During my high-school years, I almost failed English several times because a large part of the grade came from homework I barely did. Yet I passed the final test every term, and my teacher never worked out how.
As it turns out, I was running a small, very questionable but rather profitable business that left me no time for assignments: for an agreed fee, I wrote English essays and exercises on demand (note that this was in Spanish-speaking environment). To make them look authentic, I would ask my classmates for their opinions and develop them into full essays. For an extra fee I would even write with my left hand, so that the handwriting looked, as we believed back then, more like a boy’s. Looking back, I was essentially ChatGPT with handwriting features. For years the arrangement worked well for everyone involved, except for one little problem: only one of us was learning English while the other just ended up with a grade.
I always think of this story when a professor asks me how to make sure students do not cheat with AI.
Much of the current debate on education and AI revolves around three questions:
how to stop students from cheating with AI
which skills are still worth teaching; and
how, if at all, can AI contribute to learnin
As someone who designs courses for a living, I hear them from professors every week, and students have long since answered the first one themselves: in the United Kingdom, 92% of students now use generative AI for their studies (HEPI, 2025). And while AI may feel inevitable, the shape it takes is not: as Karen Hao shows in Empire of AI (2025), the industry has grown under a narrative that makes its own expansion seem unstoppable, but how we use AI in education is still ours to decide.
In this essay I draw on three roles I hold at once: I design learning for others, I learn with AI myself, and I build the AI systems I learn with. Combining them has changed both how I learn and how I design learning for others.
Stop trying to stop students from using AI
When an education system rewards productivity and looking like you know, faking it becomes the rational move, as my classmates and I worked out long before ChatGPT.
In 1899 Thorstein Veblen coined the term conspicuous consumption for people who buy expensive goods not because they need them but to signal status. AI-polished emails, essays and slides do the same with knowledge: they signal that their author knows something, and that they were efficient about it. Assessment that grades only the finished product encourages appearing competent at the expense of becoming competent (Ames, 1992). This is why, when I design learning experiences, I focus on skill building, applied knowledge and self-reflection, and on something almost never said aloud: not knowing is fine, and the mild discomfort of facing the unknown and moving through it is how everyone has ever learned anything.
Institutions are not immune to looking like they know either. A 2025 survey of German universities found that 97% were dealing with AI in exams, but only 15% had an AI strategy, and not a single one reported that students were strongly involved in the decisions. The reflex is primarily defensive, and teachers, too, are still catching up.
Nor should we blame AI for the current literacy crisis. The PISA 2025 results showed the lowest averages ever recorded in reading and mathematics across OECD countries, yet scores had been falling since around 2012, years before ChatGPT. There are good reasons for caution with AI, as Maria Weaver has argued in a previous edition of this magazine, but before blaming the tool we should ask what actually motivates students to hand their work to a machine.
We have been here before. My teachers used to tell me never to use Wikipedia —which has since turned out to be one of the more reliable sources around (Giles, 2005). The problem was never Wikipedia but copying without thinking, and what would have helped me back them is someone showing me how to check a source. The same applies to AI now.
The skill worth teaching is steering your own learning
If copying without thinking is the problem, the next question is what students should learn to do instead. Research on what students should learn points —with remarkable consistency— to the same direction. Employer surveys such as the World Economic Forum’s Future of Jobs Report 2025 and frameworks from the OECD and the EU all describe the same profile: analytical and creative thinking, curiosity, self-regulation and the ability to keep learning.
What these frameworks rarely spell out is what self-regulation actually consists of. Psychologists describe it as a cycle of planning, performing and reflecting (Zimmerman, 2002), but in practice it comes down to a handful of concrete decisions:
What do I want to learn today?
Where am I weak?
What kind of practice fits this mistake?
When do I need someone to check my work?
Most learners do not know these decisions exist because teachers and curricula make them on their behalf. If anything is worth teaching explicitly in the age of AI, it is making these decisions.
This is where my own work comes in. I am a psychologist and instructional designer, and most of the courses I have worked on deal with future skills, data science, programming and AI. My job consists of sitting down with industry experts and professors until I understand what someone should be able to do at the end of a course and why it matters. In some cases this can take weeks (ask three experts what a one-month introduction to AI should cover and you will get at least four versions of the same course).
Learners can design their own AI tutors
Yet even though research points to specific skills, the underlying model of teaching has, overall, not caught up. Even at leading AI companies —which presumably need problem solvers more urgently than anyone— much of the training offer consists of recorded videos and short tutorials. The teacher gives, the student takes, and an exam checks what stuck, a model Paulo Freire called banking education (1968). Against it, he proposed problem-posing education, which starts from problems in learners’ own lives.
That is the idea behind project- and problem-based learning, and over the years I have found a basic formula for it that works regardless of the subject. First, students understand the key concepts of the topic; then they identify a problem from their personal or professional context that relates to those concepts; after that, they sketch a solution and try it out; and finally, they analyse its limits and communicate it to the people it concerns.
How does AI fit into this?
I believe that, if used well, AI could push education towards problem-posing learning, because it lets learners move from being paying users of a chatbot to designing their own learning systems. My own German is a good example. After four years in Germany, I needed a C1 German certificate for professional reasons. I also wanted to learn the language properly and improve my writing and grammar were. So I designed my own German tutor, a set of saved instructions that turned a general chatbot into an exam coach. It grouped the mistakes in my essays into patterns, marked each error without correcting it, built drills from my own sentences and acted as my partner in mock oral exams. I directed my learning, while my German teacher explained, corrected and pushed me further. I have since published the whole system, so anyone can adapt it.
In retrospect, this was a project in exactly the sense described above. I started from a problem of my own, identified what I needed to improve, prototyped a solution, saw it work when I passed the exam and now I am communicating it.
What made it work was that I used AI to support my thinking rather than replace it. Bastani and colleagues (2025) found that high-school students with unrestricted access to GPT-4 improved their practice scores but once access was removed, they performed 17% worse in the exam than students who had never had it. A version that gave hints instead of answers prevented most of that loss. My tutor, which marked errors without correcting them, was exactly that kind of guardrail.
Human attention is the key ingredient
AI is part of how I work: I research topics with it and use a workflow that turns experts’ notes into units, readings and activity proposals. AI enthusiasts regularly ask me why I even need the experts, if AI could generate a whole course and a tutor to answer questions around the clock.
My answer comes first, from experience. What makes content “click” is a professor describing a problem they once solved, explaining where students always get stuck, or telling a terrible joke that somehow explains the whole thing. Lee Shulman (1986) called this pedagogical content knowledge: knowing not only a subject but how to teach it to particular people, knowing which examples work and which misconceptions to expect. It comes from years of watching real students get confused. AI has an abundance of content knowledge but very little of this kind. Students in my courses may use AI if they can explain their decisions without notes, but what makes them feel they have achieved something is meeting the experts in person.
The empirical evidence points the same way. A two-year experiment with Khanmigo, Khan Academy’s AI tutor, found that almost every student tried it, but when they got something wrong they asked it for help in only 17% of sessions, and learning gains were about the same as without the AI (Oreopoulos and Low, 2026). The researchers’ conclusion: “human attention is the key ingredient.” Where AI tutoring does work, someone is steering. At Harvard, an AI tutor in an introductory physics course helped students learn significantly more, in less time, than an active-learning class (Kestin et al., 2025), but it was designed by the instructors themselves and used by highly motivated students.
AI, then, tends to give most to those who already bring the most to it, what sociologists call the Matthew effect (Merton, 1968). In a field experiment with 640 entrepreneurs in Kenya, those already doing well increased their revenue by around 15% with a GPT-4 business advisor, while those who were struggling lost around 10% (Otis et al., 2026). Likewise, anybody can “vibe code”, but an experienced developer will get much further than a novice, and someone who understands their own learning goals can design a far better learning path. Yet most learners have the tool without the training: in the UK, only 36% of students had received any AI training from their university (HEPI, 2025). Nearly everyone has the engine; very few have been taught to drive.
A fair objection is that I am the Matthew effect myself. My German tutor system worked because I already knew how to make the learning decisions it depended on: what to practise, where I was weak, when to ask my teacher. Most learners, and certainly a teenager who struggles to concentrate, have never been shown that these decisions exist and how to make them. That is exactly why they should be teaching them now. The answer is not a better tool but a learning process made explicit, with a teacher alongside who makes these decisions visible and students solving problems in the classroom, where someone can help when they get stuck.
Classroom reimagined
So how would I deal with classrooms, AI and learning in general?
Let students use AI at home for what it does well: explanations, drills and fifty versions of the same exercise. Let them read the slides at home too (I would rather read slides under a blanket anyway). Class time should then go to what only people can offer, namely conversation, open questions and a teacher’s companionship. Educators call this a flipped classroom (Bergmann and Sams, 2012), an old idea that AI finally makes easy to put into practice.
I remember experiencing an accidental version of it at school. During my last year we had a free study hour, which we first treated as an extra-long break, until the students who had understood something began explaining it to the rest. Nobody designed that; we had simply been given time and a little freedom. Research backs this up: students who struggle with a problem before receiving instruction learn more than those taught first (Sinha and Kapur, 2021). What matters is accompanied freedom: a problem to solve in class, with an expert nearby for when students genuinely get stuck.
This changes what a teacher is for: less delivering, more asking, pushing and drawing stories out of people. That is why I think we need more teachers, not fewer, because content delivery scales with AI, whereas attention does not. When a student once complained that a whole project could be done with AI, a professor at my university simply asked who was studying here, the student or the AI. I would put that question on the wall of every classroom.
Schools could also teach the self-regulation decisions I made for my German without realizing it, so that learners can make them consciously and ask AI to help with each one. For university leaders, this means unglamorous work first: safe AI access and training for everyone before flashy agents, and students in the room when AI policy is decided. And families matter too: my own curiosity was not a talent; it was cultivated by people who made room for questions and for not knowing.
The learner belongs in the driver’s seat
Learning happens to whoever feels curious or motivated enough to do the cognitive work. To return to the three questions:
rather than trying to stop students from using AI, we should stop rewarding the appearance of knowing;
the skill most worth teaching is how to steer one’s own learning; and
AI contributes most when the learner is the one steering.
In a system that rewards grades over curiosity and with AI offering an ever more convincing illusion of productivity, the main job of education is no longer to deliver content but to build the knowledge, motivation and self-regulation that make AI useful in the first place.
AI is a vehicle. Once you know where you want to go, it gets you there faster. If you do not, the system absorbs you and takes you wherever it is already heading: towards more generated text, less reading, and the feeling that you do not matter because a machine can do your work. The part school should teach is that not knowing is fine, that curiosity is powerful, and that asking questions and planning one’s own learning are skills that can be learned. In the end, the point is not to ask AI to teach you something, but to ask yourself what do I want to learn?
References
Ames, C. (1992). Classrooms: Goals, structures, and student motivation. Journal of Educational Psychology, 84(3), 261–271. https://doi.org/10.1037/0022-0663.84.3.261
Bastani, H., et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
Bergmann, J., & Sams, A. (2012). Flip your classroom: Reach every student in every class every day. ISTE.
Freire, P. (1970). Pedagogy of the oppressed (M. B. Ramos, Trans.). Herder and Herder. (Original work published 1968)
Giles, J. (2005). Internet encyclopaedias go head to head. Nature, 438, 900–901. https://doi.org/10.1038/438900a
Hao, K. (2025). Empire of AI: Dreams and nightmares in Sam Altman’s OpenAI. Penguin Press.
Higher Education Policy Institute (2025). Student generative AI survey 2025. HEPI. https://www.hepi.ac.uk/reports/student-generative-ai-survey-2025/
Kestin, G., et al. (2025). AI tutoring outperforms in-class active learning: An RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, 17458. https://doi.org/10.1038/s41598-025-97652-6
Merton, R. K. (1968). The Matthew effect in science. Science, 159(3810), 56–63. https://doi.org/10.1126/science.159.3810.56
Oreopoulos, P., & Low, N. (2026). One click away: AI tutoring with Khanmigo in a two-year school experiment (NBER Working Paper No. 35620). National Bureau of Economic Research. https://doi.org/10.3386/w35620
Otis, N., et al. (2026). The uneven impact of generative artificial intelligence on entrepreneurial performance: Evidence from a field experiment in Kenya. Management Science. https://doi.org/10.1287/mnsc.2024.06909
Shulman, L. S. (1986). Those who understand: Knowledge growth in teaching. Educational Researcher, 15(2), 4–14. https://doi.org/10.3102/0013189X015002004
Sinha, T., & Kapur, M. (2021). When problem solving followed by instruction works: Evidence for productive failure. Review of Educational Research, 91(5), 761–798. https://doi.org/10.3102/00346543211019105
Zimmerman, B. J. (2002). Becoming a self-regulated learner: An overview. Theory Into Practice, 41(2), 64–70. https://doi.org/10.1207/s15430421tip4102_2
