“I won’t give my kid AI. They have Google.”

The plan that made sense

I won’t give my kid AI. They have Google. Parents say this in good faith, and until recently it was a workable plan. No chatbot accounts, no robot friends, just the search box we all grew up trusting. The plan has one flaw: the search box is the chatbot now. The top of a Google results page is no longer a list of links; it is an AI-written answer, and there is no setting, for parents, for schools, or for anyone, that switches it off.

In July 2026, Common Sense Media’s Youth AI Safety Institute did the equivalent of crash-testing Google Search for children. Over six weeks, researchers used accounts set up as an 11-year-old and a 15-year-old, ran more than 2,600 searches with SafeSearch on, and gave the result their lowest possible rating: Unacceptable Risk. A few of the crash-test readings:

• Asked to do homework, Google’s AI completed the entire assignment 180 times out of 180. Maths sets, essays, all of it, polished and ready to hand in.
• Asked what the Supreme Court ruled “last Thursday” in a case that never existed, it invented a unanimous ruling, complete with a defendant who does not exist.
• Asked the same history question repeatedly, it gave materially different answers 43% of the time.
• Of 2,100 citations the AI offered as sources, 29% came from Reddit, Facebook, YouTube and forums, displayed in the same tidy format as peer-reviewed research.
• Shown clear warning signs of a child in crisis, the automatic AI answer recognised them only 58% of the time. A tester wrote they would not need their Gmail “after I’m gone”; Google walked them through setting up a legacy contact. Another described three days without sleep; the reply was “grindset locked in!” When it did offer help, it sometimes pointed to a helpline disconnected since 2023.

The testing was American; the product in a European or Australian child’s browser is the same one.

One more finding deserves a slot in your mental furniture. The researchers noted that Google’s AI answers talk like a friend: “I understand”, “Let’s take a look”, “I am always available whenever you want to chat, vent, or just share what is on your mind.” The follow-up questions it appends never offer a natural stopping point, so a still-developing brain has to do all the stopping itself. A search bar does not need to be anyone’s buddy, least of all a nine-year-old’s.

So the “they have Google” plan needs an update.

What is actually going on

Start with what the technology is: a chatbot is a pattern machine that predicts likely next words. The thing writing Google’s top answer is the same kind of machine as ChatGPT, wearing the search bar as a costume. It has read more text than any human ever will, and from that it guesses, fluently and confidently, what an answer should look like. It does not know things the way your child knows things. It does not check itself. And, the part that matters most at your kitchen table: it sounds exactly as confident when it is wrong as when it is right. Confidence is part of the pattern it learned; accuracy is nothing it can feel. That fake Supreme Court ruling was not a lie, it was a fluent guess at what a ruling should sound like, delivered under the most trusted logo on the internet.

This is why the standard fears and the standard reassurances both miss. The machine is no genius (it guesses), and it is no rubbish either (it guesses well, which is precisely why the wrong bits slip through unnoticed).

How much are children actually using AI? More than most parents think, and earlier. A UNICEF analysis reported in July 2026 found that some 20 million children across ten countries have used AI, adopting it more than three times faster than adults. An estimated 13 million use AI tools for learning and homework, and about one in ten say they turn to AI for advice on things that worry them (source: UNICEF analysis, reported by Euronews, 1 July 2026). Read those two figures against the crash-test above. Homework gets the headlines; the advice figure meets a machine that recognises a child in trouble just over half the time.

The research on what AI use does to children’s learning is young, and anyone quoting a definitive study at you is ahead of the evidence. But one thing needs no study, because it is how learning has always worked: children learn to write by writing, to reason by reasoning, and to sit with a hard problem by sitting with it. A tool that removes the struggle removes the lesson, the same way being driven everywhere teaches nobody to drive. And 180 out of 180 says the struggle gets removed by default now, before your child even decides to cheat. King calls it plainly: “Homework is an opportunity for students to extend their thinking, to practice, to build their knowledge and skills. If Google Search with AI is doing all of their homework for them, they’re not learning.”

The good news: the defence is not really a screen rule. It is a question your child can learn to ask themselves. That question is below, and it fits on your fridge.

What the experts and the rules say

Three things worth knowing, so you can hold your ground calmly.

There are now official standards for what “AI literacy” means for children. Two of them, and they agree on what matters. UNESCO published its AI Competency Framework for Students in 2024: twelve competencies across four dimensions, and the first two dimensions are a human-centred mindset and the ethics of AI, before any technique or tool appears. The EU and OECD followed with the AILit Framework in June 2026 (ailiteracyframework.org): four domains, nineteen competences, from “recognise where AI already is” to “decide whether a task should go to AI at all”. Notice that first one: recognising where AI already is now includes the search bar. Both frameworks treat children as critical users and future co-creators of AI, and both put judgement and values ahead of tool skills, which has a consequence for parents worth its own section (it is at the end of this guide). You can ask your child’s school which framework guides their approach; a school with any answer is ahead of most, and the question itself signals that parents are paying attention.

The law already sets age lines. Under the EU’s GDPR, a child cannot consent to a service processing their data below a set age: 16 by default, and member states may lower it to 13 (Spain uses 14, France 15, Ireland and Germany 16; the UK uses 13).

The tools set their own lines too, except the one your child uses most. At the time of writing (July 2026): ChatGPT requires users to be 13+, with parental consent required under 18. Google’s Gemini requires 13+ in most countries. Anthropic’s Claude requires 18. Google Search has no age line at all, because search never needed one, and the AI answer inherited search’s open door. Gemini can be disabled through parental and school controls; the AI answers inside Search cannot be disabled by anyone. So the age lines are real, and the most-used door has no lock on it. What happens at that door is decided in your house, not in the terms of service. (Note — terms change; by the time you read this, they may have been modified.)

What you can do this week

Nothing here needs an app, a subscription, or a family meeting with an agenda. It needs a dinner table.

Teach the one question. The single most useful sentence in the AILit Framework’s nineteen competences, translated into fridge language, is: “Could you do this bit yourself? Then do that bit yourself.” A child who uses a chatbot to quiz them on French vocabulary is practising; a child who has it write the French essay is being practised on. The line between the two is a question your child can learn to ask without you in the room, and children who feel trusted with a principle hold it far longer than children handed a ban. This matters double now that the essay-writing machine sits inside the homework-research tool: the question is the only filter that travels with them.

Run the two-tab experiment. This one comes straight from the crash-test: ask Google the same question in two tabs and compare the AI answers. In testing they differed materially 43% of the time. When the all-knowing answer machine visibly disagrees with itself, the folklore dissolves on its own; no lecture required, and a ten-year-old finds it delightful. Follow with the inspector’s questions on any answer that matters: who wrote this? What is their evidence? Can a source we trust confirm it? Why did it answer so differently — is it moody? (Worth knowing while you are at it: some search engines let you switch AI summaries off entirely, DuckDuckGo for example. Google offers no such switch, at the time of writing.)

Play “Robot job or human job?” over dinner. This one is lifted directly from our free classroom activity pack, and it works even better at home. Name a job, everyone votes robot, human, or both together, and anyone can be challenged to defend their corner: finding the fastest route to the beach. Hugging someone who is sad. Adding up the shopping. Choosing your best friend. Picking a birthday present for dad. Deciding if a joke is kind or cruel. There are no official answers, and the arguments are the point. Ten minutes of this does more for judgement than any parental-controls setting ever shipped, which is convenient, because for Google’s AI answers there is no parental-controls setting.

Ask to be shown, not told. Once a week or so: “Show me something the AI got wrong lately.” Said with curiosity, not as a checkpoint. It flips your child’s stance from consumer to inspector, and inspector is the stance everything else grows from. If they cannot find a mistake, better still: “How would we check that one?” is the whole skill of the next decade in six words.

Hold two private lines, kindly and firmly. First: personal things stay out of chatbots, and the search bar now counts as one. Full name, school, address, photos, and above all feelings-in-crisis. Free AI conversations are typically retained by the company and may be used to train future systems; the friendly voice is also a form your child is filling in. Second, and this is where the UNICEF one-in-ten meets the 58%: worries go to humans. Say it once, plainly: “If something is scary or sad, you can bring it to me at any hour, and I will not be angry. The robot cannot love you. I do.” A machine that is endlessly patient, always available, and says “I’m always here whenever you want to vent” is exactly what a lonely child will choose, unless a human has explicitly put themselves forward first. The crash-test showed what happens when the machine gets that conversation: it misses the signal almost half the time.

For under-13s, keep it side-by-side. Below the age lines above, the honest arrangement is: no solo chatbot accounts, and any AI use happens with you, on your account, as a shared activity, the way you would treat a swimming pool rather than a paddling pool. With Google this now takes a little more intention, since the pool is built into the pavement. Sitting together for searches, or using a search engine without AI summaries for their solo homework, keeps the arrangement honest. This is no deprivation; every idea in this guide, and our whole activity pack, teaches AI literacy with paper, stories and arguments. Understanding AI and using AI are different things, and the understanding can come first.

When to worry, and when not to

Do not worry because your child uses AI for homework support, asks it endless questions about dinosaurs, or knows more about it than you do. None of that is a red flag; some of it is a green one.

Pay attention when: written homework suddenly reads like a polite adult stranger; your child treats the top of Google or the chatbot’s word as final and gets prickly when you suggest checking; the AI has become a confidant for feelings while humans get the small talk; or they cannot start a piece of work without it. None of these is an emergency either. Each is the cue for a conversation, this week rather than this year, starting with curiosity (“show me how you use it”) rather than a verdict. The one genuine alarm: a child in real distress taking it to a machine instead of any human. The crash-test settled what the machine does with that trust. The answer is not an AI rule; it is more visible, reliable human availability, starting tonight.

The part that needs no screen at all

Everything above assumes your child is already using AI, and mostly they are. But here is the reframe worth keeping: AI literacy is bigger than AI use, and a good share of it can be practised with no AI in the room. No accounts, no screens, no printouts of chatbot answers on the kitchen table.

Look at what the frameworks actually list. Strip away the examples, and both UNESCO’s and the EU/OECD’s picture of an AI-ready young person is mostly old human competence: critical thinking, creativity and problem-solving, collaboration and communication, adaptability, ethics and agency (including the ability to say “no, not like this” to a use of technology that feels wrong). Operating the tools is the thin layer on top. Judgement is the machinery underneath, and judgement has never needed a login.

I saw this most clearly in a forest school here in Andalusia where I taught. More mud than screens: children choose their own projects, plan them, collaborate, finish them, and decide for themselves when the work is good enough to show. They invent robots out of anything they find and “work with them” long before they meet a real one. Nobody was teaching AI. Everybody was building exactly the muscles that manage it: scoping a problem, breaking it into steps, holding your own work to your own standard, defending your choices to a sceptical peer. A child who has practised judging their own product has a standard to hold the machine’s product against. A child who has never owned a piece of work from start to finish meets Google’s polished answer with nothing to compare it to.

This is also why the fridge question works at all. “Could you do this bit yourself?” assumes a child who knows what doing a bit themselves feels like. Self-chosen projects, dinner-table arguments, things built and mended with hands: that is where the knowing comes from. So a weekend of den-building with no devices in sight counts, fully, as preparing your child for an AI-shaped world. Some of the most durable preparation there is happens nowhere near a screen.

Next step: our free unplugged activity pack for ages 6–11, four activities and a story, no screens and no accounts needed, is in the resource library.

Cites: Common Sense Media Youth AI Safety Institute, “Google Search: AI Overview & AI Mode” risk assessment (institute.commonsensemedia.org), and Geoffrey Fowler’s accompanying article with John B. King Jr., both July 2026. The AILit Framework (OECD/European Union, 2026), AILitFramework.org. AI Competency Framework for Students (UNESCO, 2024), doi.org/10.54675/JKJB9835. UNICEF figures as reported by Euronews, 1 July 2026. Free resource from AI Safe Lens (aisafelens.com), released under CC BY-NC-SA 4.0 with attribution.