An AI Lesson Plan Can Feel Like Progress. In 90 Classrooms, It Didn't Look Like It.

experiment · Frontiers in Psychology · la publicación, 6 oct 2026 · gratis

La IA que te escribe las preguntas de la clase

Un estudio siguió a 90 profesores de matemáticas en China durante un año escolar: cuando el orden de las preguntas venía del modelo, las clases que siguieron tuvieron menos preguntas de razonamiento y menos ideas de alumnos aprovechadas. Las estimaciones principales son observacionales.

Versión breve · la versión detallada sigue, unos 5 min

Pregunte al chat de weeklyAI

Pregúnteme por este estudio: a quiénes se estudió, qué encontró y qué no dice.

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: Profesores de matemáticas de secundaria
Cuántos: 90 profesores, 897 clases válidas
Dónde: Una ciudad-prefectura del este de China
Cuándo: Un año escolar completo
Tipo de estudio: analysis of what people did
Quién lo hizo: Universidades de China (Northeast Normal University y otras)
El límite que importa: Las estimaciones principales son observacionales: no muestran que delegar haya causado la diferencia.

Dos formas de preparar la clase con el modelo

Primero el borrador del modelo, después revisarlo (259 clases)frente aPrimero escribir lo propio, después consultar el modelo (403 clases)

En el primer grupo, el orden de las preguntas vino del modelo en 0.591 de los casos, hubo 0.189 de preguntas de razonamiento y 0.211 de retoma sustantiva. En el segundo, 0.210, 0.264 y 0.335.

Mire lo que se queda: la secuencia de preguntas y el orden de los ejemplos es la parte que este estudio ata a lo que los estudiantes alcanzan a decir.
Lectura de weeklyAI

Ver · 1:03

Usted prepara una clase con un chatbot. Le pide la secuencia de preguntas, la revisa, y se siente listo. Eso tiene un nombre en este estudio: dejar en manos de la IA la capa generativa, es decir, el orden de las preguntas y la secuencia de variaciones.

Noventa profesores de matemáticas de secundaria en el este de China fueron grabados en video a lo largo de cuatro oleadas de clases durante un año escolar. En total, 897 clases válidas y 4,640 aportes de estudiantes que se salieron de lo previsto. Los investigadores compararon cada plan final con los registros de las conversaciones con la IA, segmento por segmento.

Cuando la secuencia de preguntas venía del modelo, las clases que siguieron tuvieron menos preguntas de razonamiento, menos respuestas que retomaban lo que decían los alumnos, y menos uso de lo que los estudiantes dijeron y el plan no anticipaba. Las diferencias fueron pequeñas o moderadas.

Cuando del modelo salieron sólo los materiales y la redacción, esos costos no aparecieron. Al contrario: el diseño de variaciones fue más largo y variado.

El estudio no puede mostrar que delegar haya causado la diferencia. Las estimaciones principales son observacionales: comparan a cada profesor consigo mismo, pero no descartan que una semana difícil lo llevara a delegar más y atender menos. En una comparación al azar, a cada profesor se le asignó una clase para preparar primero sus propias preguntas y después consultar al modelo. Esa asignación se asoció con que retomaran más las ideas de los alumnos. Para los demás resultados de la clase, el estudio no pudo detectar cambios. El estudio no puede descartar que existan.

Su confianza después de preparar subió junto con la parte de las preguntas que entregó al modelo. Y esa confianza no dijo nada sobre la clase que vino después.

Mire lo que se queda: la secuencia de preguntas y el orden de los ejemplos es la parte que este estudio ata a lo que los estudiantes alcanzan a decir.

Qué significa para usted

Lo que usted puede vigilar es el orden de sus preguntas y de sus ejemplos, la parte que este estudio asocia con lo que los alumnos alcanzan a decir en clase. No concluya que usar la IA en la preparación perjudica: fueron 90 profesores de matemáticas en el este de China, y las estimaciones principales son observacionales.

Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005

Quién pagó: El trabajo fue financiado por el Proyecto de Planificación de Ciencias de la Educación Nacional 2025 (Subvención n.º XRA250378); los autores declararon no tener relaciones comerciales o financieras que pudieran constituir un conflicto de intereses, y el artículo no indica que ningún financiador, proveedor de equipos o de software haya intervenido en el estudio.

Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1093 palabras · unos 5 minLeerla →Cerrar

La IA que prepara la clase: lo que se delega importa más que cuánto se usa

Un año de clases grabadas en China sugiere que ceder a la máquina el orden de las preguntas se asocia con un diálogo de aula más pobre, y que un protocolo sencillo lo mejora.

Así podría verse · ilustración generada por weeklyAI.watch, no es una fotografía

Noventa profesores de matemáticas de secundaria en el este de China aceptaron que sus clases fueran grabadas en video durante un año escolar completo. Eran 90 docentes, 165 cursos y 897 clases válidas, en 18 escuelas públicas de una ciudad-prefectura. Cada vez que preparaban una clase con un chatbot, el sistema guardaba un registro de esa conversación. Después, alguien comparaba el plan final, segmento por segmento, con esos registros para ver de dónde venía cada parte. De ahí salieron dos proporciones. Una mide cuánto del orden de las preguntas y de la secuencia de variaciones salió del modelo. La otra, cuánto del material de contexto y de la redacción salió del modelo¹.

Los resultados no fueron iguales para las dos. Cuando el docente había delegado la parte de organizar las preguntas y las variaciones, sus clases mostraban menos preguntas de orden superior, menos retoma sustantiva de lo que decían los estudiantes, una demanda cognitiva que se sostenía peor y menos reconocimiento y uso productivo de los aportes no previstos². Los tamaños eran modestos: entre menos de una décima y menos de una quinta parte de una desviación estándar. Cuando lo delegado era el material y la redacción, en cambio, no apareció ese costo, y sí más cadenas de variación y más tipos de variación en el diseño². El tiempo total de preparación con la herramienta predijo algo por sí solo, pero perdió ese valor en cuanto se tuvo en cuenta qué capas se habían delegado³.

La parte más fuerte del estudio es un pequeño experimento metido dentro del año de observación. En las dos últimas olas, a cada docente se le sorteó una clase: en ella debía escribir primero, sin ayuda, su propia secuencia de preguntas y variaciones, y solo después consultar al modelo. Esa asignación bajó la delegación de la capa generativa y subió la retoma sustantiva de lo que decían los estudiantes⁴. Los docentes, además, dijeron sentirse más seguros de su preparación cuando más habían delegado la organización de las preguntas. Esa confianza no anticipó nada de lo que después se vio en el aula: una vez consideradas las dos capas, no predijo ningún resultado⁵. El estudio también midió cuánto hablaban los docentes, en entrevistas posteriores, sobre lo que pensaban sus estudiantes, y esa atención apareció como posible puente entre lo delegado y lo que ocurre en el momento⁶.

Conviene detenerse en lo que este estudio no puede decir. Las estimaciones principales son observacionales: comparan a cada docente consigo mismo, lo que elimina diferencias estables entre personas, pero no descarta que una semana difícil lo llevara a delegar más y a atender menos. La muestra viene de una sola ciudad-prefectura, un solo país, una sola materia y un solo nivel escolar, y el estudio no fue preregistrado⁷. La atención se midió después de la clase, así que su dirección queda abierta. Las preguntas de orden superior se codificaron de la clase, pero el estudio no codificó el nivel cognitivo de las preguntas tal como estaban escritas en el plan, de modo que la asociación podría describir lo que produce el modelo y no un cambio en el docente. Y no detectar un costo en la capa de redacción no es prueba de que no lo haya.

Puesto en palabras llanas, el estudio separa dos cosas que hasta ahora se medían juntas. Una es armar el texto de una explicación y reunir materiales: un costo de manejo, de esos que cansa hacer y que tiene sentido soltar¹. La otra es decidir en qué orden se preguntará y cómo se irán variando los ejemplos para que un concepto se entienda: eso es organizar e integrar, la operación que sostiene la comprensión⁸. Un chatbot puede hacer las dos. La diferencia es cuál de las dos le entregamos.

Otros trabajos que leímos apuntan en la misma dirección, aunque con otras poblaciones. Un estudio con estudiantes de secundaria que practicaban inglés oral con conversaciones y retroalimentación de un chatbot encontró mejoras en vocabulario, gramática, fluidez e interacción frente a la enseñanza tradicional; solo pudimos leer el resumen, el texto completo está detrás de una suscripción⁹¹⁰. Un ensayo aleatorizado con 95 estudiantes de ciencias comparó recibir la herramienta sin más con recibir una sesión de entrenamiento para usarla como socio y no como máquina de respuestas, y el grupo entrenado rindió mejor al corregir sus propios errores¹¹¹²¹³. Y un estudio con 531 estudiantes universitarios en España encontró que la IA cambiaba sobre todo la confianza al explicar un texto, más que la comprensión real¹⁴.

Así lo leemos nosotros. El patrón que se repite en estos trabajos no es "la máquina ayuda" ni "la máquina daña". Es que cuando una herramienta nos entrega el trabajo ya resuelto, sentimos que avanzamos rápido y confiamos en el resultado, aunque no hayamos ejercitado las decisiones que ese trabajo exigía. Eso es lo que esperaríamos ver en un aula como la suya: un docente que deja al modelo la secuencia de preguntas se siente más seguro de su preparación, y esa seguridad no le anticipa cómo va a responder cuando un estudiante dice algo que el plan no preveía. Podríamos estar equivocados. Si la confianza posterior a la preparación predijera de verdad la calidad del diálogo, o si quienes delegan más la estructura de preguntas manejaran igual o mejor lo imprevisto, esta lectura se caería. Y hay un límite de fondo: no sabemos si este patrón se sostiene fuera de las aulas de matemáticas de secundaria donde se midió.

Lo que usted puede hacer con esto es concreto. Antes de abrir el chatbot, escriba su propia secuencia de preguntas y de variaciones. Después compare qué conservó y qué cambió, y fíjese si eso altera su manera de escuchar a los estudiantes. Otra opción, si tiene con quién: pida a un colega que observe una clase y anote cuántas veces usted retomó una idea no prevista de un estudiante. Compárelo con lo que usted creía haber hecho, sin usar eso para calificar a nadie. Y si prepara con una máquina, anote aparte qué preguntas y en qué orden las diseñó usted mismo; al terminar la clase, revise si esa sensación de seguridad se correspondió con lo que de verdad ocurrió en el aula.

¿Qué parte de su próxima clase quiere seguir decidiendo usted?

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

Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Composing the wording of an explanation and assembling contextual materials are handling costs, and delegating them should free capacity. Designing the questioning sequence and ordering the variations through which a concept will be encountered is organizing and integrating. We call these the formulation layer and the generative layer."
Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Within teachers, generative-layer offloading was associated with less higher-order questioning, less substantive uptake of student talk, weaker maintenance of high cognitive demand, and lower recognition and substantive use of unplanned student contributions, with standardized coefficients between 0.08 and 0.18."
Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Total preparation time with AI predicted discourse outcomes on its own and lost that value for the lesson-level indicators once layer composition was entered."
Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005 - el artículo del que trata esta nota — el artículo completo — el pasaje: "An embedded micro-randomized comparison, in which teachers drafted their own questioning and variation sequence before consulting the model, lowered generative-layer offloading and raised substantive uptake."
Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Once both offloading layers were entered, confidence predicted no outcome (all p > 0.12), so what little it tracked was inherited from the delegation that produced it."
Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The indirect path was −0.038 with a Monte Carlo 95% confidence interval of [−0.065, −0.015], and the remaining direct path was −0.005 (p = 0.904)"
Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The sample comes from one prefecture-level city, one country, one subject, and one phase of schooling, and the study was not preregistered."
Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Generative processing consists of organizing incoming material into a coherent structure and integrating it with prior knowledge, and selection is explicitly excluded from it (Fiorella and Mayer, 2016; Mayer, 2024)."
Nageen S, Khawaja HUH, Sarwar M, Moin M, Imran Z, Jabeen S. (2026). Integration of ChatGPT to enhance English speaking skills among secondary school students in Pakistan. Discover Education. https://doi.org/10.1007/s44217-026-02278-z — solo el resumen - no se pudo obtener el texto completo — el pasaje: "The findings showed that the experimental group performed significantly better than the control group in both monologue and dialogue tasks, demonstrating notable improvements in lexical range, grammatical accuracy, fluency, interaction, and coherence."
Nageen S, Khawaja HUH, Sarwar M, Moin M, Imran Z, Jabeen S. (2026). Integration of ChatGPT to enhance English speaking skills among secondary school students in Pakistan. Discover Education. https://doi.org/10.1007/s44217-026-02278-z — solo el resumen - no se pudo obtener el texto completo — el pasaje: "The experimental group received a 12-week ChatGPT-assisted learning module in which ChatGPT was integrated to provide interactive speaking practice through voice- and text-based conversations with feedback, whereas the control group received traditional instruction."
Huang B, Xie L, Liu Y, Liu X, Tang H, Feng X, et al. (2026). From answer machine to collaborative partner: impact of structured generative AI training on scientific problem solving in STEM education. International Journal of STEM Education. https://doi.org/10.1186/s40594-026-00652-9 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "On the basis of Zimmerman’s self-regulated learning (SRL) framework and cognitive load theory (CLT), our randomized controlled trial (RCT; N = 95) investigated whether a structured, SRL-aligned scaffolding training session could improve students’ AI-assisted revision performance in electromagnetism problems."
Huang B, Xie L, Liu Y, Liu X, Tang H, Feng X, et al. (2026). From answer machine to collaborative partner: impact of structured generative AI training on scientific problem solving in STEM education. International Journal of STEM Education. https://doi.org/10.1186/s40594-026-00652-9 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Independent samples t tests revealed that the experimental group showed greater AI-assisted revision performance ( d = 0.81, p < 0.001) than did the control group."
Huang B, Xie L, Liu Y, Liu X, Tang H, Feng X, et al. (2026). From answer machine to collaborative partner: impact of structured generative AI training on scientific problem solving in STEM education. International Journal of STEM Education. https://doi.org/10.1186/s40594-026-00652-9 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "The trained students achieved an incorrect-to-correct revision rate of 69.3%, compared with 41.3% for the control group."
Marquès-Donoso A, Benguría-Aguirreche J, Neira-Calama M, Sánchez-Huete JC. (2026). Generative AI Use and Reading Comprehension: Associations with Metacognitive Experience in Higher Education—A Quasi-Experimental Study. Youth. https://doi.org/10.3390/youth6040144 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Generative AI use was associated with changes in the subjective experience of reading, particularly explanatory confidence, more consistently than in actual performance."

experiment · Frontiers in Psychology · the paper, 6 Oct 2026 · free

An AI Lesson Plan Can Feel Like Progress. In 90 Classrooms, It Didn't Look Like It.

Teachers who let a chatbot write the question sequence taught differently. The study watched lessons; it cannot prove the chatbot caused it.

Short version · the longer version follows, about 6 min

Ask the weeklyAI chat

Ask me about this study: who was studied, what it found, and what it does not say.

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: junior secondary mathematics teachers
How many: 90
Where: eastern China
When: one school year
Kind of study: analysis of what people did
Who did it: Northeast Normal University and others
The limit that matters: Observational; cannot prove the AI caused the differences.

Lesson preparation order: model draft first vs. own draft first

Model draft first, then revisedagainstDrafted own questions first, then consulted model

The model-first group had more generative-layer offloading (0.591 vs. 0.210), less higher-order questioning (0.189 vs. 0.264), and less substantive uptake (0.211 vs. 0.335).

In this study, the sequence of questions and the order of examples was the part tied to what students got to say.
weeklyAI's reading

Watch · 1:00

You open a chatbot, describe your lesson, and it hands you a finished sequence of questions and examples. You read it, change a few words, and you feel ready.

Researchers in eastern China followed 90 junior secondary mathematics teachers for one school year. They video-recorded 897 lessons across four waves. For each lesson, they compared the final plan against the AI chat logs, segment by segment, to see which parts the model had supplied.

The teachers were not a random slice of the profession. They taught in 18 public schools in one prefecture-level city, and the study was not preregistered.

When the questioning sequence and the order of examples came from the model, the lessons that followed showed fewer questions that asked students to analyze or argue. Teachers took up fewer student ideas. They made less use of what students said that the plan had not anticipated. The differences were small to moderate. A teacher's lesson with more of that kind of delegation, compared to that same teacher's own average, went with about a sixth of a step down on those measures.

When only the wording and the materials came from the model, those costs did not show up. The designed examples were longer and more varied, and teachers noticed more of what students said off script.

The study's main comparisons are observational. It followed the same teachers across lessons, which rules out stable differences between teachers, but not a bad week. A teacher who delegates more in a stressful week may also pay less attention.

One part was random. In the last two waves, some teachers were told to draft their own questions before consulting the model. Being assigned to draft their own questions first raised how often they took up student ideas. For other measures, the study could not detect a difference.

After preparing, teachers rated how ready they felt to handle whatever students would do. That confidence rose with how much of the questioning they had handed over, by about a quarter of a step per step of delegation. It told them nothing about the lesson that followed.

In this study, the sequence of questions and the order of examples was the part tied to what students got to say.

What this means for you

If you prepare lessons with a chatbot, notice which part you hand over. In this study, teachers who worked out their own question sequence and order of examples before asking the model were later seen using more of what students said. When teachers were told to do this, their use of student ideas went up.

Who paid: The work was supported by the 2025 National Education Sciences Planning Project (Grant No. XRA250378); the authors declared no commercial or financial relationships that could be a conflict of interest, and the article does not say that any funder, equipment or software provider had a say in the study.

The longer versionThe passages copied from the paper, the pictures, and every source with how much of it we read · 1123 words · about 6 minRead it →Close

# The Part of Lesson Planning Teachers Kept for Themselves

In 90 Chinese maths classrooms, what teachers handed to an AI — not how much they used it — tracked how the lesson went. A small experiment suggests an order that helps.

How it could look · illustration generated by weeklyAI.watch, not a photograph

Give a chatbot the job of writing your lesson plan, and it will write you a plan. The question a new study asked is narrower and more interesting: which part of the thinking did you hand over, and which part did you keep? Ninety junior secondary mathematics teachers in eastern China were followed across four waves of video-recorded lessons over one school year — 897 coded lessons, 1,518 preparation sessions logged against the plans they produced, and 4,640 moments in which a student said something the lesson had not anticipated. Every final plan was compared against the AI logs, segment by segment, to see where its content had actually come from.¹

The authors split preparation into two layers. The formulation layer is the handling work: the wording of an explanation, the contextual materials, the examples you gather. The generative layer is the organizing work: the sequence of questions, the order in which variations of a concept are met. Composing wording and assembling materials are costs, and handing them over should free capacity; designing the questioning sequence is the part that builds understanding, and the authors call it generative for that reason.²¹ The two ratios barely moved together — they correlated only weakly — which suggests two separable practices rather than one habit of delegating a lot or a little.

Within teachers — comparing each teacher's own lessons to her own average — the generative layer went with less higher-order questioning, less uptake of what students said, weaker maintenance of demanding tasks, and less recognition and use of unplanned student contributions. The standardized coefficients ran between 0.08 and 0.18, small-to-moderate.³ The formulation layer behaved differently: it was unrelated to four of those indicators, positively associated with the length and variety of designed variation sequences, and positively associated with recognition of unplanned contributions — the one indicator where the two layers point in opposite directions.³

Total time spent preparing with AI predicted weaker discourse on its own, and lost that predictive value for the lesson-level indicators once the layer composition was entered.⁴ That is the finding's sharpest edge: the field has been counting minutes and asking about adoption, and minutes turn out to be carrying someone else's information.

An embedded randomized comparison ran in the two later waves. One lesson per teacher was designated for a protocol in which the teacher drafted her own questioning and variation sequence *before* consulting the model, restricting the AI to materials and wording. Assignment lowered generative-layer offloading and raised substantive uptake of student talk.⁵ Observationally, the same ordering appeared: lessons prepared by taking a model draft first and revising it had mean generative-layer offloading of 0.591, higher-order questioning of 0.189 and substantive uptake of 0.211; lessons prepared by drafting first and consulting the model afterward had 0.210, 0.264 and 0.335.⁶

Confidence is where this study gets uncomfortable. After preparation and before teaching, teachers rated how ready they felt to handle students departing from the plan. That confidence rose with generative-layer offloading by about 0.23 standard deviations per within-teacher standard deviation of the ratio — a real, sizeable rise. It carried no positive information about the lesson that followed. Once both offloading layers were entered, confidence predicted no outcome at all, so whatever little it tracked was inherited from the delegation that produced it.⁷

Attention to student thinking is the proposed carrier. Measured from interviews conducted after the lesson, the share of a teacher's remarks that addressed student thinking followed the same within-teacher pattern, and it in turn predicted substantive use of unplanned contributions — an indirect path of −0.038, with the direct path essentially zero.⁸ The authors are careful here, and so should we be: the interviews came *after* the lesson, so a teacher who took up student thinking had more student thinking to talk about. The measurement does not establish which came first, and the authors report it as a mechanism probe rather than a result.

Other limits matter for what you should take from this. The primary estimates are observational; comparing a teacher to herself removes stable traits but not a bad week in which she both delegates more and attends less. Higher-order questioning is coded from the lesson, but the questioning sequence largely originates in the plan, and the study did not code the cognitive level of questions as written — so that association could describe shallower model output rather than any change in the teacher. The variation measures are an operational segmentation, and longer, more varied sequences are not automatically better teaching. The sample is one prefecture-level city, one country, one subject, one school phase, and the study was not preregistered.⁹

Here is how we read it. Two things seem to be happening at once, and the study can see both but cannot fully separate them. First, a finished-looking plan feels like mastery, so the person holding it reads fluency as readiness. Second, and quieter: designing the sequence is a rehearsal. Whoever builds the order of questions has already walked through the lesson once in their head, and that walk is what makes a teacher fast when a student goes off-script. Hand over the design and you may still deliver a competent lesson — you have simply stopped practicing the move that improvisation requires. We would expect this to apply well beyond a maths classroom, and beyond teaching: to anyone who uses a tool to produce the structure of a piece of work and then wonders why the work feels harder to defend in the room. How would you know we are wrong? If teachers who delegated the questioning design felt no more confident than those who wrote their own, or if their confidence tracked how the lesson actually went, the pattern would not hold.

What you can do with it costs nothing and takes a few minutes. Next time you plan anything with an AI tool — a lesson, a workshop, a talk — write your own key questions and the order of your examples first, then let the tool do the materials and the wording. Afterwards, ask yourself one plain question: what did I actually decide here, and what did I accept? Notice whether your sense of readiness rests on having thought it through or merely on holding a complete-looking document. If it is the second, you have found the gap this study is pointing at. Then watch what happens in the room when someone says something you did not plan for.

What did you decide, and what did you accept?

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

Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005 - the article this story is about — the whole article — the passage: "Composing the wording of an explanation and assembling contextual materials are handling costs, and delegating them should free capacity. Designing the questioning sequence and ordering the variations through which a concept will be encountered is organizing and integrating. We call these the formulation layer and the generative layer."
Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005 - the article this story is about — the whole article — the passage: "Generative processing consists of organizing incoming material into a coherent structure and integrating it with prior knowledge, and selection is explicitly excluded from it (Fiorella and Mayer, 2016; Mayer, 2024)."
Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005 - the article this story is about — the whole article — the passage: "Within teachers, generative-layer offloading was associated with less higher-order questioning, less substantive uptake of student talk, weaker maintenance of high cognitive demand, and lower recognition and substantive use of unplanned student contributions, with standardized coefficients between 0.08 and 0.18."
Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005 - the article this story is about — the whole article — the passage: "Total preparation time with AI predicted discourse outcomes on its own and lost that value for the lesson-level indicators once layer composition was entered."
Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005 - the article this story is about — the whole article — the passage: "An embedded micro-randomized comparison, in which teachers drafted their own questioning and variation sequence before consulting the model, lowered generative-layer offloading and raised substantive uptake."
Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005 - the article this story is about — the whole article — the passage: "Lessons prepared by taking a model draft first and revising it (n = 259) had mean generative-layer offloading of 0.591, higher-order questioning of 0.189, and substantive uptake of 0.211. Lessons prepared by drafting first and consulting the model afterward (n = 403) had mean generative-layer offloading of 0.210, higher-order questioning of 0.264, and substantive uptake of 0.335."
Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005 - the article this story is about — the whole article — the passage: "Once both offloading layers were entered, confidence predicted no outcome (all p > 0.12), so what little it tracked was inherited from the delegation that produced it."
Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005 - the article this story is about — the whole article — the passage: "The indirect path was −0.038 with a Monte Carlo 95% confidence interval of [−0.065, −0.015], and the remaining direct path was −0.005 (p = 0.904)"
Chen, H., Wang, N., Chen, C. et al. (2026). The layer of lesson preparation delegated to AI predicts classroom discourse quality: a four-wave video study of 90 junior secondary mathematics teachers. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1962005 - the article this story is about — the whole article — the passage: "The sample comes from one prefecture-level city, one country, one subject, and one phase of schooling, and the study was not preregistered."