How to teach your child to fact-check an AI

A whole country bet on AI in the classroom, then quietly hit undo. That is worth a parent’s attention, and not for the reason you might think.

A country that hit undo on its AI textbooks

In 2025 South Korea did something no other education system had tried at that scale. It put AI-powered digital textbooks into primary and secondary classrooms for English, maths and computer science, built by commercial publishers with roughly 800 billion won of industry investment behind them. The pitch was personalised learning for every child.

By August 2025 the National Assembly had stripped those books of their legal status as textbooks, narrowing the official definition back to printed books and e-books and pushing the AI versions down to optional “educational materials”. Adoption tracked the mood. It slid from 37 percent of schools in the first semester to about 19 percent in the second, with only around 2,095 schools still using them by September, according to reporting by Rest of World.

The complaints were ordinary and damning. A maths teacher told Rest of World the books were poor quality and “clearly hastily put together”. Students said lessons were delayed by technical faults and were not actually tailored to their level. Parents worried about screen time; a teachers’ union flagged data protection.

Here is the part for parents, and it is not schadenfreude. If a national programme, government-funded and publisher-vetted, could still ship AI schoolwork that teachers called rushed and unreliable, then the free chatbot your child opens for a homework question, with no vetting behind it at all, will get things wrong too. The useful question stops being “is this tool accurate?” It becomes “does my child know how to check it?”

Confident and wrong is the hard combination

The problem was never that AI is sometimes wrong. Everything is sometimes wrong. The problem is that AI is wrong in a fluent, self-assured voice that sounds identical to when it is right. There is no wobble, no “I think”, no pause. That is exactly the tone a child reads as authority.

And catching it is genuinely hard, even for people you would expect to manage. A 2026 study in arXiv looked at 63 senior computer engineering students describing their run-ins with AI hallucinations. These are tech-literate young adults, and still only 46 percent used any active verification strategy. Roughly a quarter did not try to spot errors at all. Most who did catch something relied on gut feeling, sensing an answer “didn’t make sense”, rather than actually checking. The single most common failure they reported was the AI citing sources that were wrong or did not exist, at 23 percent of cases.

If trained engineering undergraduates coast on intuition, a ten-year-old certainly will. The authors’ conclusion is blunt and useful: AI literacy has to move past prompt-writing to “specific protocols for fact-checking and lateral reading”. In plain terms, checking has to be taught. It does not arrive on its own.

The good news: checking is learnable, and fast

Children can pick this up, and quicker than you would guess. A 2026 study presented at the CHI conference had 48 children aged 10 to 14 build and test their own chatbots, with small supports built in: confidence signals, a fact-checking step, asking the same question more than once, and comparing two models side by side. The children showed clear gains in spotting hallucinations and in their understanding of how the tools work. They started inventing their own moves, probing for inconsistencies and cross-checking against outside sources.

The honest caveat matters too. The same study found the children leaned too heavily on visible cues and used the checking tools patchily. A confidence badge is not judgement. But the direction is unmistakable: kids learn to check by doing it, not by being warned about it.

Australia has written the same instinct into policy. Its national Framework for Generative AI in Schools, from the federal Department of Education, asks schools to teach students how these tools actually work, including their limitations and biases, rather than treat them as an oracle. You can teach the identical lesson at the kitchen table, and probably sooner than the school will.

What to actually do this week

You do not need a curriculum. You need a few small habits, borrowed from what worked in that study, that turn your child from a consumer of answers into a checker of them.

• Ask it twice. Have your child pose the same question again in a fresh chat. If the answer shifts, that instability is the tell. Real facts do not flinch.
• Catch it together, on purpose. Ask the AI something you both already know cold, ideally where it is likely to slip, and let your child spot the mistake. Watching a confident machine get it wrong is far more memorable than any lecture.
• Open the source. When the AI cites something, click through and read it. Given that made-up citations were the top error in the study above, this one habit does a lot of work.
• Name the confident voice. Point out that the tone never changes whether it is right or wrong, so tone tells you nothing. Correctness and confidence are different things.

None of this is about raising a child who distrusts AI, and none of it is about raising one who trusts it either. The goal is smaller and sturdier: a child whose first reflex, faced with any slick answer, is to ask “how could we check that?” That reflex is the thing worth building now, while the stakes are a maths worksheet and not something that matters more. South Korea can rewrite its textbook law. Your kid just needs the habit.