What scientific evidence reveals about the real impact of AI on learning (and what schools and publishers need to decide)
EduIA pulse #1 | research on AI and learning in education by BlinkLearning
What Research Tells Us About AI and Learning This Month
AI in education is transforming learning, but there’s a question everyone in education is asking right now: does AI really teach, or does it simply provide answers?
This month, four key studies bring us closer to an answer. Spoiler: AI can be transformative in the classroom, but only if it’s well designed and used. And what “well” actually means is more complex — and more interesting — than it seems.
1. AI improves outcomes… but are students actually learning?
“The Evidence Base on AI in K-12: A 2026 Review” Fesler, Martinez, Agnew & Loeb — AI Hub for Education, Stanford University (2026)
Stanford has just published the most comprehensive review to date on the impact of AI in K-12 education. They analyzed more than 800 academic studies and found that only 20 provide strong causal evidence. A finding that calls for collective humility.
What do those 20 studies tell us? That AI in education improves student performance while they have access to it—in math, programming, and writing. But when they are assessed without that support, the effects fade or disappear. Researchers call this the difference between assisted performance and durable learning.
Two findings deserve special attention. First: pedagogical design matters more than the technology itself. Chatbots that guide students step by step—rather than giving direct answers—lead to stronger learning outcomes. Second: AI in education works very well for teachers. Teachers who used ChatGPT to prepare lessons reduced their planning time by 30% without losing quality, according to Roy et al. (2024). And less experienced tutors were the ones who benefited most from AI-powered feedback systems.
Current evidence suggests that tools designed to foster independent reasoning are more likely to support durable learning.” — Fesler et al., Stanford, 2026
2. Beyond the grade: how to measure learning with AI
New tools to understand AI and learning outcomes” — OpenAI + University of Tartu + SCALE Initiative, Stanford (March 2026)
OpenAI published this month the results of a study involving more than 300 university students who used ChatGPT’s Study Mode to prepare for microeconomics and neuroscience exams. In microeconomics, students with access to the tool scored 15% higher than the control group.
But the most valuable part of the article isn’t those results—it’s the honesty about its limitations. The OpenAI team acknowledges that an exam score is not enough to determine whether AI in education improves learning. That’s why they developed the Learning Outcomes Measurement Suite, a longitudinal measurement framework designed together with Stanford and the University of Tartu (Estonia), where it is being tested with nearly 20,000 students aged 16 to 18.
This system measures things that, until now, no one was measuring: autonomous motivation, persistence in the face of difficulty, metacognition, and long-term retention. A paradigm shift the sector urgently needed.
“What truly matters is whether improvements persist over time.” — OpenAI team, 2026
3. The real impact lies in the learning sequence
“Effective Personalized AI Tutors via LLM-Guided Reinforcement Learning” Chung, Zhang, Kung, Bastani & Bastani — Universidad de Pennsylvania + National Taiwan University (2025)
This study is one of the most rigorous published to date: a randomized controlled trial with 770 high school students in Taipei over five months, teaching Python. All students had access to the same AI chatbot. The only difference: half received practice problems in a personalized sequence using reinforcement learning, adapted in real time to their level.
The result was an improvement of 0.15 standard deviations on the final exam without AI assistance, which—according to some studies—is equivalent to an additional 6 to 9 months of schooling. And without increasing instructional time or teacher workload.
The most surprising part: students in the personalized group didn’t solve more problems or harder ones. What changed was their engagement: they were more involved, persisted longer, and used the chatbot more productively—asking questions to understand rather than requesting direct answers.
The benefit was greater for beginner students and those from lower-performing schools, opening up significant opportunities for educational equity.
“Personalizing the learning experience—not the chatbot itself, but the sequence of problems—can sustainably improve engagement.” — Chung, Bastani et al., 2025
4. The risk of AI: cognitive offloading
“Thinking—Fast, Slow, and Artificial: how AI is reshaping human reasoning and the rise of cognitive surrender” — Shaw & Nave, The Wharton School, University of Pennsylvania (2025)
The most conceptual article of the month—and perhaps the most important for any educator. Wharton researchers propose the Tri-System Theory: humans no longer think only with System 1 (fast intuition) and System 2 (slow, deliberate reasoning), as Kahneman suggested. There is now a System 3: AI.
And System 3 has a serious problem: it produces what the authors call “cognitive surrender.” In three experiments with 1,372 participants, when the AI provided an incorrect answer, participants still followed it in 80% of cases. Most concerning, access to AI increased participants’ confidence in their answers—even when they were wrong.
Who is most vulnerable? Individuals with higher trust in AI and a lower tendency to question their own thinking. What helps? Incentives and immediate feedback reduce cognitive surrender—though they don’t eliminate it entirely.
For educators, the key question is: are we teaching students to think with AI, or to delegate to it without thinking?
“We don’t just use AI—we think with it. And that changes who authors our decisions.” — Shaw & Nave, Wharton, 2025
The big idea of the month: access to AI is not enough
These four studies, published between 2025 and 2026, all point in the same direction: access to AI is not enough—and can even be counterproductive if not managed properly.
What makes the difference is design: tools that guide without giving answers, that adapt difficulty to the learner, that measure learning over time, and that keep critical thinking active. This is not technological magic. It is intelligent pedagogy.
How to integrate AI with intention: the Blinklearning approach
At BlinkLearning, we have spent years working on exactly that: technology in service of learning, not the other way around. And each month, the research reminds us why this principle matters.
The key question for your school
The key question for your school
If your school or institution is using AI in the classroom, how do you know whether students are actually learning more—or simply performing better with support? Do you have a way to measure it?
Nos encantaría leer tu experiencia en los comentarios. 👇
📌 EduIA Pulse es la publicación mensual de BlinkLearning sobre investigación en IA y educación. Si te ha resultado útil, compártela con alguien de tu equipo.
🔗 Fuentes:
— Fesler et al. (2026). The Evidence Base on AI in K-12. Stanford / AI Hub for Education.
— OpenAI + U. Tartu + SCALE (2026). Learning Outcomes Measurement Suite.
— Chung, Bastani et al. (2025). Effective Personalized AI Tutors via LLM-Guided RL. U. Pennsylvania / NTU.
— Shaw & Nave (2025). Thinking—Fast, Slow, and Artificial. Wharton, U. Pennsylvania.