The question isn't whether AI can teach. It's whether students will let it

Philip Oreopoulos is a distinguished professor in Economics of Education Policy at the University of Toronto. Can artificial intelligence actually help children learn? As students head back to school this September, it is hard to find anyone without a strong opinion. Last month, my research lab at the University of Toronto released results from two studies through the National Bureau of Economic Research: randomized experiments with thousands of students in ordinary math classes. What we found supports neither the hype nor the panic. It points somewhere more interesting. Learning requires effort, but AI's most impressive talent is eliminating effort. Students learn math by wrestling with problems, making mistakes and working out why. Watching someone else lift weights does not make you stronger. Opinion: The cheating generation is asking us to help them stop One solution is to add guardrails: get AI tutors to coach rather than answer, ask students to explain their reasoning, and break problems into steps while holding back the solution. This is the right idea. But it runs into a second, more stubborn problem: students have to want the help. Students tend to prefer the easier option. In our first study, we followed middle school students across 18 Tennessee schools for two years as they did daily math practice on one of the world's largest learning platforms, with a built-in AI tutor designed to coach them through mistakes. Achievement rose modestly and at low cost. But the tutor itself went largely unused. Nearly every student tried it at least once. The typical student then consulted it on only about a third of practice days, and turned to it after fewer than 20 per cent of the exercises where they made mistakes, the moments it was designed for. Help was one click away. Students didn't click. In order to encourage students to take the harder but more rewarding route, we built our own practice platform, constructed for research, and randomly assigned more than 6,000 students to math practice that was identical except for the AI. Students with the tutor slowed down. They attempted fewer problems but answered more of them correctly. After making a mistake, they were more likely to get the next question right. And when the AI was combined with a rule requiring students to keep trying until demonstrating they could answer three questions correctly in a row, and having to work with the AI tutor after every mistake rather than taking another guess, students remembered more of what they practiced when tested a week later. The gains were modest, not transformational – some of the first experimental evidence that a well-designed AI tutor produces learning beyond what the same practice delivers without one. These effects came from design choices, not from the technology itself. The next step, then, is not a smarter tutor but one students want to use even when the work is hard, and know how to use when they do. That means letting them talk to it instead of typing, since many kids won't write out what confused them but will say it. It means patience, humour and playfulness. It means meeting students at their level and connecting practice to hockey or baking or whatever they love. These sound like gimmicks. They are the whole game. A tutor students ignore teaches no one. Opinion: We've convinced students they shouldn't sound like themselves when they write. No wonder they use AI Skeptics will say a screen is no substitute for a human being, and they are right. A good teacher working one-on-one beats an AI tutor, and no one is proposing to replace teachers. But that is not the choice on offer. A teacher facing 25 or 30 students cannot sit beside each of them at the moment they get stuck. Private tutoring works, and it costs thousands of dollars a year, which is why it flows overwhelmingly to the families who need it least. For most students, most of the time, the realistic alternative to an AI tutor is not a caring adult. It is no help at all: a kid staring at a problem, alone, deciding whether to guess or give up. Maybe the human element will prove irreplaceable – maybe engagement with a machine, however charming, has a ceiling that a person does not. What our evidence shows so far is that the question has an answer: with structure, guardrails, and designs that draw students in, AI tutoring can add real learning – and without them, it mostly adds a window nobody opens. So as students head back to school amid bans on one kind of screen and mandates for another, the question this year isn't whether AI belongs in the classroom. It's whether we will do the slow, unglamorous work of finding out what makes it help – before we decide it can't, or assume it already does.