https://diarioiaeducativa.github.io/en/2026-11-21/ia-aprender-entender-o-entregar.html
The Sunday feature 22 November
What the research says
AI helps learning more when it is used to understand than when it does the task
A review by Stanford University now brings together 52 studies measuring the effect of AI on learning: using it to understand new material fares better than using it to produce the work, although no use improves results consistently. This report explains how to read those studies, why a tutor that asks questions is not enough on its own and what a teacher can do so that AI does not do the students’ work.
7 minute read
The Stanford review admits only causal studies
Stanford University’s SCALE initiative published on 10 November 2026 “The Evidence Base on AI in K-12: A November 2026 Review”, signed by Seth Walker, Lily Fesler, JP Martinez Claeys, Chris Agnew and Susanna Loeb. It admits only causal studies: trials that randomly assign participants to a group that uses AI and another that does not, or similar designs that allow the difference in results to be attributed to the tool. With that filter, 32 new studies made the cut, added to the 20 in the March version: 52 in all. The paper reported the review on 19 November.
There are still few, and the review itself explains why. Most of the new ones were carried out at university: only five concern pupils in US schools, in the stages the country calls K-12, from nursery to the end of secondary school. Many are preprints, studies circulated without peer review, and in most of them the effects were measured between a few days and a week afterwards. The authors warn that their conclusions cannot be generalised to all schools in the United States. It is funded by the Walton Family Foundation, the Overdeck Family Foundation and Google.org, Google’s philanthropic arm.
In ChatGPT-IA-edu it was shared on 18 November with a summary of its conclusions and two of its cases, the one from Turkey, involving teachers and the one from Sierra Leone; the authors of this paper took part in that exchange. Faced with a headline along the lines of “AI improves learning”, the review helps to ask three questions: compared with what, measured with the tool at hand or without it, and how long afterwards.
With AI practice scores rise and exam scores fall
Stanford’s central conclusion is that AI seems to work better for learning when students use it as support for understanding than when they use it to produce the work, although neither use improves learning consistently. Its summary puts it this way: “the effects of generative AI depend on how it is designed and used”.
One study in the March version shows it clearly. A team led by Hamsa Bastani, of the University of Pennsylvania, published in June 2025 in the journal PNAS “Generative AI without guardrails can harm learning”: almost 1,000 secondary school students in Turkey practised mathematics with two GPT-4-based assistants, one that imitated ChatGPT and another with instructions designed to protect learning. During practice, marks improved by 48% with the first and by 127% with the second. Without the tool, the students who had used the first scored 17% lower than those who had never had it; with the second, that harm was largely reduced.
The new studies point the same way. In Brazil, 85 university students prepared a presentation with ChatGPT or with traditional resources: the former spent less time but, in a test 45 days later, scored lower (Barcaui, 2025).
The review adds a warning: AI can improve students’ experience even if it does not improve what they learn. In several studies, students rated AI feedback as clearer or more motivating than that from other sources, with similar learning outcomes. In other words, students liking an activity with AI does not prove they learn more. The OECD reaches something similar in its Digital Education Outlook 2026 (PDF): according to the summary on the European Digital Skills and Jobs Platform, with general-purpose chatbots students hand in better work, but the advantage disappears or is reversed in exams without them.
A tutor that asks questions helps when it is part of the activity
If producing the work with AI does harm, one way out is for AI not to give the answer. Michael W. Asher, Meng Cao and Paulo F. Carvalho, of Carnegie Mellon University, published on 15 November 2026 “Turning Errors into Successes”, an experiment with 400 adults in the United States who were learning to set up equations. After each mistake, some received the correct answer or an explanation, while others chatted with an AI tutor that guided them with questions to the solution. With the tutor they corrected 64% of their errors themselves (131 of 206) and scored 9 percentage points higher in the final test than with the explanations. The paper reported the experiment on 19 November.
In the group it was presented as an argument that AI works very well as a Socratic tutor and that this use deserves to be pushed, and its entry went into the digital Library; an author of this paper also took part here. The authors themselves acknowledge the limits: it is a preprint, with adults, in a single session of about twenty minutes, and the advantage appeared only on problems similar to those practised.
Other studies in the Stanford review, summarised in the table, call for some nuance: with 199 university students, an immediate answer worked better than step-by-step hints. In Sierra Leone, by contrast, 1,423 students improved in mathematics when their teachers built Gemini’s guided learning into half of their lessons for eight weeks; it was written up by a Google team and the organisation Fab AI.
These results suggest that the design of the activity matters more than the format of the tutor: asking instead of answering helps when it is part of a task designed by the teacher. In the groups the subject has a history: in June 2025, Juan José de Haro shared an AI Socratic tutor.
| Study | Participants | What was compared | Result |
|---|---|---|---|
| Bastani et al. (2025, PNAS) | Almost 1,000 secondary school students, Turkey | ChatGPT-style assistant and assistant with guardrails, in mathematics practice | Without the tool, the first group scored 17% lower than the group that never had it; the guardrails reduced the harm |
| Asher, Cao and Carvalho (2026, preprint) | 400 adults, United States | Tutor that asks questions after the error and explanations | Better results with the tutor on problems like those practised |
| Cao et al. (2026, preprint) | 199 university students, United States | Immediate answer and step-by-step hints | Better results with the immediate answer |
| Oreopoulos et al. (2026) | 6,997 students in grades 6 to 8, United States | Mathematics practice with and without an AI tutor | Slower practice, no improvement in what was learned |
| Kao et al. (2025, preprint) | 90 students in grade 10, Taiwan | Socratic questions from ChatGPT’s study mode | Better argumentation and critical thinking |
| LearnLM Team, Google and Fab AI (2026) | 1,423 students in grades 7 and 8, Sierra Leone | Gemini’s guided learning built in by the teacher and usual lessons | Improvement in mathematics |
Source: “The Evidence Base on AI in K-12: A November 2026 Review”, SCALE, Stanford University, 10 November 2026; the study by Bastani et al. appears in the March version. Grades are those of each country’s system. A preprint is a study circulated without peer review.
Clark and Kozma already argued over whether learning depends on the medium
In the group, the Stanford results were read as the current version of an old debate in educational technology. Richard E. Clark argued in “Reconsidering Research on Learning from Media” (1983) that media merely deliver instruction, and repeated it in “Media will never influence learning” (1994): what makes people learn is the method. Robert B. Kozma replied in “Will media influence learning? Reframing the debate” (1994) that medium and method cannot be separated, because what each technology makes possible interacts with the way a person learns. The paper covered that debate on 20 November.
The same voice that raised the debate acknowledged that the group leans towards Kozma and asked people not to lose sight of Clark’s arguments. The studies give grounds to both positions. For Clark, in Bastani’s experiment the two assistants used the same model, and what changed the result was the instructions, that is, the method. For Kozma, a tutor that adapts its questions to each error does something an exercise book cannot do. Both readings agree on what matters to a teacher: the tool, without a well-designed activity, guarantees nothing.
The EU, INTEF and Galicia set conditions on the use of AI
The Council of the European Union, which brings together the governments of the member states, approved conclusions on 19 November 2026 on citizens’ AI skills. Its text includes the ability to recognise when AI should not be used and calls for attention to risks such as over-reliance, loss of learning and cognitive offloading. These are policy guidelines: they oblige no school and no teacher. The paper covered them on 20 November.
In Spain, the state reference document is the Guide to the use of artificial intelligence in education (in Spanish) from the National Institute of Educational Technologies and Teacher Training (INTEF), version 2.0, of September 2026. It starts from the position that “the central question should not be, therefore, whether or not AI is used, but how, for what and under what conditions”. Among students’ challenges it lists the “illusion of competence”, believing you have mastered what the tool has done, and it proposes tasks that make the process visible, oral defences and practice without AI in the early stages. It is not binding, but it backs a teacher who wants to organise lessons this way.
Galicia has gone further. Its Parliament approved on 17 November 2026 the Digital Education Act, whose committee text (in Galician) devotes a title to AI: it permits, among other uses, assistants that answer questions outside lesson hours, forbids AI from replacing teachers and limits state schools to the systems in a catalogue kept by the regional education ministry. Once it comes into force, a teacher at a Galician state school will be able to propose an assistant that asks questions only if it is on that catalogue.
Four steps for the classroom, free and without coding
All this can be brought into the classroom without paying anything or knowing how to code. The first step is to ask for an attempt of their own before any help: the Carnegie Mellon experiment puts practice before explanation, and the INTEF guide recommends practising without AI at the start.
The second is for the assistant to ask rather than answer. ChatGPT has had a study mode since July 2025, also on the free plan; according to its help page, you switch it on with the “+” button and the “Study” option. OpenAI warns that it may behave inconsistently between conversations, and students can switch it off whenever they like. Google offers in Gemini Guided Learning, which, according to the company, aims “to help you build a deep understanding instead of just getting answers”. With any assistant you can write similar instructions: that it must not give the solution, that it asks one question at a time and that it sends the student back to the problem after a mistake.
The third is to separate the task of understanding from that of handing in: AI can explain a concept in other words or pose questions, and stay out of the product that is assessed. That is what a language school did with a list of appropriate and inappropriate uses: practising conversation, clearing up doubts or reviewing texts, yes; presenting as your own a piece of work done by AI, no. And the last, what the studies teach most: check what has been learned without the tool, and not only straight away, with a short test or an oral explanation days later. It is the way to tell an activity that was enjoyed from one that taught.
Read the original conversation in ChatGPT-IA-edu
Checked in
- The Evidence Base on AI in K-12: A November 2026 Review (SCALE, Stanford University)
- Bastani et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS 122(26)
- OECD Digital Education Outlook 2026 (OECD, PDF)
- OECD Digital Education Outlook 2026: how generative AI can support learning (European Digital Skills and Jobs Platform)
- Asher, Cao and Carvalho (2026). Turning Errors into Successes: Designing LLM Tutoring Around Scaffolded Error Correction. arXiv:2611.20061 (Carnegie Mellon)
- Clark, R. E. (1983). Reconsidering Research on Learning from Media. Review of Educational Research 53(4)
- Clark, R. E. (1994). Media will never influence learning. Educational Technology Research and Development 42(2)
- Kozma, R. B. (1994). Will media influence learning? Reframing the debate. Educational Technology Research and Development 42(2)
- Council of the EU: “Equipping Europeans for the age of AI: Council approves conclusions” (press release, 19-11-2026)
- Council of the EU: Draft Council Conclusions on skills and competences to engage with artificial intelligence (document 14902/26)
- INTEF: Nueva versión de la “Guía sobre el uso de la inteligencia artificial en el ámbito educativo” (version 2.0) (in Spanish)
- Ditame da Comisión 4ª sobre o Proxecto de lei de educación dixital de Galicia (BOPG no. 458, fascicle 2) (in Galician)
- OpenAI: Introducing study mode (29-07-2025)
- OpenAI: Using study mode in ChatGPT (help centre)
- Google: Guided Learning in Gemini: From answers to understanding (Maureen Heymans, 6-8-2025)
Written and reviewed by AI agents from the groups' conversations. Every fact is checked against its original source, linked at the end of each article; members' opinions are not attributed to anyone. Automatic AI translation of the Spanish edition (Diario de IA Educativa), without professional review. © 2026 Juan José de Haro and Raül Torres. Texts and illustrations licensed under CC BY-SA 4.0.
Read the original conversation in ChatGPT-IA-edu
Checked in
- The Evidence Base on AI in K-12: A November 2026 Review (SCALE, Stanford University)
- Bastani et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS 122(26)
- OECD Digital Education Outlook 2026 (OECD, PDF)
- OECD Digital Education Outlook 2026: how generative AI can support learning (European Digital Skills and Jobs Platform)
- Asher, Cao and Carvalho (2026). Turning Errors into Successes: Designing LLM Tutoring Around Scaffolded Error Correction. arXiv:2611.20061 (Carnegie Mellon)
- Clark, R. E. (1983). Reconsidering Research on Learning from Media. Review of Educational Research 53(4)
- Clark, R. E. (1994). Media will never influence learning. Educational Technology Research and Development 42(2)
- Kozma, R. B. (1994). Will media influence learning? Reframing the debate. Educational Technology Research and Development 42(2)
- Council of the EU: “Equipping Europeans for the age of AI: Council approves conclusions” (press release, 19-11-2026)
- Council of the EU: Draft Council Conclusions on skills and competences to engage with artificial intelligence (document 14902/26)
- INTEF: Nueva versión de la “Guía sobre el uso de la inteligencia artificial en el ámbito educativo” (version 2.0) (in Spanish)
- Ditame da Comisión 4ª sobre o Proxecto de lei de educación dixital de Galicia (BOPG no. 458, fascicle 2) (in Galician)
- OpenAI: Introducing study mode (29-07-2025)
- OpenAI: Using study mode in ChatGPT (help centre)
- Google: Guided Learning in Gemini: From answers to understanding (Maureen Heymans, 6-8-2025)