested over Zoom, plus a separate group of 49 adults ages 18 to 34 for comparison. The children heard a story about a SmartBot, watched it write a story and make a picture on the spot, practiced four rounds, and then judged 128 items: short texts, and pictures of objects, art and faces.
The children did not simply guess at random. They leaned toward “human”: they were more likely to call an AI-made item human-made than the reverse, and that bias was statistically reliable2. By category they were best with objects (54 percent) and worst with faces (43 percent), with art at 51 percent and texts at 47 percent3. Their answers fell clearly below a coin toss for faces and for texts, while objects were the one category where they did better than a coin toss3.
What makes a face hard? A face carries no obvious label, and a good AI face looks like any other face. The article describes a study in which adults also tend to over-identify AI faces as human, and in which AI faces are hard to tell from real ones4. The authors write that children’s failure on faces, and their habit of taking AI faces for human ones, matters for the risk of manipulating children online5. A parent reading that sentence should hear it as a risk the authors name, not as something the study measured.
Before this study, the question had not been asked of children. The article describes studies in which adults land near or slightly above chance4. The authors state that no study had looked at whether children can do this, or at what makes one child better at it than another6. This is the first look, and first looks are small.
Two things did not predict a child’s score: age, and scores on tests of reasoning and vocabulary7. A 6-year-old and a 10-year-old were, on average, about equally likely to be fooled. The one factor that did move with accuracy was the parents’ report of how often the child used devices at home — tablets, phones, televisions, voice assistants and the like. More use went with worse accuracy, a moderate negative link89. The study cannot say which way the arrow points. Heavy device use may dull the eye for what is artificial; or children already less attentive to sources may drift toward more screens; or some third thing may drive both.
The children were not blank about AI. Nearly 95 percent said the SmartBot was not alive10. About half, though, said it could make decisions, think for itself and choose to do things10. They also reached for reasons. Asked why they chose one source, they talked most about specific details of the item, then about whether it looked “real,” then about what the SmartBot or the teacher would plausibly do. One child trusted a photo because of its detail: only a camera, the child reasoned, could capture burnt corn that precisely. The reasoning was sound. The clue was not.
The scale of what children meet is worth holding next to that. One survey cited by the authors found that about half of parents of 3- to 12-year-olds say their child has used generative AI, including to look things up and to make stories or art11. Another found that 88 percent of children have access to or own a smart speaker, and 45 percent of 6- to 9-year-olds speak to one daily12. The tools are built to imitate: generative systems are designed to produce content that mimics what people make13. A child’s judgment is being tested by design, not by accident.
The limits are as important as the result. The 37 children came from one small city in the eastern United States, and most families reported incomes above $100,000 a year14. The authors say plainly that the findings may not hold for children in other places or other economic circumstances, with different exposure to AI14. The sample is small, which makes the size of the device-use link uncertain. The AI used to make the test items was older: ChatGPT 4.0, Gemini 1.0 and an earlier face generator. Today’s tools may be harder to spot, or easier; the study does not say. And the parents all reported that their children had not used generative AI, which the authors read as a sign that parents may be underestimating what their children already touch.
Here is weeklyAI’s reading. The children in this study were not lazy or careless. They were doing what people do: judging by look and feel. A picture seemed real, a passage had detail, a voice sounded sure. Those shortcuts are exactly what these systems are built to satisfy. What the study suggests is that the shortcut and the skill are different things, and that time with a screen may train the first while leaving the second untouched. weeklyAI would expect, in homes like the reader’s, that a child who is quick with a tablet is not thereby quick at spotting a fake — and that the gap may be widest in the houses with the most devices. This reading would be wrong if children who use screens heavily turned out to catch up with light users as they grow, or if simply handling AI tools a lot made a child reliably better at spotting their output. Watch your own child at the kitchen table: when a picture or a story comes up, does your child reason from how it looks, or from where it came from? The second question is the one that travels.
A second expectation, and this one is cheerier. The children in the study were already reaching for explanations, already naming the SmartBot’s tricks, already saying “this looks unreal.” That is raw material. The authors themselves argue against making detection the whole lesson, since detection may get harder as the tools improve; they point instead to teaching children to weigh where information comes from, to check it against more than one source, and to notice that feeling sure is not the same as being right15. They also point to labeling: platforms that mark AI-made content, or that keep a clear line between what a machine can do and what a person is16. Those are things a family can ask for, from a platform and from a school. This reading would be wrong if children who are taught to ask “who made this?” did no better than children taught to look harder at the picture. Nobody has tested that yet, and the study does not claim otherwise.
One more caution. The children were told that half the items were AI-made. In the real world, almost nothing is labeled. The authors note that this makes their finding, if anything, an underestimate of how often children would take machine output for human work.
The piece opened on a child’s finger hovering between two answers. What the reader now knows is that the finger was not the problem, and neither was the child. The habit worth building is smaller and more portable than a lesson in spotting fakes: before anything is passed along or believed, ask where it came from, and how you could find out. Try it once this week, out loud, on something you both already know is fake. Then ask your child the question the study leaves open: how would you check?
Where each piece of context comes from, and how much of it we read
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - the article this story is about — the whole article — the passage: "On average, children performed at or below chance in discriminating human-created from AI-generated content across modalities, and their overall accuracy was significantly below that of adults."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - the article this story is about — the whole article — the passage: "A one-tailed t -test revealed that participants were significantly below chance in their overall discernment accuracy (t(36) = −2.12, p = .02), and systematically over-endorsed stimuli as human rather than AI (t(36) = −2.15, p = .02)."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - the article this story is about — the whole article — the passage: "Average accuracy was highest for objects (M = 54%, SD = 19%), then art (M = 51%, SD = 18%), texts (M = 47%, SD = 19%), and faces (M = 43%, SD = 18%)."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - the article this story is about — the whole article — the passage: "Empirical work shows that adults consistently perform near or slightly above chance when discerning AI-generated texts and images ( Chein et al., 2024 ; Dugan et al., 2023 ; Miller et al., 2023 ) and that this discernment ability may be linked to factors such as fluid intelligence and time spent on digital devices ( Chein et al., 2024 )."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - the article this story is about — the whole article — the passage: "Children’s inability to distinguish human and AI-generated faces, and their overattributing of AI faces as human, has important consequences for the potential of manipulation of children online."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - the article this story is about — the whole article — the passage: "However, no study has investigated this ability in children, or whether individual characteristics make children better able to detect AI-generated content."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - the article this story is about — the whole article — the passage: "Age, scores on the Raven’s Colored Progressive Matrices (RCPM), and scores on the Peabody Picture Vocabulary Test (PPVT) were not significantly correlated with detection accuracy (see Fig. 4 )."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - the article this story is about — the whole article — the passage: "Children’s individual discernment varied and was significantly and negatively associated with parent reports of children’s at-home technology use."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - the article this story is about — the whole article — the passage: "Parent-report technology use was significantly negatively correlated with detection accuracy, R = −0.52, p < .001, 95% confidence interval [−0.73, −0.24]."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - the article this story is about — the whole article — the passage: "Most children said the SmartBot was not alive (94.6%). However, around half of children said the SmartBot could make decisions (48.6%), think for itself (48.6%), and choose to do things (51.4%)."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - the article this story is about — the whole article — the passage: "Surveys also show that children as young as 3 are using generative AI tools like ChatGPT—about half of parents of 3- to 12-year-olds report that their child has used generative AI, including for seeking information or advice and for creative activities such as storytelling or art ( Bickham et al., 2024 )."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - the article this story is about — the whole article — the passage: "Children’s access to AI is also becoming increasingly widespread, with one survey showing that eighty-eight percent of children have access to, or own, a smart speaker, and forty-five percent of 6-to-9-year-olds speak to their smart speaker daily ( Bickham et al., 2024 )."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - the article this story is about — the whole article — the passage: "Generative AI platforms are designed to produce content that mimics human-created content, raising questions about how children at different developmental stages will understand these systems’ capabilities and limits, and in turn how children evaluate and trust their content."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - the article this story is about — the whole article — the passage: "Further, our participant pool was recruited solely from a small city in the eastern United States, with high average annual income and caregiver education levels. While such results are still informative, they may not generalize to children in different geographic or socioeconomic contexts with different exposure to, experience with, or beliefs about AI."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - the article this story is about — the whole article — the passage: "Rather than focusing primarily on teaching children to detect AI-generated content—a skill that may become increasingly unreliable as AI advances—AI literacy interventions may need to emphasize broader critical thinking strategies, including how to evaluate information quality, seek verification from multiple sources, and recognize that confidence or intuition may not suffice for assessment of AI content."
- Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303 - the article this story is about — the whole article — the passage: "AI watermarking and transparency, where platforms are required to label AI-generated content as such or maintain clear distinctions between AI’s capabilities and human-like traits ( Collyer-Hoar & Rubegni, 2025 ; Kurian, 2025 ), may reduce the risk that children are misled by non-human generated content."
Langer, A., Martinez, S., Marshall, P. J. et al. (2026). Children's susceptibility to content generated by artificial intelligence. Technology in Society. https://doi.org/10.1016/j.techsoc.2026.103303
Who paid: The article says the work was supported by a National Science Foundation Graduate Research Fellowship (Grant No. 2038235) and NIH R01 HD098097 (PI: Chein); it does not say whether the funders had any say, and it names no equipment or software lent to the study.
The findings of other studies mentioned here come from this document; those studies were not opened one by one.

