On this page
Plenty of AI tools print a solution in numbered lines. That is not the same as explaining it. A useful explanation shows why each line follows from the last, so you could rebuild it without the screen. Here is what that looks like on a Class 11 problem, and what to do when a step is skipped.
What should a step-by-step AI explanation include?
Each step should answer four questions: which principle is being used, why it applies here, what it gives you, and whether the result makes sense. If any of those is missing, the step is a claim, not an explanation.
There is good evidence that the "why" matters most. In a classic study of students learning mechanics from worked examples, Chi and colleagues (Cognitive Science, 1989) found that stronger learners generated many explanations of the example steps and related them to principles in the text, while weaker learners explained too little and later relied heavily on the examples. A good AI explanation does part of that work out loud and leaves the rest for you.
Accuracy matters too. In a randomised trial in Harvard's introductory physics course (Scientific Reports, 2025), the researchers put comprehensive step-by-step answers into GPT-4's prompts rather than relying on it to generate solutions alone, and had it guide students through each part of each problem in turn. Students using that AI tutor had median learning gains over double those of students in an in-class active-learning lesson.
A worked example: a block on a rough incline
Problem: a 2 kg block is released on a 30° incline with a coefficient of kinetic friction of 0.2. Take g = 9.8 m/s². Find its acceleration.
A good step-by-step explanation:
Forces: weight mg down, normal force N perpendicular to the incline, friction up the incline. Why: friction opposes sliding.
Axes along and perpendicular to the incline. Why: the block moves along it, so one axis has zero acceleration.
Perpendicular: N = mg cos 30° ≈ 17.0 N. Why: no acceleration that way, so forces balance.
Does it slide? Assuming static friction is similar, tan 30° ≈ 0.577 is more than 0.2, so the pull down the slope beats friction.
Along the incline: ma = mg sin 30° − μN, so a = g(sin 30° − μ cos 30°) = 9.8 × (0.5 − 0.173) ≈ 3.2 m/s². Why: Newton's second law along the motion.
Sanity check: less than the frictionless 4.9 m/s², as it should be. Mass cancels, so the 2 kg never mattered.
Now a version that skips reasoning: "Friction = μmg = 3.92 N. Net force = mg sin 30° − 3.92 = 5.88 N. So a = 2.94 m/s²." It looks tidy. It is wrong: it quietly set N = mg instead of mg cos 30°, and never said why.
How do you spot when an AI skips reasoning?
Watch for these signs as you read:
A number appears that you cannot trace to the question or an earlier line.
Phrases like "simplifying, we get" or "clearly" replace a named law.
No diagram or axes are described before forces are added.
Units vanish halfway through.
The answer is never checked against a limiting case.
A formula is used without saying when it applies (here, only if the block slides).
For more checks, see our guide on how to check an AI's physics solution.
What should you ask the AI at each step?
Treat every line as something you are allowed to question. These prompts work with any AI tutor:
"Which law gives this line, and why does it apply here?"
"Where did this number come from?"
"Don't show the next step. Let me try it, then check my work."
"What happens if the angle is 90°? Does that make sense?"
The third prompt matters most. In a field experiment with nearly a thousand high school maths students (PNAS, 2025), students who practised with an unrestricted ChatGPT-style tool had grades 17% lower once access was removed than students who never had it, while a version with learning safeguards largely avoided that harm. Without guardrails, the authors write, students used GPT-4 as a "crutch" during practice. Trying the next step yourself first is a simple guardrail of your own.
This is the kind of teaching we build for. Fermi's AI tutor holds a line of reasoning across a session, notices where you are stuck and decides whether to explain, ask or show. If your mistakes tend to come from choosing the wrong method, the Fermi Problem Library has worked problems with a step-by-step solver that explains the reasoning rather than printing an answer.
Whatever tool you use, finish by closing the screen and redoing the problem. Read how to use a step-by-step solver without just copying it and what the research says about worked examples.
Tonight, take one solution you trusted this week and ask "why?" at every line. The line you cannot answer is the one to study.
