Make AI Explain the Step You're Stuck On, Not Hand You the Answer

I grade a lot of problem sets. The difference between a student who used AI well and one who used it badly is obvious within two lines. The good ones can explain why line three follows from line two. The bad ones copied a clean derivation they can't reproduce on a quiz. Same tool, opposite outcomes.

The trap with hard math and science problems is that the AI can almost always produce a final answer, and a final answer feels like understanding. It isn't. Understanding is being able to redo it without the AI in the room. Here is how to aim the tool at that goal instead of the shortcut.

Stop asking for the answer. Ask for the obstacle.

When you paste a problem and say "solve this," you get a finished solution and you learn almost nothing, because the hard part of learning is doing the steps yourself. Flip the request so the AI works on your stuck point, not the whole problem.

Compare these two prompts for the same integral:

• Weak: "Evaluate the integral of x times e^x dx."
• Strong: "I'm trying to integrate x e^x dx using integration by parts. I chose u = x and dv = e^x dx. Is that the right choice, and why? Don't finish the problem, just tell me if my setup is sound and what principle guides the choice of u."

The second prompt gets you the actual lesson: pick u to be the part that gets simpler when you differentiate it. Now you own a rule you can apply to the next twenty problems. The first prompt gets you one answer you'll forget by dinner.

The reveal-one-step technique

My favorite move for a genuinely confusing problem is to make the AI go slowly and stop. Tell it explicitly: "Give me only the first step and the reasoning behind it. Then wait. Do not continue until I tell you I understand."

Then you try the next step yourself and check it: "My next step was to multiply both sides by the conjugate. Is that valid here? If not, what did I miss?"

This turns a wall of algebra into a conversation where you are doing the work and the AI is confirming or correcting. It is slower. That slowness is the point. Fast copying does not build the mental muscle that a closed-book exam tests.

Ask it to explain the same idea three ways

Textbooks give you one explanation. If it doesn't land, you're stuck. An AI can reframe on demand, which is genuinely useful for concepts that have a symbolic form, a visual form, and a plain-language form.

For something like standard deviation, try: "Explain standard deviation three ways: the formula and what each symbol does, an intuitive plain-English version, and a concrete example with the numbers 2, 4, 4, 4, 5, 5, 7, 9 worked through."

Cross-checking the three against each other is where the concept clicks. If the plain-English version and the worked example seem to disagree, that gap is exactly what you didn't understand, and now you can ask about it directly.

Where AI is unreliable in STEM, and how to catch it

You have to know the failure modes or the tool will quietly teach you wrong things.

Arithmetic and algebra slips. Language models are text predictors, not calculators. They drop signs, misplace factors of two, and botch multi-step arithmetic with total confidence. In 2026 the major chat tools (ChatGPT, Claude, Gemini) can run code or call a calculator for numeric work, but they don't always choose to. For anything where the number matters, verify: recompute by hand, or paste the expression into Wolfram Alpha or Symbolab, which are built for symbolic math and will show reliable steps. Use the chat AI for the reasoning and a dedicated math engine for the computation.

Confident wrong physics. It will happily apply the wrong formula (using kinematics where energy conservation is cleaner, or an ideal-gas assumption that doesn't hold). Defend yourself by asking: "What assumptions does this solution rely on, and when would they break?" A good answer teaches you the boundaries of the method. A vague one is a signal to check your textbook.

Invented sources. If you ask for a citation or a specific theorem name, it may fabricate one that sounds real. Never cite anything from an AI you haven't independently confirmed exists.

A verification habit that doubles as studying

Turn checking into a learning step instead of a chore. After you and the AI reach an answer, do one of these:

1. Plug it back in. For equations, substitute your solution into the original. Physics answers should survive a units check and a sanity check (a car doesn't accelerate at 500 m/s^2).
2. Change one number and redo it yourself. If you can solve the same problem with different values without help, you learned it. If you can't, you didn't, and you now know exactly where to focus.
3. Explain it back. Type the solution in your own words and ask the AI to poke holes: "Here's my explanation of why this works. What did I get wrong or leave out?" Teaching the idea back is one of the most reliable study techniques there is, and the AI makes a patient audience.

The academic-integrity line, stated plainly

Using AI to understand a graded problem set is a gray area at many schools and flatly banned at some. Two rules keep you safe and honest.

First, read your course's AI policy before you touch the tool, and when in doubt ask the instructor. "I didn't know" is not a defense anyone accepts. Some professors allow AI as a tutor but not on the specific problems you submit; some allow it with a disclosure line; some ban it entirely on assessments.

Second, apply the reproduce test. If you could sit down tomorrow with a blank page and redo the problem, the AI tutored you. If you couldn't, you outsourced the assignment, and you'll find out the hard way on the exam where the tool isn't allowed. The whole point of a problem set is to build the ability the exam measures. Anything that skips that ability is cheating yourself even when it isn't caught.

A quick starter workflow

1. Attempt the problem cold. Write down where you get stuck.
2. Ask the AI about that specific obstacle, not the whole problem.
3. Make it go one step at a time and wait for you.
4. Move the computation to Wolfram Alpha or a calculator; keep the AI for reasoning.
5. Verify the answer, then redo a variant unassisted.

Do that and the AI becomes the patient tutor you can't always get in office hours. Skip it and you get a very expensive answer key that will let you down in exactly the moment that counts.