This lesson is not about getting caught. It is about a quieter loss: doing well on the work and learning nothing, then finding out later, in an exam hall or a job, that the ability was never yours.
The uncomfortable research
A reasonable belief is that reading a good explanation teaches you. The evidence says otherwise, and this is one of the better-established findings in learning science.
The generation effect: material you struggle to produce yourself is remembered substantially better than material you read, even when the read version was clearer.
The testing effect: retrieving an answer from memory strengthens it more than re-reading. Studying feels productive; retrieving feels hard and works better.
Desirable difficulties: conditions that slow learning down and make it feel worse — spacing sessions, mixing problem types, struggling before being told — produce better long-term retention. The feeling of ease during study is a poor predictor of what you will retain.
Put those together and the danger becomes clear. AI is very good at removing exactly the difficulty that does the teaching. The clearer the explanation, the smoother the experience, and the less your memory has to do.
A controlled study at a Turkish high school in 2024 found students given an unrestricted chatbot for maths practice performed better during practice and then *worse* on an unaided exam than students with no AI. A tutor-style version that withheld answers avoided the drop. The tool was not the variable. Whether it did the work was.
The line
The practical test is not what the tool is, but which cognitive step it performed.
Fine — the AI handles something that was never the skill you are building:
- Explaining a concept a different way after you tried to understand it
- Generating practice problems, then you solve them
- Quizzing you and marking your attempts
- Being the confused student you explain the topic to, which is one of the strongest study techniques there is
- Formatting, tidying references, translating a passage you already understand
- Debugging code you wrote, after you have read the error yourself
Not fine — the AI did the thinking the assignment was measuring:
- It wrote the essay, or the paragraph, or the argument
- It produced the solution and you transcribed it
- It summarised the reading instead of you reading it
- It generated the code and you submitted it without being able to explain any line
The grey area is real and depends on what is being assessed. If you are graded on argument, having AI polish your grammar is fine. If the course is academic writing, it may not be.
Three habits that keep the learning
Attempt first, always. Write your answer badly before asking. This is the single highest-value rule, because it forces retrieval before you get help, and it means you can see the difference between your version and a better one — which is where the learning actually is.
Ask for the method, not the answer. "What approach applies here and why?" beats "solve this." Then apply it yourself and get the arithmetic wrong on your own.
Close the tab and reproduce it. If you cannot rebuild the argument or re-derive the result without help, you did not learn it. You watched it. Twenty minutes later, from a blank page, is the honest test.
Also: it will be wrong sometimes
Everything from the hallucination lesson applies double here. Models make mathematical errors, misstate historical dates, and confidently explain concepts backwards. A student who cannot yet tell a good explanation from a fluent wrong one is the least equipped to catch it — which is another reason to attempt first. Having your own answer gives you something to notice a disagreement against.
What is genuinely good about this
A patient explainer, available at 2am, in your own language, that never sighs when you ask the same question a fourth time, is a real gift — especially if nobody in your house studied this subject and there is no money for a tutor. That access is worth defending. It is worth using in the way that leaves the knowledge in your head.
Before you move on