On this page
AI chatbots make homework faster. The question that matters for exams is different: when the tool is gone, can you still do it? Researchers have started to test this directly, and the answer depends on how the AI is used. This post covers what cognitive offloading is, what the studies found, and how to use AI so it builds your skills instead of replacing them.
What is cognitive offloading?
Psychologists Evan Risko and Sam Gilbert (2016) define cognitive offloading as using physical action to change what a task demands of your mind, so as to reduce mental effort. Writing a shopping list, setting a reminder and tilting your head to read sideways text are all offloading. It is often sensible. The risk comes when the effort you offload is the effort that would have made you learn.
Practice problems exist to make you do the hard step yourself. If a chatbot does that step, the homework gets done but the practice does not happen.
What happened when students used ChatGPT for maths practice?
The clearest evidence comes from a field experiment published in PNAS by Bastani and colleagues (2025). At a large high school in Turkey, nearly 1,000 students in grades 9 to 11 did practice sessions in one of three conditions: no AI, a standard ChatGPT-style interface ("GPT Base"), or a version with prompts designed to protect learning ("GPT Tutor"), which gave teacher-designed hints instead of giving away answers.
During practice, AI helped a lot: grades rose by 48% with GPT Base and 127% with GPT Tutor. But on a later exam without AI, students who had used GPT Base scored 17% lower than students who never had access. The safeguards in GPT Tutor largely removed that harm. The authors found that students often used GPT Base as a crutch, asking for and copying solutions.
Do other studies point the same way?
The evidence is still young, but several studies line up.
In a randomised study of 117 university students writing essays, Fan and colleagues (2024) found that the ChatGPT group improved their essay scores the most, but did not gain more knowledge or show better transfer than other groups. The authors warn that AI may encourage dependence and "metacognitive laziness", meaning less planning and checking of your own work.
A team at MIT recorded brain activity (EEG) while 54 people wrote essays with an AI assistant, a search engine or no tools. Their preprint (Kosmyna and colleagues, 2025) reported the weakest brain connectivity in the AI group, and that AI users struggled to quote their own essays. It is a small study and was posted as a preprint, so treat it as a signal, not a settled result.
The pattern is not "AI is bad for learning". A randomised trial by Kestin and colleagues (2025) found that students learned more in less time from an AI tutor built on teaching principles than from an active-learning class. The difference is design: an AI that makes you think versus one that thinks for you.
How to use AI without losing the practice
Attempt first, always. Give every problem a real try, at least to the point of writing down what principle you think applies. Then ask for help.
Ask for a hint, not the answer. Say "Give me only the next step" or "What should I check?" This is close to the GPT Tutor design that avoided the harm.
Close the chat and redo it. After AI helps on a problem, solve a similar one on paper with the chat closed. If you cannot, you have not learned it yet.
Explain it back. Tell the AI how you would solve the problem and ask it to find the flaw. That keeps the reasoning in your head.
Test without the tool. Once a week, do a timed set with no AI at all. Your score there is the real measure.
Our post on AI tutors versus answer chatbots explains the design differences in more detail.
An AI that teaches rather than answers
The research points to one feature that matters most: whether the AI does your thinking or guides it. Fermi's AI tutor is built to teach rather than answer. It notices where you are stuck and decides whether to explain, ask or show, so the hard step stays yours.
For your next homework set, write your first step on paper before you open any AI tool.
