Teach me. Don’t just hand me the answer.

You ask AI to explain something difficult. The explanation is clear. The examples make sense. You nod at the screen with the confidence of a person who has absolutely got this.

Then you try it without the explanation. Your new expertise has apparently stepped out for lunch.

Recognizing a good answer and producing one yourself are different achievements. If your goal is learning, give your assistant a job that leaves some of the thinking to you. The useful test comes later: what can you explain or do when the help is gone?

Choose one visible result

“Make me brilliant at photography” is an ambitious opening. It is also difficult to test. Try something smaller: “Help me understand aperture, shutter speed and ISO well enough to explain a camera-setting choice in my own words.”

Tell the assistant what you already know, where you get stuck and how much time you have. Ask it to check your starting point before planning lessons. A short exchange can reveal whether you need an introduction, a missing prerequisite or practice applying something you understand in theory.

Keep a real reference nearby, such as a reputable textbook, course or equipment manual. AI can help you work through it. Its feedback still needs checking when an explanation seems doubtful or contradicts the source.

Let the lesson leave room for an attempt

Ask for a short explanation, then a question you must answer before seeing the solution. If you are stuck, request one useful hint. If that is not enough, ask for a worked example and then try a similar problem yourself.

The point is to make your understanding visible. Endless guessing is not a virtue. Neither is watching an assistant complete every step while you supply the occasional enthusiastic “Exactly.”

For the photography example, have it describe a simple situation and ask what you would change. Explain your choice before asking for feedback. Then ask it to identify the specific misunderstanding, show a correction and give you a fresh situation. You now have something more informative than a reassuring summary.

The research makes the distinction matter

In a 2025 study involving nearly 1,000 students at one Turkish high school, access to an answer-oriented GPT-4 tool improved assisted math practice scores but reduced performance on subsequent unassisted tests. A more carefully designed tutor avoided that detectable harm, but its unassisted results were not significantly better than the no-AI group. [1]

That tutor had teacher-supplied solutions and guidance about likely mistakes, not simply a request to give hints. The study covered particular math lessons and immediate tests. It does not prove that AI harms all learning, or that one clever prompt guarantees a good tutor. [1]

OpenAI’s preliminary March 2026 report also gives a mixed picture. In randomized research with more than 300 college students, Study Mode variants showed gains in microeconomics, while neuroscience results were not distinguishable from the comparison group. Analysis was ongoing, and the company left longer-term durability unresolved. [2]

The practical response is to check your own progress instead of assuming that access to a powerful model has settled the learning question.

Close the chat and try again

After a lesson, put the explanation away. Write a short account of what you learned from memory, or solve a fresh problem. Reopen the source to check it. Mark the part you missed and revisit it later.

There is evidence for the value of retrieval practice beyond AI. In experiments by Henry Roediger and Jeffrey Karpicke, undergraduates who practiced recalling prose passages remembered more after a delay than students who repeatedly studied them. That supports testing your recall; it does not establish a universal timetable or prove mastery of every practical skill. [3]

For a hands-on skill, include hands-on practice. A tidy explanation of camera controls is a useful start. Making choices with a camera gives you a different kind of evidence about what you can do.

A prompt you can adapt

“Help me learn [specific skill]. My goal is to [observable result]. First check my starting level. Teach one idea at a time, then ask me to attempt a question before showing the answer. Offer hints when needed and explain errors clearly. Ask me to explain the idea in my own words. Finish with a new task I can try without your help and a short plan for later review. Use [trusted reference] and flag uncertainty.”

Treat this as a practical routine to try, not a proven formula. Adjust it when the questions are too easy, the hints are unhelpful or you need instruction from a qualified teacher. Let the assistant know when its feedback is unclear, and check important corrections against your reference.

A good learning session should leave you with more than a beautifully organized chat. It should leave you able to do something you could not do before, even after you close it.

Sources

[1] Bastani and colleagues 2025. Generative AI without guardrails can harm learning

[2] OpenAI 2026. New tools for understanding AI and learning outcomes

[3] Roediger and Karpicke 2006. Test Enhanced Learning