On this page
Matching the answer key tells you whether you got the answer, not whether you can say why the tempting alternatives fail. A sharper ten-minute test: ask an AI to defend a wrong answer as convincingly as it can, then knock its argument down.
Why argue against a wrong answer at all?
A refutation text states a common misconception, refutes it, and explains the correct idea. In a meta-analysis by Noah Schroeder and Aurelia Kucera in Educational Psychology Review, refutation texts showed a moderate advantage (g = 0.41) over other learning conditions across 44 independent comparisons with 3,869 learners.
A second line of research looks at learning from other people's mistakes. A 2025 meta-analysis by Ecenaz Alemdag, Anja Eichelmann and Susanne Narciss in Review of Educational Research pooled 42 papers (177 effect sizes) comparing erroneous examples with correct examples or problem-solving, and found only a weak overall benefit from studying erroneous examples (g = 0.136). The benefit was larger when learners were prompted to explain the error, or were given an explanation, than when they were given no explanation.
These findings suggest the explaining is what helps, and arguing against an AI makes you do it.
How do you ask an AI to defend a wrong answer?
Pick a wrong answer you once believed, or one your classmates fall for.
Paste this prompt: "Play a Class 11 student who believes [wrong claim]. Argue for it as convincingly as you can. Do not reveal the correct answer. When I reply, concede a point only if I have proved it with a law or definition, and tell me which step of my reply was weakest."
Write your rebuttal on paper before asking for any help.
Then ask: "Drop the role. Which parts of my rebuttal did I prove, and which did I only assert?"
Check the final position against your NCERT textbook.
The "concede only if" line matters. A 2023 study of sycophancy by Mrinank Sharma and colleagues found that five state-of-the-art AI assistants consistently showed sycophancy, favouring a user's beliefs over the truth, across four varied tasks. An AI that folds the moment you push back proves nothing. The same move works outside science, as when you use AI as a debate partner on GM crops and gene editing.
Worked example: is acceleration zero at the top of a throw?
The AI's case: "At the highest point the ball stops. If it is not moving, nothing is changing, so the acceleration must be zero."
A weak rebuttal: "No, acceleration is always g." True, but remembered, not argued. A strong rebuttal finds the broken step:
Acceleration is the rate of change of velocity, not the velocity itself. Velocity being zero at one instant says nothing about how fast it is changing.
If acceleration were zero at that instant and stayed zero, the velocity would stay zero and the ball would hang in the air. It falls, so its velocity is changing.
The ball's weight does not switch off when it pauses, so by F = ma the acceleration is still g, downward.
On a velocity-time graph, the line crosses zero without bending. Its slope, which is the acceleration, is the same before, at and after the top.
Worked example: does a catalyst push equilibrium towards products?
The AI's case: "A catalyst speeds up the reaction. The Haber process uses an iron catalyst to make more ammonia. So a catalyst gives more product."
Your rebuttal can come straight from NCERT Class 11 Chemistry, Unit 6 Equilibrium, section 6.8.5: a catalyst lowers the activation energy of the forward and reverse reactions by exactly the same amount and does not affect the equilibrium composition. It gets you to equilibrium faster, not to a different one. The same chapter says the Haber catalyst gives a satisfactory rate at temperatures where the ammonia yield is reasonably favourable, and that temperature, not a catalyst, changes Kc. The AI's own evidence turns against it.
How do you know your rebuttal is good enough?
You named the step in the argument that fails, not just the final claim.
You used a definition or law, not "the book says so".
You showed something the wrong answer predicts that does not happen, like the ball hovering.
You could explain it to a younger student in three sentences.
The AI conceded for a reason you could repeat, not because you kept pushing.
To see which topics you only thought you knew, rate your confidence before checking answers with AI. If the AI and your textbook disagree, use the five-step check for when the answer key and AI disagree.
Parents need no physics to help: ask your child to read the AI's argument aloud, then their rebuttal, and ask which step was wrong and how they know.
Getting stuck at the same step usually points to a concept gap. The Fermi Concept Library has its own lesson for each concept in physics and chemistry, and Fermi's AI tutor is built to teach rather than answer: it notices where you are stuck and decides whether to explain, ask or show.
