Key takeaways
- Used as an answer machine, AI improves your exercises but can lower your results when you work without it.
- Used as a tutor that guides without giving the solution away, it helps without that cost.
- Factual errors, or "hallucinations," are real: verify the points that matter at the source.
- The right split: AI explains and quizzes you, you produce the work.
What the research shows
During the 2023-2024 school year, researchers from the University of Pennsylvania ran an experiment with nearly 1,000 high schoolers in Turkey, in math class. One group had access to GPT-4 as a plain chat. A second had a version designed as a tutor, which gives hints rather than answers. A third had no AI.
During exercises with AI, both groups soared: +48% for the plain chat, +127% for the tutor. But on the exam, without AI, the plain-chat group scored 17% worse than students who never had AI at all. The tutor group matched them.
The authors' explanation: with the plain chat, many students simply asked for the answer. The AI acted as a crutch. And they didn't notice: they believed they had learned just as much as everyone else.
An AI that does the work for you mostly teaches you to ask it to do the work.
The real risk: the illusion of competence
Reading a clear explanation produces a genuine feeling of understanding. But understanding while reading and being able to redo it alone are two different things. It's the same trap as re-reading: the fluency of the explanation passes itself off as mastery. The remedy is the same too: test yourself, without help, right after the explanation. We cover this in our article on active recall.
7 ways to learn with AI without fooling yourself
1. Ask for hints, not solutions
When you're stuck on an exercise, ask for what would help you move forward, not the answer:
Don't give me the solution. Ask me a question that helps me find the next step, then wait for my answer.
2. Get yourself quizzed
After a lesson, ask for a quiz and answer without looking at your notes. One question at a time, so you don't read ahead before answering:
I just learned about proof of stake. Ask me one question at a time, wait for my answer, correct me, then move to the next one.
3. Explain first, get corrected after
Write your own explanation, then ask the AI to spot the errors and gaps in it. You do the work, it checks: that's the right division of labor.
4. Ask for examples and counter-examples
A concept is understood as much by what it isn't as by what it is. "Give me an example, then a case that looks like one but isn't."
5. Vary the level
Start with "explain it to me like a beginner," then step up: "now with the technical details." You're checking along the way that you're following.
6. Check the important facts
Numbers, dates, quotes, references: ask where the information comes from, and go read the source.
7. Keep control of the plan
A conversation answers whatever you ask, in the order you ask it. Without a plan, you quickly go in circles. Set a goal and a path, and use AI within it.
The hallucinations: how not to get caught out
Language models produce plausible text, not guaranteed truths. They can invent a reference, a date or a figure with the same confidence as a correct one. Recent models make fewer mistakes, but none is infallible, especially on niche topics or very recent events.
- Numbers, dates and quotes: always worth checking.
- Ask for the source, then open it: a reference that doesn't exist is a red flag.
- Be wary of very precise answers on a niche topic or recent news.
- Cross-check against a primary source: official documentation, research paper, legal text.
AI chat or a course built with AI?
Both have their place. A chat is unbeatable for a one-off question; a structured course wins when you want to go deep on a subject.
| An AI chat | A course built with AI | |
|---|---|---|
| Curriculum | You build it yourself, question by question | A plan from the start to your goal |
| Verification | Up to you to remember to test yourself | Exercises, quizzes and reviews built in |
| Memory | The conversation is forgotten | Spaced reviews of each concept |
| Sources | You ask, then you verify | Read and sorted when the course is built |
| Flexibility | Total, on any subject | A guide for your questions, inside the course |
In summary
AI is a formidable explainer and a patient examiner. It becomes a trap when it produces the work for you. Keep the effort that makes you learn, testing yourself, writing, redoing, for yourself, and hand it what it does better than anyone: explaining differently, any time, at your level.
Tell us what you want to learn: Dropee builds the plan, the exercises and the reviews, with an AI guide for your questions.
Frequently asked questions
Can you learn with ChatGPT?
Yes, as long as you use it to explain and quiz you, not to do the exercises for you. Studies show that an "answer machine" use improves results with AI but can hurt them without it.
Will AI replace teachers?
It can take over part of the explaining and the grading, at scale. Choosing the right goals, holding a high bar, and human support still carry real value.
How do you avoid AI's mistakes?
Check numbers, dates and quotes at the source, ask where each important piece of information comes from, and cross-check against a primary source. Be especially careful with niche topics and recent events.
What's the best prompt for learning?
A prompt that makes you work: "Ask me questions one at a time about [topic], at my level, wait for my answer, correct me, then move to the next one."
Sources
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. et Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26). doi.org/10.1073/pnas.2422633122
- Ji, Z., Lee, N., Frieske, R. et al. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 1-38. doi.org/10.1145/3571730
- Roediger, H. L. et Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249-255. doi.org/10.1111/j.1467-9280.2006.01693.x
- Bjork, R. A. (1994). Memory and metamemory considerations in the training of human beings. Dans J. Metcalfe et A. Shimamura (dir.), Metacognition: Knowing about knowing (p. 185-205). MIT Press.
Passionate about teaching and new technology, I write to help everyone learn smarter.