experiment · International Journal of Educational Technology in Higher Education · la publicación, 1 ene 2026 · gratis
"Sentí que estaba tomando prestada su estructura en lugar de poner a prueba mi propio razonamiento"
Un ensayo con 1,176 estudiantes de primer año de ciencia comparó cuatro formas de dar retroalimentación. Pedir al alumno que juzgue su borrador antes de ver los comentarios de la IA rindió más que la IA sola. El estudio no puede separar ese orden del tiempo extra que exigió.
Versión breve · la versión detallada sigue, unos 6 min
Pregunte al chat de weeklyAI
Pregúnteme sobre la IA en su vida: le señalo las notas que weeklyAI ha publicado sobre el tema, esta semana y antes.
Las conversaciones se guardan mientras exista weeklyAI, para mejorar la publicación. Se responde en el idioma en que usted escribe.
- El estudio, de un vistazo
- Quiénes
- Estudiantes de primer año de ciencia (biología, química y física)
- Cuántos
- 1,176 estudiantes de primer año en 48 secciones de curso
- Dónde
- Cuatro universidades que el artículo no nombra
- Cuándo
- No lo dice el pasaje
- Tipo de estudio
- experiment
- Quién lo hizo
- No lo dicen los pasajes
- El límite que importa
- Los grupos con autoevaluación hicieron más trabajo guiado y los grupos con IA recibieron capacitación extra: no se puede separar.
Cuatro formas de dar retroalimentación sobre un texto científico
La crítica directa de la IA mejoró más la calidad del argumento de inmediato
El grupo que se autoevaluó primero rindió mejor en aprovechar los comentarios, regular su revisión, aprender conceptos y en la tarea posterior sin IA
El combinado rindió mejor en aprovechar los comentarios, regular su revisión, aprender conceptos y en la tarea posterior sin IA
El combinado tuvo el promedio más alto de mejora inmediata del argumento; en la tarea sin IA ambos superaron a la IA directa
El orden de los pasos cambió lo que quedó aprendido.
Un estudiante lo describió así. La herramienta le mostró casi de inmediato lo que faltaba en su texto. Pero a veces sentía que estaba tomando prestada la estructura de la máquina en lugar de poner a prueba su propio razonamiento. El estudio toma esa frase de un alumno del grupo que recibió los comentarios de la IA de forma directa.
El estudio repartió al azar 48 secciones de cursos de biología, química y física, en cuatro universidades, entre cuatro diseños. Un grupo solo recibió comentarios de compañeros. Otro recibió comentarios de una IA al instante. Un tercero se autoevaluó primero y después leyó a la IA. El cuarto combinó autoevaluación, comentarios de dos compañeros y crítica de la IA, y cerró con una breve nota explicando qué sugerencias aceptó, rechazó o cambió. Eran 1,176 estudiantes de primer año.
La IA directa subió más la calidad del argumento de inmediato que los compañeros. En una tarea posterior, escrita sin ninguna IA, no superó a los compañeros. Ahí los otros dos grupos sí sacaron ventaja.
El estudio atribuye parte de esa ventaja a que esos alumnos aprovecharon mejor los comentarios y regularon mejor su propia revisión. Quienes se autoevaluaron primero y quienes usaron el diseño combinado obtuvieron mejores resultados en esa tarea sin IA. Entre esos dos grupos no hubo diferencia clara.
Hay algo que el estudio no puede separar. Los grupos de autoevaluación y combinado hicieron más trabajo guiado de juicio, y los grupos con IA recibieron además capacitación extra sobre el sistema. Parte de la diferencia pudo venir de ese tiempo y esa preparación, no solo del orden de los pasos.
Tampoco dice nada sobre otros contextos. Fueron cursos de primer año de ciencia en cuatro universidades que el artículo no nombra. No muestra si el mismo patrón aparece en secundaria, en otras materias ni en otros países.
Lo que sí deja es una pista concreta. Busque rutinas de retroalimentación que pidan al alumno nombrar las debilidades de su propio borrador antes de que aparezca cualquier comentario de la IA, y que después le pidan decir qué sugerencias tomó, cuáles dejó y cuáles modificó.
Qué significa para usted
Si en su curso le piden revisar un texto con ayuda de una herramienta que responde por escrito, conviene mirar si antes le piden anotar qué le falta a su propio borrador. En el ensayo, quienes hicieron ese paso previo y luego explicaron qué sugerencias tomaron escribieron mejores argumentos en una tarea posterior sin ninguna ayuda. Vale preguntar si ese orden forma parte de la actividad, y no dar por hecho que el tiempo extra de preparación no influyó.
Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9
Quién pagó: El artículo indica que la investigación no recibió ninguna subvención específica de organismos de financiación públicos, comerciales o sin ánimo de lucro, y no menciona préstamos de equipos o programas más allá de la interfaz de IA alojada institucionalmente por las propias universidades.
Versión detalladaLos pasajes copiados del artículo, las ilustraciones y cada fuente con cuánto leímos de ella · 1174 palabras · unos 6 minLeerla →Cerrar
La crítica de la IA ayudó más cuando el alumno juzgó primero su propio borrador
Un experimento con casi 1,200 estudiantes universitarios comparó cuatro maneras de dar retroalimentación. El orden de los pasos cambió lo que quedó aprendido.
Imagine usted a un estudiante de primer año de biología con un borrador de 400 palabras sobre un experimento. Tiene que decir qué sostiene, con qué evidencia y por qué. Alguien le devuelve comentarios y él reescribe. La pregunta del estudio es quién da esos comentarios y en qué orden. En cuatro universidades que el artículo no nombra, mil 176 estudiantes de primero de carrera, repartidos en 48 grupos de biología, química y física, siguieron uno de cuatro caminos durante diez semanas y tres rondas de escritura1.
Los cuatro caminos fueron estos: solo compañeros; la crítica de una inteligencia artificial generativa sin más; primero una autoevaluación escrita y después la crítica de la máquina; o el híbrido, que suma autoevaluación, dos compañeros, la crítica de la máquina y una nota final donde el alumno explica qué aceptó y qué rechazó2. En el camino reflexivo, el estudiante escribe antes qué le parece débil de su texto3. La herramienta era un modelo de lenguaje alojado por las propias universidades, con instrucciones fijas para señalar fortalezas y debilidades, marcar evidencia faltante o débil, cuestionar afirmaciones sin respaldo, sugerir cambios y hacer preguntas, sin reescribir el texto por el alumno.
La ganancia promedio en la calidad de la argumentación, del primer borrador a la versión final, fue distinta según el grupo. La ganancia promedio fue de 2.84 puntos sobre 16 en el grupo de compañeros, 3.41 con la crítica directa de la máquina, 4.07 en el reflexivo y 4.31 en el híbrido. Frente a la crítica directa, el diseño reflexivo sacó 0.66 puntos más y el híbrido 0.90 más4. En la tarea de transferencia, hecha al final del estudio y sin ninguna ayuda de IA, el grupo de crítica directa no superó al de compañeros, mientras que el reflexivo y el híbrido sí quedaron por encima de ambos5.
El detalle que más pesa está en el orden. Los estudiantes que primero escribieron su propio diagnóstico dijeron después, en entrevistas, que ya no aceptaban los comentarios de la máquina como respuestas: comparaban, verificaban y decidían6. En el grupo híbrido hubo quien lo dijo con claridad incómoda: ahorró tiempo, pero tuvo que preguntarse si el texto seguía siendo su explicación o solo estaba puliendo la versión que la herramienta parecía preferir7. Las cuentas del estudio sugieren que parte de la ventaja de los diseños más activos pasa por cuánto se apropia el alumno de la crítica y cuánto regula su propia revisión8.
Conviene mirar los límites antes de llevarse la receta a otro lado. Los grupos reflexivo e híbrido hicieron más trabajo guiado de evaluación, así que el estudio no puede separar el efecto del orden de los pasos del tiempo y el esfuerzo extra que exigieron9. Los estudiantes con IA recibieron además una capacitación sobre el sistema que los demás no tuvieron. Eran cursos de primero de universidad en ciencias, y no se sabe si el patrón se sostiene en secundaria, en otras materias o en otros países10. La calidad de los comentarios entre compañeros no se calificó por separado.
Piense en lo que hace un profesor cuando revisa un trabajo. No aplica una regla mecánica: sopesa el caso, distingue entre cumplir el trámite y respetar el criterio de quien aprende. Lo curioso es que la herramienta hace lo contrario: responde rápido y con seguridad, sin dudar. El estudio no midió si esa seguridad cambia la forma en que los docentes califican. Y ahí hay una escena que se repite en muchas aulas: el alumno entrega, el profesor pregunta qué cambió y por qué, y el alumno contesta "lo que me dijo la IA".
Otros trabajos que leímos van en la misma dirección, y conviene decirlo con cuidado porque de algunos solo pudimos leer el resumen. Un ensayo aleatorizado con 362 estudiantes de enfermería comparó un tutor digital con IA, una enseñanza mixta y la clase tradicional durante 16 semanas; el grupo con IA obtuvo mejores calificaciones y mayor retención111213. Solo leímos el resumen; el texto completo no estaba disponible.
Otro estudio revisó la calificación automática de exámenes de enfermería: en preguntas de opción múltiple el acuerdo con los profesores fue alto, pero en respuestas escritas el sistema sobreestimó las notas y aprobó a 40 de 59 exámenes que los profesores habían reprobado141516. La conclusión de sus autores es llana: la revisión del profesor sigue siendo necesaria cuando la decisión importa17.
Un tercer trabajo, con 310 universitarios que aprendían inglés, comparó un flujo de trabajo en el que el alumno escribe primero, luego recibe la crítica de la máquina y al final evalúa él mismo; el grupo con IA superó al control y sus diarios mostraron más planeación y seguimiento, aunque también advirtieron el riesgo de delegar el pensamiento si no hay reglas claras181920. También aquí leímos solo el resumen.
Así lo leemos nosotros. El patrón que se repite es sencillo: cuando la herramienta resuelve la duda antes de que hayamos formado nuestra propia opinión, la costumbre de consultarla puede ir sustituyendo el hábito de juzgar por cuenta propia, sin que nadie lo decida. En su casa o en su salón eso se vería así: el alumno entrega mejores borradores y, aun así, en la siguiente tarea abre la herramienta antes de pensar. Sabríamos que nos equivocamos si esos estudiantes consultan la IA menos o igual que los demás y si sus borradores posteriores muestran el mismo criterio propio. Lo que usted puede hacer con esto es pequeño y concreto: antes de abrir la herramienta, pida que escriba en dos o tres líneas qué le parece débil de su texto. Después, que compare: en qué coincidió, en qué no, y qué decidió por su cuenta.
También vale para quien enseña. Si la herramienta se presenta como la voz que sabe y el estudiante solo acata, la revisión se vuelve más uniforme y menos discutida; si se presenta como una opinión más entre varias, el estudiante puede defender sus decisiones y hasta rechazar sugerencias. Una manera de saberlo: al revisar un trabajo, pregunte al estudiante qué cambió y por qué antes de mirar lo que dijo la máquina. Anote en qué casos acertó la herramienta y en qué casos el juicio del profesor fue mejor.
Lo que el estudio propone, en palabras de sus autores, es que la crítica de la IA vaya precedida de una autoevaluación breve y estructurada: fortalezas, debilidades, dudas y prioridades de revisión21. No hay una cifra de cuánto tiempo debe durar ese paso, ni cuántas preguntas bastan. Tampoco hay evidencia de que funcione igual en un salón de secundaria, en una clase de historia o en una universidad de otro país. Esa parte queda por hacer.
Ese borrador de 400 palabras sigue ahí, con sus tachones. Lo que cambió fue quién juzgó primero. ¿Qué le pediría usted a un estudiante que escriba antes de abrir la herramienta?
De dónde sale cada dato de contexto, y cuánto leímos de cada documento
- Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The analytic sample comprised 1,176 first-year undergraduate students from 48 course sections across four universities and three science domains."
- Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The study compares four feedback designs—peer feedback only, direct GenAI-supported feedback, reflective GenAI-supported feedback, and a hybrid design combining self-evaluation, peer feedback, and GenAI critique"
- Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "In reflective GenAI-supported feedback, students first evaluate their own work against explicit criteria or identify areas of uncertainty before engaging with AI-generated critique."
- Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The hybrid condition yielded the highest adjusted mean for immediate argument-quality gain, whereas both reflective and hybrid conditions outperformed direct GenAI-supported feedback on delayed AI-free transfer."
- Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Direct GenAI-supported feedback improved immediate argument-quality gain relative to peer feedback, but reflective and hybrid designs produced stronger outcomes on feedback uptake, self-regulated learning, conceptual learning, and delayed AI-free transfer."
- Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Because I had already written down what I thought was weak, I was not just accepting the AI’s comments. I was checking where it agreed with me and where it made me rethink the argument."
- Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "It definitely saved time, but I had to stop and ask whether this was still my explanation or whether I was just polishing the version the tool seemed to prefer."
- Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Multilevel mediation analyses indicated that feedback uptake and self-regulated learning partially explained these advantages."
- Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "First, the reflective and hybrid conditions involved greater amounts of guided evaluative activity, meaning that feedback-source configuration and reflective workload were not fully separable."
- Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "Third, the study was conducted in first-year university science courses, and the extent to which the findings generalize to other disciplines, levels of study, or institutional contexts remains uncertain."
- Zhao C, Zhu J, Liu J, Zhao W, Pang Y. (2026). Effectiveness of a generative AI-powered digital tutor integrated with a knowledge graph in anatomy education for nursing students: a randomized controlled trial. BMC Medical Education. 10.1186/s12909-026-09469-0 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "A three-arm parallel randomized controlled trial was conducted among 362 first-year nursing students at Cangzhou Medical College, of whom 301 completed the study."
- Zhao C, Zhu J, Liu J, Zhao W, Pang Y. (2026). Effectiveness of a generative AI-powered digital tutor integrated with a knowledge graph in anatomy education for nursing students: a randomized controlled trial. BMC Medical Education. 10.1186/s12909-026-09469-0 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "The AI-Enhanced Group achieved higher module scores (82.5 ± 7.3) compared with the Blended Group (77.6 ± 8.1) and Traditional Teaching Group (74.1 ± 8.5) (F = 28.64, p < 0.001)."
- Zhao C, Zhu J, Liu J, Zhao W, Pang Y. (2026). Effectiveness of a generative AI-powered digital tutor integrated with a knowledge graph in anatomy education for nursing students: a randomized controlled trial. BMC Medical Education. 10.1186/s12909-026-09469-0 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "Furthermore, the AI-Enhanced Group demonstrated higher knowledge retention rates (82.3%) and case-based inference accuracy (91.2%) than the comparison groups."
- Feng Q, Liu M, Tang T, Zhang L, Li H. (2026). Agreement and Bias in ChatGPT-Assisted Grading of Undergraduate Nursing Examinations: A Method-Comparison Study. Nurse Educator. 10.1097/NNE.0000000000002372 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "Multiple-choice grading showed minimal bias (-0.05 points) and excellent agreement with faculty grading (intraclass correlation coefficient = 0.949)."
- Feng Q, Liu M, Tang T, Zhang L, Li H. (2026). Agreement and Bias in ChatGPT-Assisted Grading of Undergraduate Nursing Examinations: A Method-Comparison Study. Nurse Educator. 10.1097/NNE.0000000000002372 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "ChatGPT overestimated the constructed-response subtotal and total scores by 7.48 and 7.43 points."
- Feng Q, Liu M, Tang T, Zhang L, Li H. (2026). Agreement and Bias in ChatGPT-Assisted Grading of Undergraduate Nursing Examinations: A Method-Comparison Study. Nurse Educator. 10.1097/NNE.0000000000002372 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "Of the 59 examinations with faculty total scores below the institutional passing threshold of 60 points, ChatGPT classified 40 as passing."
- Feng Q, Liu M, Tang T, Zhang L, Li H. (2026). Agreement and Bias in ChatGPT-Assisted Grading of Undergraduate Nursing Examinations: A Method-Comparison Study. Nurse Educator. 10.1097/NNE.0000000000002372 — solo el resumen - el texto completo está tras una suscripción — el pasaje: "Faculty review remains necessary for consequential assessment decisions."
- Chen J, Alibakhshi G. (2025). AI ‐Powered Applications’ Effects on English Language Learners’ Cognitive, Metacognitive, and Resource Management Strategies, and Language Achievement. Journal of Computer Assisted Learning. 10.1002/jcal.70171 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "The AI group completed a 6-h AI-literacy workshop and 12 weeks of guided practice using ChatGPT, Poe, and Bard within a draft-first [right arrow] AI critique [right arrow] human evaluation workflow; the control group received the same curriculum without AI."
- Chen J, Alibakhshi G. (2025). AI ‐Powered Applications’ Effects on English Language Learners’ Cognitive, Metacognitive, and Resource Management Strategies, and Language Achievement. Journal of Computer Assisted Learning. 10.1002/jcal.70171 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "The AI group outperformed the control across all four outcomes; a MANCOVA indicated a significant multivariate effect, with follow-up ANCOVAs showing partial [eta-squared] values of approximately 0.16-0.22."
- Chen J, Alibakhshi G. (2025). AI ‐Powered Applications’ Effects on English Language Learners’ Cognitive, Metacognitive, and Resource Management Strategies, and Language Achievement. Journal of Computer Assisted Learning. 10.1002/jcal.70171 — solo el resumen - no se pudo obtener el texto completo — el pasaje: "The journals illuminated mechanisms--planning/monitoring, time efficiency, confidence, and calibrated help-seeking--and flagged the risks of cognitive offloading without explicit guardrails."
- Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9 - el artículo del que trata esta nota — el artículo completo — el pasaje: "The results instead suggest that AI critique should be preceded by a brief, structured self-evaluation that asks students to identify strengths, weaknesses, uncertainties, and revision priorities before viewing external feedback."
Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9
Quién pagó: El artículo indica que la investigación no recibió ninguna subvención específica de organismos de financiación públicos, comerciales o sin ánimo de lucro, y no menciona préstamos de equipos o programas más allá de la interfaz de IA alojada institucionalmente por las propias universidades.
experiment · International Journal of Educational Technology in Higher Education · the paper, 1 Jan 2026 · free
Judge Your Own Draft First, Then Ask the AI
A trial with 1,176 first-year science students found that AI feedback helped most when students had already named their own draft's weaknesses. The study cannot fully separate the order of steps from the extra guided judging those groups did, and the AI groups also got extra training on the tool.
Short version · the longer version follows, about 6 min
Ask the weeklyAI chat
Ask me about AI in your life: I point you to the stories weeklyAI has published on it, this week and before.
Conversations are saved for as long as weeklyAI exists, to improve the publication. Answers come in the language you write in.
- The study at a glance
- Who
- First-year university students in biology, chemistry and physics courses
- How many
- 1,176 first-year undergraduate students from 48 course sections
- Where
- Four universities (country not stated in the passages)
- When
- Not stated in the passages
- Kind of study
- experiment
- Who did it
- Huseyin Ates (institution not stated in the passages)
- The limit that matters
- The extra guided judging and extra AI training cannot be separated from the order of steps
Four ways of getting feedback on a scientific argument
The AI group improved their arguments more on the immediate rewrite, by a small margin
The self-judging group did better on taking in feedback, steering their own revision, understanding concepts, and writing a new argument with no AI
The combined group did better on taking in feedback, steering their own revision, understanding concepts, and writing a new argument with no AI
On the later task written with no AI, the AI group did no better than the classmates-only group
When the learner's own judgment comes before the machine's, the machine becomes a second opinion.

"It helped me see what was missing almost immediately, but sometimes it felt as though I was borrowing its structure instead of testing my own reasoning." That is a first-year biology student describing what happened when an AI read the draft and told the student what to fix.
Behind that sentence sits a pattern. Researchers ran a trial with 1,176 first-year students in biology, chemistry and physics courses at four universities. The students worked in 48 class sections, and each section was assigned by chance to one of four ways of getting feedback on a scientific argument.
One group swapped drafts with two classmates only. A second uploaded the draft and read AI comments straight away. A third wrote down their own view of the draft's strengths, weaknesses and uncertainties before the AI spoke. A fourth did all of it: their own evaluation, two classmates' comments, the AI's critique, and a short memo on which suggestions they took, rejected or changed.
Speed had its uses. The direct-AI group improved the quality of their arguments more than the peer-only group did, by a small margin.
The later test told a different story. Students wrote a new argument with no AI available. There, the direct-AI group did no better than the peer-only group. The reflective and hybrid groups beat both. The study's analysis ties part of that advantage to how much feedback students took in and how deliberately they steered their own revision.
Consider what the two stronger groups also had. They did more guided judging of their own work, and the AI groups got extra training on the tool. The study cannot say the order of steps alone produced the gain, and a quick AI prompt may not reproduce it.
The trial covered first-year university science courses only. It says nothing yet about other subjects, other levels of study or other kinds of university.
So watch for a routine, in any course, that asks a learner to name their own draft's weak points before any AI comments appear, and to say afterwards which suggestions they kept, dropped or reworked.
What this means for you
The trial covered first-year science courses only, and the extra judging those groups did cannot be separated from the order of steps. What a reader can watch for in any course is a routine that asks students to name their draft's weak points before the AI comments appear. Ask whether the feedback is built that way, and treat a quick AI prompt as no proof that the learning will last.
Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9
Who paid: The article states that the research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors, and it names no equipment or software lenders beyond the universities' own institutionally hosted AI interface.
The longer versionThe passages copied from the paper, the pictures, and every source with how much of it we read · 1116 words · about 6 minRead it →Close
Judge Your Own Draft First: The Step That Made AI Feedback Teach More
A four-university experiment found that students who wrote down their own critique before opening the machine's comments learned more — and kept more of it when the AI was gone.
Four ways of handling feedback, one written assignment, and four different results: that is the shape of a new experiment on generative AI in first-year university science courses. Students in one group received only comments from classmates. A second group uploaded a draft and read what the AI said. A third wrote down their own judgment of the draft before the AI spoke. A fourth did all of it — self-evaluation, two peers, the AI — and finished with a short memo explaining which suggestions they took, which they refused, and why.
The students were 1,176 first-year undergraduates in 48 course sections at four universities, studying biology, chemistry or physics. Whole sections, not individual students, were assigned to one of the four designs, and the comparison ran across ten weeks and three writing cycles. On the immediate rewrite, the group that judged its own draft first and the group that did everything both beat the group that went straight to the machine: their average gains from draft to revision were 4.07 and 4.31 against 3.41, while peer-only feedback came in at 2.84. The gap between the two self-judging designs was small and not statistically reliable.
The more telling result came later. On a new argumentation task, written with no AI available, students in the reflective and hybrid groups still outperformed the direct-AI group, and neither showed an advantage over the other. Direct AI feedback did beat peer-only feedback on the immediate rewrite — it was not harmful — but it did not beat peer-only feedback on the delayed, AI-free task. The study's process measures point to where the difference came from: students in the self-judging designs engaged more with feedback and regulated their revision more, and those two things partly explained their better results.
What the study cannot separate is time and effort from sequence. The reflective and hybrid conditions asked for more guided evaluative work, so the extra minutes students spent judging their own drafts are tangled with the order of steps. Students in the AI conditions also received additional training on the system, which may have added preparation time the peer-only group did not get. Peer-comment quality was not scored separately. The setting was first-year science courses at four universities, and the author notes that generalization to other disciplines, year levels or institutions remains uncertain.
Consider what the reflective group actually did, because the mechanism is the point. Before they saw any AI critique, they wrote down what they thought was weak in their draft and where they were unsure1. Then they read the machine's comments against that written judgment. One chemistry student put it plainly: "Because I had already written down what I thought was weak, I was not just accepting the AI's comments. I was checking where it agreed with me and where it made me rethink the argument"2. That is comparison, not compliance — the difference between using a critique and absorbing it.
The direct group's experience was different in kind. Students described the AI as fast and often useful, but some felt steered: "It definitely saved time, but I had to stop and ask whether this was still my explanation or whether I was just polishing the version the tool seemed to prefer"3. The author's warning about this design is not that it fails — it improved drafts — but that it may reduce the need for students to make their own judgments if the process encourages passive uptake4. A tool that answers the question "what should I change?" can quietly take over the question "is this good?"
This study sits in a family of work asking not whether AI helps but how it is arranged. A randomized trial with 310 undergraduates in an English-language program, which we could read only in its summary, tested a draft-first, AI-critique, human-evaluation workflow and found the AI group ahead on all four outcomes; its authors flagged the risk of cognitive offloading without explicit guardrails567. A separate trial in nursing education, also summary-only, found a generative AI tutor blended with a knowledge graph improved module and final scores over traditional teaching89. And a grading study, again summary-only, found the same kind of system scored multiple-choice reliably but overestimated written answers by 7.48 points and passed 40 of 59 students who had actually failed — a reminder that the machine's judgment is not the standard1011.
Here is how we read it. The pattern in this study is not about which tool is better; it is about who goes first. When the learner's own judgment comes before the machine's, the machine becomes a second opinion. When it comes after, the machine becomes the opinion. We expect this to show up in homes and classrooms far from first-year physics: the student who opens the chatbot before thinking will produce a cleaner draft and carry less away from it, and the gap will be widest on the exam or essay where no chatbot is allowed. You would know we are wrong if students who skipped the self-judgment step did just as well on the later, AI-free task — the study found they did not, but a single experiment is not the last word.
The practical move is small enough to try this week. Before anyone opens AI feedback on a piece of writing, ask them for three lines: what I think is weak here, and why. Then have them read the machine's comments and mark which ones they agree with, which they reject, and which changed their mind. The order of steps is the intervention; the license is not. The study's own recommendation points the same way — a brief structured self-evaluation before viewing external feedback12 — and it costs nothing but a few minutes and the willingness to be wrong on paper first.
What the experiment leaves open is whether a shorter version of that first step would work as well, and whether the advantage holds across a full semester rather than three cycles. The author also notes the sample may reflect consent-based selection, since class participation was required but research participation was voluntary. Those are questions for the next study, not reasons to doubt what this one measured.
The draft on the screen looks better after the machine's comments. The question worth keeping is the one the reflective students asked themselves before they read them: what did I already think was wrong? Ask it first, and the feedback has something to argue with.
Where each piece of context comes from, and how much of it we read
- Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9 - the article this story is about — the whole article — the passage: "In reflective GenAI-supported feedback, students first evaluate their own work against explicit criteria or identify areas of uncertainty before engaging with AI-generated critique."
- Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9 - the article this story is about — the whole article — the passage: "Because I had already written down what I thought was weak, I was not just accepting the AI’s comments. I was checking where it agreed with me and where it made me rethink the argument."
- Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9 - the article this story is about — the whole article — the passage: "It definitely saved time, but I had to stop and ask whether this was still my explanation or whether I was just polishing the version the tool seemed to prefer."
- Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9 - the article this story is about — the whole article — the passage: "GenAI-supported feedback is pedagogically promising but also potentially risky: it may help students improve written products and, when designed dialogically, support deeper reasoning, yet it may also reduce the need for students to make their own evaluative judgments if the feedback process encourages passive uptake"
- Chen J, Alibakhshi G. (2025). AI ‐Powered Applications’ Effects on English Language Learners’ Cognitive, Metacognitive, and Resource Management Strategies, and Language Achievement. Journal of Computer Assisted Learning. 10.1002/jcal.70171 — only the abstract - the full text could not be fetched — the passage: "The AI group completed a 6-h AI-literacy workshop and 12 weeks of guided practice using ChatGPT, Poe, and Bard within a draft-first [right arrow] AI critique [right arrow] human evaluation workflow; the control group received the same curriculum without AI."
- Chen J, Alibakhshi G. (2025). AI ‐Powered Applications’ Effects on English Language Learners’ Cognitive, Metacognitive, and Resource Management Strategies, and Language Achievement. Journal of Computer Assisted Learning. 10.1002/jcal.70171 — only the abstract - the full text could not be fetched — the passage: "The AI group outperformed the control across all four outcomes; a MANCOVA indicated a significant multivariate effect, with follow-up ANCOVAs showing partial [eta-squared] values of approximately 0.16-0.22."
- Chen J, Alibakhshi G. (2025). AI ‐Powered Applications’ Effects on English Language Learners’ Cognitive, Metacognitive, and Resource Management Strategies, and Language Achievement. Journal of Computer Assisted Learning. 10.1002/jcal.70171 — only the abstract - the full text could not be fetched — the passage: "The journals illuminated mechanisms--planning/monitoring, time efficiency, confidence, and calibrated help-seeking--and flagged the risks of cognitive offloading without explicit guardrails."
- Zhao C, Zhu J, Liu J, Zhao W, Pang Y. (2026). Effectiveness of a generative AI-powered digital tutor integrated with a knowledge graph in anatomy education for nursing students: a randomized controlled trial. BMC Medical Education. 10.1186/s12909-026-09469-0 — only the abstract - the full text could not be fetched — the passage: "The AI-Enhanced Group achieved higher module scores (82.5 ± 7.3) compared with the Blended Group (77.6 ± 8.1) and Traditional Teaching Group (74.1 ± 8.5) (F = 28.64, p < 0.001)."
- Zhao C, Zhu J, Liu J, Zhao W, Pang Y. (2026). Effectiveness of a generative AI-powered digital tutor integrated with a knowledge graph in anatomy education for nursing students: a randomized controlled trial. BMC Medical Education. 10.1186/s12909-026-09469-0 — only the abstract - the full text could not be fetched — the passage: "Final comprehensive scores were also significantly higher in the AI-Enhanced Group (84.2 ± 6.8) (F = 32.16, p < 0.001)."
- Feng Q, Liu M, Tang T, Zhang L, Li H. (2026). Agreement and Bias in ChatGPT-Assisted Grading of Undergraduate Nursing Examinations: A Method-Comparison Study. Nurse Educator. 10.1097/NNE.0000000000002372 — only the abstract - the full paper is behind a subscription — the passage: "ChatGPT overestimated the constructed-response subtotal and total scores by 7.48 and 7.43 points."
- Feng Q, Liu M, Tang T, Zhang L, Li H. (2026). Agreement and Bias in ChatGPT-Assisted Grading of Undergraduate Nursing Examinations: A Method-Comparison Study. Nurse Educator. 10.1097/NNE.0000000000002372 — only the abstract - the full paper is behind a subscription — the passage: "Of the 59 examinations with faculty total scores below the institutional passing threshold of 60 points, ChatGPT classified 40 as passing."
- Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9 - the article this story is about — the whole article — the passage: "The results instead suggest that AI critique should be preceded by a brief, structured self-evaluation that asks students to identify strengths, weaknesses, uncertainties, and revision priorities before viewing external feedback."
Ates, H. (2026). Human-centered GenAI feedback design in higher education: a multisite experiment on direct, reflective, and hybrid approaches to scientific argumentation. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00614-9
Who paid: The article states that the research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors, and it names no equipment or software lenders beyond the universities' own institutionally hosted AI interface.
