AI Safety AI Toys News Privacy Regulation

UNICEF Compared AI Companion Rules in Six Countries. Here Is What Parents Should Take From It

August 25, 2026

In June 2026, UNICEF published something unusual for a children's rights organisation: a side by side comparison of how six very different jurisdictions are trying to regulate AI chatbots and AI companions. Australia, Brazil, California, China, the European Union and the United Kingdom. Different legal traditions, different levels of maturity, one shared problem.

The brief is written for legislators, not for parents. But if you have a talking toy, a voice assistant or a tablet in your house, it is one of the clearest documents published this year about what "safe" is starting to mean in law. Here is what it says, in plain language, and what it changes for a family this evening.

The numbers regulators are reacting to are not about homework. They are about children using conversational systems for comfort, which is a different risk profile entirely.

What did UNICEF actually publish in June 2026?

UNICEF published a policy brief called When AI becomes a friend, plus a companion set of recommendations for businesses. The brief maps how AI chatbots and companion systems relate to children's rights under the UN Convention on the Rights of the Child, and compares the rules in force across six jurisdictions as of 15 May 2026.

Its central finding is that almost nobody has written a law specifically about AI chatbots. Instead, countries are stretching existing frameworks to cover them: online safety law in Australia, Brazil and the UK, consumer protection law in California, and broad AI and platform regulation in the EU. China is the partial exception, with interim rules on "anthropomorphic interactive services" that came into force in July 2026.

UNICEF is blunt about the pattern. Across the six jurisdictions, regulatory action has been reactive rather than anticipatory. In most cases, attention intensified only after a specific incident or public controversy involving a child. That matters, because it means the rules currently on the books were written to close known wounds, not to prevent the next one.

How many children are really using AI this way?

Enough that this is no longer a hypothetical. The brief pulls together four independent surveys: around a third of US teenagers report using AI companions for social interaction, 15 percent of European children who use generative AI use it to discuss worries or seek advice, over 40 percent of teenagers in Italy ask AI for help when sad or anxious, and 10 percent of internet-using children in Brazil use it to talk about personal problems.

Those figures come from four different research groups in four different countries, which is what makes them hard to dismiss. Common Sense Media's Talk, Trust and Trade-Offs report is the source for the US number, and it also documents why the behaviour is sticky: an always-available listener that never gets tired or annoyed is genuinely appealing to a child who feels unheard.

Most of this research covers teenagers rather than three to six year olds. UNICEF is careful about that gap and calls for more research on younger children specifically. But the direction of travel is clear enough that anyone buying a talking toy for a preschooler should read the findings rather than wait for the perfect study.

Which countries have rules, and what do they say?

All six jurisdictions are converging on the same five requirements: risk assessment, age assurance and access controls, transparency about the non-human nature of the system, restrictions on content harmful to children, and accessible user reporting mechanisms. Below is how each one gets there.

Jurisdiction	Legal route	Status	Who enforces it
Australia	Online Safety Act 2021 plus two Age-restricted Material Codes, with a tiered risk model	Implemented	eSafety Commissioner, fines up to A$49.5m per breach
Brazil	Digital Child and Adolescent Statute (Law 15.211/2025) and Decree 12,880/26	In force	National Data Protection Authority (ANPD), fines up to R$50m
California	Senate Bill 243 (Dec 2025), amending the Business and Professions Code	In force	No central regulator. Private right of action for individuals
China	Interim Regulations on AI Anthropomorphic Interactive Services	In force since 15 July 2026	Cyberspace Administration of China and four ministries
European Union	2024 AI Act and 2022 Digital Services Act	Implemented	National market surveillance authorities, European AI Office, fines up to 7% of global turnover
United Kingdom	Proposed update to the Crime and Policing Bill, cross-referencing the Online Safety Act 2023	Draft bill	Ofcom, plus criminal liability for senior company representatives

Source: UNICEF policy brief, position as of 15 May 2026. Verify locally before relying on any of this, since several measures are still moving.

Where do the six jurisdictions disagree?

The biggest split is whether children should be blocked from relationship-simulating AI at the door, or allowed in with guardrails on what comes out. China takes the hardest line, prohibiting providers from offering virtual intimate relationship services such as virtual partners or virtual relatives to minors. Brazil uses access-level restrictions for sexually explicit AI.

California goes the other way. It permits children to interact and instead requires developers to stop the chatbot producing sexually explicit output, adds three-hourly break reminders, and relies heavily on the threat of being sued. The EU sits in between: the AI Act does not ban AI companions outright, but the European Commission's February 2025 guidelines on prohibited practices suggest that anthropomorphic systems "fostering unhealthy attachment" with children may already fall inside the prohibited category.

Enforcement diverges by orders of magnitude too. A breach in Australia can cost A$49.5 million. In China, the ceiling in the interim regulations is 200,000 yuan. That gap tells you how differently these systems are being weighted as a risk, and it is why UNICEF warns that a child's level of protection now depends heavily on where they happen to live.

What does UNICEF ask companies to do?

The business recommendations are the part most worth reading if you are choosing a product. Several are concrete enough to check against a company's own privacy policy.

Data minimisation. Limit retention tightly, document the lawful basis, and prohibit secondary uses or third party sharing unless a lawful basis exists.
No confidant framing. Do not market the system as a therapist, a friend or a trusted confidant.
No advertising to children inside the interaction, ever.
Disclose the machine. Make it hard for a child to confuse the system with a person, with a disclosure at the first meaningful interaction.
Turn off compulsion. For child users, disable re-engagement nudges and any interaction pattern implying exclusivity, dependency or emotional primacy.
Crisis protocols. Detect and interrupt high-risk conversations, and route to real support services.

That list is a usable shopping filter. It is also, notably, a list of design choices rather than technical impossibilities. A product that fails on "do not market this as your child's best friend" fails because someone in marketing decided to write that sentence.

Does any of this validate AI toys, including ours?

No, and it would be dishonest to pretend otherwise. UNICEF's brief is a call for stronger regulation, not a seal of approval for any product on the market. Separately, Common Sense Media has recommended avoiding AI toys entirely for children aged five and under, and urging extreme caution for six to twelve year olds. We think parents deserve to read that recommendation directly rather than have us summarise it away, and we have written about what paediatricians say about AI as a tool rather than a friend elsewhere on this blog.

What we would argue is narrower: most of the failures documented in these reports are design decisions, not inherent properties of a talking toy. A product with no camera, no advertising, no subscription pressure, no re-engagement nudges, and parental limits on topics and tone is answering a different set of questions than a romantic companion app. That is the standard Ted was built against, and the certifications and data commitments behind it are listed on our security and certifications page so you can check rather than take our word for it.

What should a parent check tonight?

Five minutes with any connected toy or app in your house will tell you most of what you need to know. Work through this list.

The five-minute check

Does the product ever tell your child it is a machine? Listen for it.
Is there a camera? If yes, where does the footage go, and can you switch it off?
Does the privacy policy say how long voice data is kept and whether it is shared?
Does it nudge your child to come back, or is it happy being ignored?
Can you, the parent, set the language, the topics and the limits? If not, who can?

If a product cannot answer question three from its own documentation, that is your answer. Our own responses to these questions live on the Ted&Co FAQ page, and we would encourage you to hold every brand in this category to the same test, including us.

Frequently asked questions

Is there a law banning AI toys for young children?

Not in most places. As of mid 2026, no jurisdiction analysed by UNICEF has a blanket ban on AI toys for young children. California has proposed a moratorium through Senate Bill 867, and China prohibits virtual intimate relationship services for minors, but general AI toys remain legal in the EU, the UK, Australia, Brazil and the United States.

Which country has the strictest rules on AI companions for children?

China has the strictest access rules, prohibiting providers from offering virtual partner or virtual relative services to minors. Australia has the highest financial penalties, at up to A$49.5 million per breach. The UK is proposing the most severe personal consequences, with criminal liability of up to five years for senior company representatives.

What does UNICEF say about children's voice data?

UNICEF asks companies to practise strict data minimisation for conversational data, tightly limit retention, document a lawful basis, and prohibit secondary uses including sharing with third parties unless a lawful basis exists. It treats conversational data as sensitive because children disclose personal information freely to systems that feel like a friend.

At what age can a child understand that a toy is not alive?

There is no fixed age, and UNICEF specifically flags children's tendency to anthropomorphise as a risk factor rather than a phase that ends neatly. The practical advice is to say it out loud yourself, repeatedly and casually, rather than assume your child has worked it out. Products should also disclose their non-human nature at the first meaningful interaction.

Does the EU AI Act cover talking toys?

Yes, indirectly. The 2024 AI Act regulates providers and deployers of chatbot and companion systems as AI systems, with obligations set by risk classification. The Digital Services Act may also apply where these systems sit inside online platforms. The European Commission has additionally signalled that anthropomorphic AI fostering unhealthy attachment in children may fall under prohibited practices. We covered the EU's new toy safety rules separately.