Studying with AI works better when it asks the questions

A thirteen-year-old is stuck on enzymes. She asks the chatbot, and ninety seconds later she has a tidy explanation with a lock-and-key analogy and a friendly closing line. She reads it. It makes sense. She thinks, right, got it, and shuts the laptop.

On Friday the test asks about enzymes and nothing comes back.

Most arguments about children and AI are about permission: allowed or banned, cheating or not. The more useful argument is about direction. In almost every AI study session tonight, the machine produces and the child receives. Flip that, and the same tool starts doing something the evidence supports.

What American students are mostly using it for

RAND published findings from its American Youth Panel in March 2026, by Heather L. Schwartz and Melissa Kay Diliberti, drawing on 1,214 young people surveyed in December 2025. They range from 12 to 29, so read this as students broadly, not only schoolchildren.

The headline numbers, as reported by eSchool News: the share using AI for homework rose from 48% in May 2025 to 62% in December, with middle schoolers jumping from 30% to 46%. Meanwhile 67% said using AI for schoolwork harms critical thinking, up from 54% earlier in the year, and even among the students who use it, 60% said so.

Now what they use it for. Better explanations of assignments came top at 38%, then brainstorming at 35%, looking up facts at 33%, and drafting or revising writing at 33%. Three of those four are the machine talking and the child listening. That is the part worth changing, and it needs no ban.

The catch is that a good explanation feels like learning

This trap predates chatbots. In 2009, Jeffrey Karpicke, Andrew Butler and Henry Roediger surveyed 177 college students in the journal Memory about how they actually studied. Their conclusion: “A majority of students repeatedly read their notes or textbook (despite the limited benefits of this strategy), but relatively few engage in self-testing or retrieval practise while studying.” Students, they wrote, “lack metacognitive awareness of the mnemonic benefits of testing”.

Rereading produces a warm familiarity the brain misreads as knowing, and a fluent AI explanation is rereading with better production values: clearer than the textbook, faster than the teacher, already tidied up, so your child never does the messy work of assembling the idea themselves. Whether the help lands or evaporates depends on who does the work, as we have written before.

The least surprising finding in learning science

A 2017 meta-analysis in Review of Educational Research by Olusola Adesope, Dominic Trevisan and Narayankripa Sundararajan put it plainly: “practice tests are more beneficial for learning than restudying and all other comparison conditions”. Not more enjoyable. More beneficial.

Lab findings often die on contact with a real classroom, so Pooja Agarwal, Ludmila Nunes and Janell Blunt went looking, in a 2021 systematic review in Educational Psychology Review. They screened nearly 2,000 abstracts, coded 50 experiments run in actual schools covering 5,374 students, and found that 57% showed medium or large benefits from retrieval practice. Read that honestly: a solid majority is not a guarantee, and only 6% of the experiments came from outside Western, educated, industrialised, rich and democratic countries.

Four ways to turn the chatbot around

No app or subscription needed. Just the child producing the sentences.

• Shut the book first. Retrieval only counts when the answer comes out of memory. Notes open on the other tab defeats the exercise.
• Ask it to ask. Something like: “Ask me one question at a time about the water cycle. Do not give me the answer. Wait for mine, then tell me what I missed.” The constraints that matter are one at a time and wait, because without them the model will helpfully answer its own question.
• Answer out loud or on paper before reading the reply. A half-formed spoken answer is retrieval. Skimming the model’s version and nodding is not.
• Come back to the same questions cold a few days later, and judge the second attempt rather than the first. That last one is ours, not a research finding, but it is the cheapest way to see if anything stuck.

Where the machine makes a shaky examiner

Handing the marking to a chatbot has a real failure mode: it can tell your child they are right when they are not. Liang Zhang of the University of Memphis and Edith Aurora Graf, formerly of Educational Testing Service, ran four models over 30 maths problems in work accepted for the AIME-Con conference in October 2025. Asked to mark solutions step by step against human experts, GPT-4o reached only “fair” agreement (a kappa of 0.366) while OpenAI’s o1 reached “substantial” (0.737).

That was adult-level maths, not a Primary 5 worksheet, so do not over-read it. The transferable point is sturdier: which model is marking matters more than which app it is wearing, and a confident wrong “correct!” is worse for a child than no quiz at all. So tell your child arguing back is allowed. If it marks them wrong and they still think they were right, that disagreement is the most valuable thirty seconds of the evening.

One more honest note: retrieval practice was proven with humans, paper and classrooms, and very little of that work tested a chatbot as the quizmaster. Pointing a well-evidenced method at an untested tool is a reasonable bet, not a proven one.

The version already sitting in your child’s school account

Singapore parents have a free head start here that most have never opened. GovTech’s account of the AI features in the Student Learning Space describes a “Test Myself” mode inside the Adaptive Learning System, and notes that “if a student gives a wrong answer, ALS may offer hints or explanations to help them understand their mistake before they try again”. Hints on a wrong answer is the behaviour you want, and it is what a general chatbot will not do unless told. Worth one question at the next parent-teacher meeting: is Test Myself switched on for this subject?

(Our commercial position, plainly: Mentus AI builds mentors designed to ask rather than answer, so we have an interest in this argument. The research above is not ours, and we have linked it.)

The question at the end of homework is not whether AI was involved. It is what your child had to pull out of their own head tonight. If the answer is nothing, the evening was pleasant and empty, however good the explanation was, and whatever cognitive debt it quietly ran up.