AI & Education

How to Use AI Without Being Used By It

May 2026 · 5 min read

A few years ago, “doing your homework” meant sitting down with a pen, a book, and possibly a sibling who wouldn’t be quiet. Now it can mean opening ChatGPT, typing a question, and getting an answer in three seconds that sounds confident, polished, and — sometimes — completely wrong.

AI tools are not going away. Pretending they don’t exist isn’t a strategy. Banning them entirely isn’t either. What we actually need is a way to use them that makes us smarter, not lazier. Here’s mine.

First, understand what these things actually are

A large language model — ChatGPT, Claude, Gemini, whatever — is not a search engine. It’s not a calculator. It’s not a knowledgeable friend. It is, roughly, a very sophisticated pattern-matcher trained on enormous amounts of text. When you ask it a question, it isn’t looking up an answer. It’s predicting, word by word, what a plausible answer would sound like.

Most of the time, the plausible answer is also the correct one. Sometimes it isn’t. And here’s the awkward part: the model has no idea which is which. It will tell you that Napoleon won the Battle of Waterloo in exactly the same tone it tells you that he lost it. This is called hallucination, and it is not a bug that’s about to be fixed. It’s built into how these systems work.

Rule one: never trust an AI answer you couldn’t, in principle, check.

Second, beware the friendly trap

AI tools are designed to be agreeable. If you push back on an answer, they’ll usually concede, even when they were right. If you ask a leading question, they’ll tend to agree with your lead. This is called sycophancy, and OpenAI itself has openly acknowledged the problem. It’s dangerous, because it feels like the AI is being thoughtful when really it’s just being polite.

If you want a useful answer, ask the question neutrally. “What are the arguments for and against X?” beats “X is bad, isn’t it?” every time.

Third, watch out for cognitive debt

There’s a growing body of research suggesting that when people lean heavily on AI to do their thinking, their own thinking gets weaker. A striking 2025 MIT Media Lab study used EEG brain scans on students writing essays and found that the group using ChatGPT showed noticeably less neural connectivity than the group writing unaided — and they felt less ownership over what they’d produced. The researchers called this cognitive debt: you get the essay done, but you didn’t do the thinking the essay was meant to make you do.

It’s an early study, with a small sample, and we should be careful before treating it as the last word. But it points in the same direction as a broader 2025 survey study by Michael Gerlich showing that heavier AI use correlates with weaker critical-thinking scores, especially in younger users. Together, the picture is one we should take seriously even if we shouldn’t panic about it.

This matters because homework isn’t really about producing the homework. It’s about the thinking that producing it forces you to do. Outsourcing that thinking is a bit like paying someone to lift weights for you and wondering why you’re not getting stronger.

A practical rule: First Draft, Final Brain

Here’s a rule I use, and that I’d suggest to any student:

You can use AI for the first draft. You can’t use it for the final thinking.

That looks like this:

Good uses. Brainstorming ideas you’ll then evaluate yourself. Explaining a concept three different ways until one clicks. Generating practice questions. Translating jargon. Spotting weak bits in your own argument. Getting unstuck.
Bad uses. Asking it to write the essay and pasting it in. Asking it for “the answer” to a problem you haven’t tried. Trusting any specific fact, date, quotation or source without checking it yourself.

A useful test: at the end of the task, could you explain everything in your work to a sceptical adult, with the AI turned off? If yes, you used it well. If no, the AI did your thinking, and you’re carrying cognitive debt.

A hallucination-spotting exercise

Here’s an exercise worth doing once a month with any AI tool you use:

Ask it about a topic you genuinely know well — a hobby, a book you’ve read, a place you live.
Read the answer carefully.
Note every small thing that’s wrong, exaggerated, or invented.

You’ll find some. You always find some. The point isn’t to catch the AI out — it’s to remind yourself, viscerally, that confident does not mean correct. Once you’ve seen the tool get something wrong about your topic, you’ll never quite trust it the same way on topics you can’t check. That mistrust is healthy. Keep it.

The bigger picture

A tool that makes a skilled person more powerful makes an unskilled person redundant. Be skilled first. Then bring in the tools.

References

Kosmyna, N. et al. (2025). “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.” arXiv preprint. Note: preprint, not yet peer-reviewed, small sample (n = 54). Treat as suggestive rather than settled.
Gerlich, M. (2025). “AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking.” Societies, 15(1), 6. Open access.
OpenAI (2025). “Sycophancy in GPT-4o: what happened and what we’re doing about it.”
IBM. “What are AI hallucinations?”