Will AI Destroy Schools and Universities?

If a student can hand in a brilliant essay and still not understand the subject, what are we actually measuring?

That is where I keep getting stuck in the conversation about AI and education. An essay can make sense from beginning to end, with an argument that barely happened in the student's head.

I understand why that worries people. Writing an essay is supposed to involve thinking: deciding what you believe, finding something that supports it, noticing where the argument falls apart. If a student skips that work, a better-looking submission may hide how little they have learned.

But I also think we risk stopping the conversation there. AI could help students avoid learning, and it could give a teacher more ways to help them learn. I don't think preserving the way we were taught is always the same as protecting what mattered about it.

A phone's contact list is a small example of how a tool changes what we need to remember. You can call someone without knowing their number. Remembering a number is different from understanding an argument, but I recognise the temptation: once something else can do a task well enough for our immediate purpose, it becomes easier to stop practising it. In a classroom, that purpose might be handing in the assignment. The skill it was meant to develop can disappear from view.

Which things can we reasonably look up, and which do we need to understand well enough to recognise when an answer is wrong?

In a 2025 study involving nearly a thousand high-school maths students, access to a general GPT-4 interface improved performance during practice. When the tool was taken away, that group performed worse than students who had practised without it. A version with safeguards for learning was designed not to return the answer and largely reduced that negative effect. Bastani and colleagues, PNAS

I don't read one maths experiment as proof that using AI makes students incapable of thinking. It does make me want to ask what happens after the help is removed.

Then I imagine a class of thirty students covering the same topic. They are unlikely to be stuck in exactly the same place. One might need a simpler explanation. Another might understand the explanation but need more practice. A third might need a reason to care about the problem at all.

Teachers already respond to those differences. AI could help them prepare several explanations, adjust the difficulty of an exercise, or build an example around a student's interests. The teacher would still need to check whether the material was correct and useful.

A student who loves football might work through percentages using league statistics. They still have to understand percentages, but football gives them a reason to care about the problem. Later, give them a percentage problem that has nothing to do with football and see whether they can use what they learned.

A 2025 randomised trial with college students found greater learning gains in less time using a custom AI tutor than in the active-learning classes used for comparison. The tutor was deliberately built around teaching practices. That is encouraging, though a result with a particular tutor and group of students cannot settle how AI will work across an entire education system. Kestin and colleagues, Scientific Reports

If preparing different materials takes less time, perhaps a teacher could spend more of it talking with a student. I would want to check whether that actually happens. Generating more worksheets is not much of an achievement if checking them leaves the teacher with even less time.

Those conversations could also change how a student sees their progress. Can they explain why they chose an approach? Can they apply it to a problem they haven't seen? What do they still understand when we come back to it later? A conversation can reveal something that a polished submission leaves hidden, and help the student notice their own improvement.

Tests and essays can still have a place. I would just be less comfortable treating the finished piece of work as sufficient evidence on its own.

My concern is that schools and universities can take too long to respond while students are already changing how they work. I want them to test these possibilities sooner, with teachers involved. Start with one part of a course and decide what students should be able to do afterwards. Try a different way of helping them, then check their understanding, including without AI and again later.

I wonder whether smaller schools might have an advantage here. A smaller group may find it easier to agree on a limited experiment and change it when something fails, though size alone tells us very little about whether a school is willing to learn from what happens.

The harder part may be admitting how much of our reaction is personal. I recognise this in business. When a way of doing things has helped you succeed, it is easy to feel protective of it. You worked at it and learned the rules, so a different approach can feel like it is dismissing that effort.

I can understand that feeling, and I think we need to examine it. A student taking a different route isn't necessarily taking a shortcut, and an educator questioning AI isn't necessarily afraid of change. There are things worth preserving, including the effort involved in learning to think. There are also habits we may defend because they are familiar to us.

I don't know exactly what schools and universities will look like as this develops. I am hopeful they can make room for students to learn differently while taking the evidence of learning seriously. We are still working out how.

If you were teaching a class tomorrow, what would convince you that a student understood something after using AI?