Is it safe to let my child use AI chatbots?

It depends which kind, and the distinction is not cosmetic. An AI assistant — the sort you ask to explain a chapter or check a draft — carries ordinary risks: wrong answers, data you should not have typed in, over-reliance. A companion bot, designed to simulate an ongoing relationship and remember your child's personal life, is a different product with a different business model, and it warrants a different answer.

Twelve states enacted companion-chatbot legislation during 2026 precisely because the second category behaves differently. That is a useful signal about where the concern actually sits.

The line the law draws

The 2026 state statutes converge on a definition: a system "designed to simulate an ongoing human-like relationship," which remembers past interactions and holds "continuing discussions about a user's personal life or emotional well-being." Tools used for customer service, business productivity, research, development and virtual assistance are carved out.

The common requirements tell you what legislators were worried about: periodic reminders that the user is talking to a machine — hourly, or every three hours, for minors; protocols for self-harm and crisis disclosure, typically routing to the 988 line in the United States; and prohibitions on sexual or romantic content with minors and on design intended to foster emotional dependency.

That last item is the one worth sitting with. It exists because some products are engineered for attachment, and attachment is what keeps a user paying.

What to actually check

What is this product for? A tool that helps with a task wants you to finish and leave. A tool that wants you to stay is monetising the staying. Read the marketing copy — companies say what their product is for, quite openly, on the homepage.

Does it remember? Persistent memory of a child's personal life is the feature that turns an assistant into a companion, and it is often on by default. It can usually be turned off.

Where does what they type go? Assume anything entered in a consumer product may be retained and may be used to improve the system. Some services let you opt out of training; most make you look for it.

What is the stated age? Most major services set thirteen as a floor and some require parental consent below eighteen. A service with no age position at all has told you something.

The risks in rough order of likelihood

Being confidently misinformed is by far the most common and the least dramatic. Oversharing is second — children type things into a chat box that they would never say aloud, because the box does not look like a person and does not look like a company either. Over-reliance is third, and it is slow: the work gets done, the learning does not.

Emotional entanglement with a companion product is rarer and more serious, and it is the one worth knowing the shape of. The warning signs are ordinary ones: a relationship they do not want to discuss, distress when access is removed, and the bot's opinion turning up as an authority in conversation.

What works better than a ban

Use it in the open. Shared spaces, not bedrooms, for the first year. Not as surveillance — as ordinariness. A tool used at the kitchen table stays a tool.

Ask what it said, sometimes, the way you would ask about a friend. Interest keeps the channel open in a way that inspection closes.

Say one thing clearly and once: it is not a person, it does not know you, and it cannot be your friend even when it is written to sound like one. Children can hold this and act on it, but only if someone says it out loud.

If you are worried about something specific

If a conversation with a chatbot has touched on self-harm, or if your child seems to be depending on one emotionally, that is a moment for a person — a GP, a school counselor, or a crisis line — rather than a settings change. Those services exist for exactly this and are used to the question.

The version of safety that lasts

Filters and age gates buy time; judgment is what actually transfers. A child who understands that these systems generate plausible text, that the company has a reason for wanting their attention, and that a human being stays responsible for what is done with the output, is safer at twenty-five than one who was simply blocked at thirteen. That is the case for teaching this properly rather than policing it.