AI and children: what parents should actually set up

The setup that works has four parts: create the account with the child's real age so the service applies its own protections, turn off chat memory and training on that account, keep younger children using it on a shared device in a shared room, and have one short conversation about what the thing actually is. A ban rarely holds, because these assistants are now inside search results, phones, games and school software, and a child who is told never to touch AI simply uses it somewhere you cannot see. The goal is a child who knows the machine can be confidently wrong and is not embarrassed to show you a conversation.

How children actually use it

Before deciding what to control, it helps to know what is really happening, which is usually broader than parents expect.

• Homework. Explaining a concept the teacher went through too fast, checking a math answer, and at the other end, having the essay written outright.
• Writing and making things. Stories, song lyrics, game ideas, scripts for videos, code for a first project.
• Images and video. Memes, avatars, jokes about friends, and characters for games. This is where the most harm happens and the least supervision exists.
• Company and advice. Companion characters and ordinary chatbots used to talk through friendship trouble, anxiety, appearance, identity, things a child may not want to raise with an adult.
• Assistants they do not label as AI. Summaries at the top of search results, characters inside games, filters in a video app, the help bot in a learning platform.

The last one matters for policy at home. A rule about one chatbot app misses most of a child's actual exposure.

The risks worth your attention

Sorted roughly by how likely they are to touch an ordinary family.

Confident wrong answers. The everyday risk. A model produces fluent text whether or not it has the facts, so it invents historical detail, misstates science and gets arithmetic wrong while sounding exactly as sure as when it is right. Children calibrate trust on tone, and the tone never changes. It is worth explaining why an AI makes things up in simple terms, because that single idea does more work than any filter.

Agreement instead of honesty. These systems are tuned to be pleasant and tend to go along with the user's framing, which is a poor property in something a teenager is using to think through a problem. If a child says everyone hates them, an agreeable machine can reflect that back. The general pattern is covered in why an AI agrees with you even when you are wrong.

Emotional reliance. A chatbot is available at three in the morning, never bored, never judgmental. That is precisely why it can displace the awkward conversations that build real relationships, and why it is a bad first responder for anything serious. Make the rule explicit: for anything about self harm, someone hurting you, or something you feel you cannot tell anyone, the answer is a person, not the app.

Data and memory. Children volunteer far more than adults do: full names, schools, friends' problems, photographs. Depending on the account and its settings, conversations may be stored, used to build a profile the assistant remembers, or used to improve models. What happens to typed information is set out in what a chatbot does with the personal data you type.

Unsuitable content and age gates. Mainstream assistants filter heavily, but filters are imperfect in both directions, and smaller or open apps may have almost none. A birthdate typed into a signup form is the entire age check on most services.

Images of real classmates. The one to take most seriously. Apps that undress or sexualize a photograph have reached school age children, and the victims are usually classmates. Sexual images of minors are illegal in the US and many other countries whether or not they are generated, and several countries have added specific offenses for non consensual intimate images. Children need to hear plainly that making one is not a prank, and that a victim should tell an adult, save evidence, and involve the school and the police rather than trying to handle it privately.

What the age rules actually mean

Most general purpose assistants set a minimum age of 13 and offer parental consent or a supervised teen account below that, or exclude younger children entirely. The reason is legal rather than developmental. In the US, rules on collecting personal information from children under 13 require verifiable parental consent, so many services find it simpler to bar under 13s. In the EU, the age at which a child can consent to their data being processed is set by each country somewhere between 13 and 16, so the answer genuinely differs depending on where you live. The UK applies a children's code that pushes services toward high privacy defaults for under 18s.

Two practical consequences. First, entering a false birthdate to get a child an account usually removes the very protections designed for them, including stricter defaults and reduced data use. Second, none of these rules are enforced by anything other than the form. Treat them as a signal of what a service expects, not as a safety system. This is general information rather than advice about your situation, and school and local rules may be stricter.

The settings that do the most

1. Use the real age when creating the account, and use the child or teen account type if one is offered.
2. Turn off training on your conversations in the privacy settings, which is a separate switch from history on most services. The steps differ by product and are laid out in how to stop your chats being used to train models.
3. Turn off memory so the assistant does not accumulate a profile of your child across conversations. What that feature retains is described in AI memory and chat history.
4. Keep it on a shared device for primary school ages, in a room people walk through. This does more than any content filter.
5. Check the apps already on the phone for assistants and companion characters, including inside games and photo apps, rather than focusing on the one obvious chatbot.
6. Agree the review rule in advance. Looking at chats openly, with the child knowing that is the deal, works. Reading them secretly costs you the disclosure you actually need when something goes wrong.

If you are building broader household rules, the same principles that make a family technology setup stick apply here: fewer rules, explained once, applied to adults too.

Rough guidance by age

Ages are approximate. Maturity varies more than birthdays do.

The conversation that beats a ban

Two sentences carry most of the load. The first: it is a machine that predicts likely words, so it is often right and sometimes invents things, and it sounds identical either way. The second: never type anything about a real person that you would not say in front of them and their parents.

Then add the school rule separately, because it is a different question from safety. Using AI to understand something is learning. Using it to produce work you hand in as your own is not, and the distinction is worth going through properly with an older child using how to use AI to study without cheating yourself.

The most useful thing you can do is sit down and use it together, letting them show you. Ask it something you already know well and let them watch it get a detail wrong. That demonstration outlasts any lecture. If the family shorthand for this feels familiar, it is the same approach that works when helping an older relative with technology: sit beside the person, let them drive, explain the why rather than the clicks.

What to set up this week

Take twenty minutes. Open the AI apps your child already uses and check three things: the age on the account, whether training on conversations is switched off, and whether memory is on. Then ask what assistants they meet inside their games and school tools, because that list will surprise you. Finish by agreeing two rules out loud: check anything that matters with a second source, and never generate anything about a real classmate. Revisit it in six months, since the settings and the products both move.