Four Questions to Ask Before Putting an AI Tool in Front of a Child

Shammika Munugoda, Founder of PebbleYard, presents four questions that parents and teachers should ask before giving a child an AI tool. This article originally appeared in Insight Jam, an enterprise IT community that enables human conversation on AI.

A trial published in PNAS last year gave nearly 1,000 high school maths students access to GPT-4 while they practiced. Their practice scores jumped 48%. Then the researchers took the AI away and set the same exam for everyone. The students who’d had it scored 17% lower than students who never touched it.

That gap between performing and learning is the most useful thing I’ve read about AI in education. It’s also why I think “does AI improve learning?” is the wrong question. The better one is simpler: who did the thinking?

The same trial had a second group. Their AI tutor was set up to give hints instead of answers. On practice problems, they did 127% better than the control group, which sounds like a win until the exam. No difference from students who’d had no AI at all. The guardrails stopped the damage. They didn’t produce learning.

Where I’ve seen AI move a learning outcome in a controlled trial, a human was in the loop. Stanford researchers ran a preregistered trial with 900 tutors and 1,800 students in the US, and the AI never spoke to a student. It whispered suggestions to the tutor mid-session. Students were 4 percentage points more likely to master the topic, and 9 points more likely for students of lower-rated tutors. The cost worked out to roughly $20 per tutor per year.

Side by side, the pattern is clear. Harm showed up when the AI handed over answers. Gains showed up when it helped an adult do their job better, and the student still had to do the effortful part.

None of this should surprise anyone who has read the learning research. Pulling something from memory without peeking does more for long-term retention than rereading it, and practice testing ranks near the top in Dunlosky and colleagues’ 2013 review of study techniques. It’s old, boring, and it works.

The catch is that someone has to write the practice. A teacher with 28 students and a unit on fractions this week doesn’t have an hour to write 20 good questions, and a parent has less. That’s where AI is genuinely useful. Not as a clever tutor, but as a fast first draft of practice material tied to what the class is learning right now. The learning still comes from the child dragging the answer out of their own head. The model just saved the adult an evening.

I should say where I’m coming from. I’ve built software for 25 years, and this year I built PebbleYard, a practice tool for primary-age kids where a parent or teacher can ask for a lesson, quiz, and spelling set on whatever is being studied. Getting a model to write a decent quiz was the easy part. The hard part was deciding everything it wouldn’t be allowed to do.

So the child never talks to it. Only an adult can ask for anything, and kids see finished, reviewed material. A person approves every draft before a student sees it. The model gets a topic and a year level and nothing about any child. It declines a short list of subjects, including anything sexual, hateful, or dangerous. Sessions are short, and they end, because a tool built to keep a child playing is solving a different problem.

Every one of those choices costs something. It’s less magical, and it needs an adult in the room. I’d make the same calls again, mostly because of that first trial.

Here’s what I don’t know. I don’t have outcome data from my own users yet, since the product is a few months old, so I’m leaning on trials with bigger samples and better controls than I could run. Be suspicious of any vendor who says otherwise, me included. The trials I’ve cited are also secondary-school maths and US tutoring. Evidence on AI-written practice for primary-age children is thin, and anyone calling it settled is guessing.

Review isn’t optional, either. A reviewer recently caught a line in one of our language lessons describing Chinese as using “pictures called characters”. It’s close enough to pass a skim and wrong enough to teach something false. Whether a person or a model drafts a sentence like that, the review step is what catches it. Models are very good at being plausible.

If you’re evaluating an AI tool for a classroom or a child, four questions tell you most of what you need. Who’s doing the thinking, the student or the tool? Who is accountable for what a child actually sees? What data leaves the building? And does it know when to stop?

An AI tool that answers those well may not demo as impressively as one that doesn’t. In the PNAS trial, the unrestricted version looked great right up until the exam. Ask about the exam.