Opinion. The views below are the editorial board's.
Every new tool for thinking comes with an illusion attached. Calculators let us believe we were good at arithmetic. Search engines let us believe we knew things we had only located. Generative AI's illusion is more convincing than either, because it doesn't just hand us an answer. It hands us an explanation, in fluent prose and in our own idiom, and then it waits politely for the next question.
That fluency is the trap. When an answer reads well, we take it as understanding. When we feel we understand, we stop doing the slow work that produces understanding. And when we stop doing that work, we need the tool a little more the next time. Call it the illusion circle: fluency feels like knowledge, the feeling replaces effort, and the missing effort deepens reliance. Each turn of the loop is comfortable. That's why it is hard to notice we're in one.
This isn't a hypothetical
The clearest evidence comes from a field experiment with nearly 1,000 high school math students in Turkey, published in the Proceedings of the National Academy of Sciences. Students who practised with an unrestricted ChatGPT-style assistant did 48% better on practice problems than students without AI. When the AI was taken away, the same students scored 17% worse on the exam than the control group, according to the researchers' summary. Their chat logs showed why: most students simply asked for the answer.
The students in that study weren't lazy. They were rational. The tool made the work feel finished, so they finished.
People can feel the circle closing
Students themselves sense it. In RAND's 2025 national surveys, 55% of high school students and 48% of middle schoolers agreed that greater use of AI will harm critical-thinking skills. So did 61% of parents. Among school district leaders, the people deciding how AI enters classrooms, only 22% agreed, RAND reported.
That gap should worry us more than any single statistic. The people closest to the learning feel the circle tightening, and the people furthest from it don't.
Breaking the loop
None of this is an argument against AI in schools. The same PNAS study tested a second assistant, one that gave hints instead of answers and included teacher input. Students using it did 127% better in practice, and their exam scores were no worse than the control group's. Design matters.
Schools can act now:
- Buy tools that withhold answers by default. A tutor that asks a question back is doing its job. One that hands over solutions is doing the student's job instead.
- Grade the process, not just the product. Drafts, reasoning and oral explanation are harder to outsource than a final answer, and more honest evidence of learning.
- Teach the illusion out loud. Only 19% of students told RAND that their teachers had given them guidance on using AI for schoolwork. Students can't resist a loop no one has named for them.
The illusion circle closes quietly, one helpful answer at a time. The good news is that it can be opened just as quietly, with a better-designed tool, a better-designed assignment, and a teacher who asks, "How do you know?"