Who does the thinking?
The practice trap
In a study of nearly 1,000 maths students in Turkey, those given a plain chatbot did 48% better during practice, then 17% worse than students with no AI once the chatbot was taken away for the exam. They had used it to get answers, so the practice felt easy and left little behind. Learning needs effort at the right moments.
Tutor-style AI works better
The same study tested a version set up like a tutor, giving hints instead of answers. Students using it did far better in practice and no worse in the exam. A 2025 Harvard trial went further: physics students using a carefully designed AI tutor learned more than twice as much as in an active-learning class, in less time. Design and use make the difference.
Desirable difficulty
Psychologists use the phrase 'desirable difficulty' for effort that feels slow but makes memory stick: recalling instead of rereading, trying before being told, mixing topics up. AI is very good at removing difficulty. Your job is to keep the useful kind and let it remove only the pointless kind, such as hunting for a definition.
A simple rule
Before you ask AI anything while learning, try first: write your own answer, guess, or sketch what you think. Then ask the AI to check, hint or explain the gap. The attempt is where the learning happens. The AI's job is to make each attempt more useful, not to replace it.
Think of something you looked up with AI recently. Could you explain it to someone now, without looking?