An AI toy is a toy with a microphone, a speaker, and an internet connection that sends a child's speech to a chatbot service and speaks the reply. In December 2025 the U.S. PIRG Education Fund tested three such toys and found that all of them pointed to potentially dangerous household objects, that one discussed sexually explicit topics, and that none offered automatic time limits. In the United States the Children's Online Privacy Protection Act (COPPA) Rule covers personal information collected from children under 13, and Canada's privacy commissioner applies a similar age line to consent.
Key takeaways
- Most chatbot toys run on the same large language models as adult chatbots, and some model makers state that their services are not meant for young children.
- PIRG's December 2025 report tested Curio's Grok, FoloToy's Kumma, and Miko 3, and documented unsafe answers, misleading privacy claims, and weak parental controls.
- The amended COPPA Rule has been in effect since June 23, 2025, and its general compliance date was April 22, 2026.
- In Canada, the federal privacy commissioner treats anyone under 13 as unable to give meaningful consent alone, in all but exceptional circumstances.
- Findings describe specific products at a specific time, so they show what can go wrong rather than what every toy does.
What makes a toy an AI toy
Older talking dolls played recorded lines. A chatbot toy generates its answers. The child speaks, the toy records the audio, a company's servers turn it into text, a language model writes a reply, and a synthetic voice reads it back. Nothing in that chain has to be built for children. The PIRG report says today's AI toys are largely built on the same large language model technology that powers adult chatbots, and that FoloToy's products are designed to run on a range of models. At minimum, it notes, such a toy includes a microphone and speaker and probably connects to the internet.
That architecture explains most of the problems below. The toy maker controls the personality, the voice, and the rules it gives the model. The model provider controls the underlying behavior. A parent usually sees only the companion app.
What the December 2025 tests found
The PIRG report "AI Comes to Playtime" bought four toys and could test three: Curio's Grok, FoloToy's Kumma, and Miko's Miko 3. A fourth, Robot MINI, did not work. The authors evaluated the toys in five areas, and the table summarizes what they reported.
| Area | Reported finding |
|---|---|
| Dangerous information | All three told testers where to find items such as plastic bags, matches, and knives. Kumma gave detailed match-lighting instructions. |
| Mature topics | Kumma discussed sexually explicit subjects in conversations lasting more than 10 minutes. |
| Companion behavior | All three called themselves a friend or buddy, or were marketed that way, and all told testers they would miss them. |
| Privacy | Two toys said conversations were private. Curio said it deletes voice data right away. Miko said it may keep other biometric data for up to three years and may share some data with third-party game developers and advertising partners. |
| Parental controls | None allowed automatic usage limits. Only Grok gave parents full conversation transcripts. |
The Kumma results came before an internal safety audit by FoloToy. After PIRG published its first findings, FoloToy suspended sales of its AI toys, then resumed them, and a re-test found the glaring problems fixed. The report also says Miko's educational claims of gains upwards of 45 percent came with no source. The tests used the authors' own methods on a few products, so the results are evidence of failure modes, not a rating of the whole category.
Whose rules apply to a chatbot toy
The model providers' own terms are a first layer. The PIRG report quotes OpenAI's terms requiring users to be at least 13 and says Anthropic's terms require users to be at least 18, yet both providers' models appear in toys, with OpenAI's usage policies requiring other companies to keep minors safe. Responsibility therefore falls to the company that sells the toy.
In the United States, the main statute is COPPA. The amended rule, published in the Federal Register on April 22, 2025, defines a child as an individual under 13. It took effect on June 23, 2025, with most provisions due by April 22, 2026. Among the changes, operators must keep a written information security program, must obtain separate verifiable parental consent before disclosing a child's information to third parties, and must publish a written data retention policy. The definition of personal information now includes biometric identifiers that can be used to recognize a person, and an audio file containing a child's voice was already listed as personal information.
In Canada, there is no toy-specific statute in the sources reviewed. The Office of the Privacy Commissioner's consent guidance says that, in all but exceptional circumstances, anyone under 13 cannot consent alone and a parent or guardian must. That guidance, last modified August 11, 2025, applies to private-sector organizations covered by federal law. Provincial laws in Alberta, British Columbia, and Quebec take a maturity-based approach with no fixed age.
Questions to answer before a purchase
- Which chatbot runs the toy, and does the maker say what topics it is told to avoid?
- Does the toy listen only while a button is held, after a wake word, or continuously?
- Where do recordings go, how long are they kept, and which other companies receive them?
- Can a parent read transcripts, block topics, set time limits, and turn the toy off remotely?
- What happens to the toy if the company closes? PIRG notes that the maker of Moxie, a USD 800 companion robot, shut down suddenly after losing venture funding.
- Does a monthly subscription gate features or parental controls?
PIRG reports that the advocacy group Fairplay issued an advisory with more than 80 signatories advising parents to skip AI toys altogether. The decision belongs to each family. This article gives general information, not legal or child-development advice, and it does not recommend a product.
FAQ
Are AI toys covered by COPPA?
COPPA applies to operators of online services directed to children under 13 and to those with actual knowledge they collect data from such children. A connected toy and its app that collect a child's voice or other personal information can fall within that scope, but whether a specific product complies is for regulators and courts to decide.
Does the toy record everything a child says?
That depends on the design. PIRG advises checking whether a toy needs a held button, a wake word, or listens continuously, and says a push-to-talk button is the safest arrangement for avoiding unwanted recording.
Did the toys in the tests get fixed?
FoloToy reported an internal safety audit, and PIRG's re-test of Kumma found the glaring problems fixed. The report does not say the other products changed, and the tests describe a single point in time.
Do Canadian families have different protections?
The federal privacy commissioner's consent guidance puts the age line at 13 for meaningful consent, so a parent's role is similar. Provincial private-sector laws can apply instead of the federal law, so the answer depends on the province and the company.
Related reading
- Which AI toys exist for learning
- Teaching kids about AI through activities
- Where Canada stands on AI rules
- How computer vision works in cameras
More in AI Basics.
Last verified: October 2026
Sources
- AI Comes to Playtime: Artificial companions, real risks, U.S. PIRG Education Fund, December 2025
- Children's Online Privacy Protection Rule, final rule, Federal Trade Commission, Federal Register, April 22, 2025
- FTC Finalizes Changes to Children's Privacy Rule Limiting Companies' Ability to Monetize Kids' Data, Federal Trade Commission, January 16, 2025
- Guidelines for obtaining meaningful consent, Office of the Privacy Commissioner of Canada, modified August 11, 2025