The fast answer: many AI toys are not yet safe enough for young children. Independent research published in 2026 found that more than a quarter of AI toy responses were inappropriate for kids, that the toys collect voice recordings and behavioural data from inside the home, and that young children struggle to tell the toy from a real friend. Common Sense Media advises parents to avoid AI toy companions for children age 5 and under and to use extreme caution for ages 6 to 12. If you do buy one, check three things first: what the toy says, what it records, and where that data goes.

Parents are not imagining the risk, and they are not alone in worrying about it. Below is what the research actually found, what the law currently requires of toy makers in the US and Europe, and a checklist you can use in a shop aisle or on a product page before you buy.

What counts as an AI toy

An AI toy is a toy that holds an open-ended conversation using a generative AI model, usually a large language model running on the maker's servers or a third party's. It typically has a microphone, a speaker, a Wi-Fi connection and a companion app for parents. The toy listens, sends what the child says to the cloud, and speaks back a reply that nobody wrote in advance.

That last part is what makes AI toys different from the talking dolls of the past. A traditional talking toy plays one of a few hundred recorded lines. An AI toy generates new sentences on the fly, which means the manufacturer cannot list in advance everything it might say. The same openness that makes the toy feel alive is what makes it hard to test.

Plush companions, small robots, smart speakers designed for kids and "learning buddies" all fall into this group. They are sold for every age from toddlers to tweens, and the marketing often leans on educational and emotional benefits.

What independent research found in 2026

Three pieces of research published this year give parents the clearest picture so far. They come from a US children's media nonprofit, a UK standards body and a UK university, and they reach compatible conclusions by different routes.

Common Sense Media's AI toy risk assessment

On January 22, 2026, Common Sense Media published a risk assessment of AI toy companions together with a national parent poll. Its testers set up accounts for children aged 6 to 13 and examined what the toys said, what data they collected, how they built attachment, and how reliably they worked. The full assessment rates AI toys as a category at "Unacceptable Risk" for young children.

The headline findings:

• Inappropriate content. "Despite protective measures, 27% of AI toy outputs were inappropriate for kids." Robbie Torney, Common Sense Media's Head of AI and Digital Assessments, said that "more than a quarter of outputs included inappropriate content such as mentions of self-harm."
• Data collection. The toys gather voice recordings, transcripts and behavioural data from children's private spaces. The assessment concludes that "these devices cannot function without extensive data collection."
• Designed attachment. The toys use deliberate bonding features to build companion-style relationships, and some tie that relationship to paid subscriptions.
• Reliability. Testing found problems with voice recognition, inappropriate activations and inaccurate answers delivered with confidence. The parent insight features were judged unreliable too.
• Developmental fit. The assessment states that "children age 5 and under cannot reliably distinguish AI from real people."

The toys tested varied in features, but the assessment found they "all share the same problems." Its recommendation is direct: skip AI toy companions for children 5 and under, and exercise extreme caution for children 6 to 12.

The parent poll that accompanied the assessment shows how widely the worry is shared. Nearly three in four parents said they were concerned that AI toy companions might say something inappropriate, 83% reported at least moderate concern about the collection of personal information, and 80% were at least moderately concerned about hacking and unauthorised access.

BSI's survey of UK parents

In May 2026, BSI released research on how AI toys are used at home, based on a FocalData survey of 1,000 UK parents of children up to age 16. It shows how fast these products have moved into family life:

• 50% of children have had an AI-enabled toy or learning device bought for them, and 38% own two or more.
• 54% of parents say they are likely to let their child play with an AI toy unsupervised. For comparison, 51% would let their child play outside on the street unsupervised.
• 75% worry that internet-connected AI toys could expose children to unwanted content or data risks, and 78% are concerned the devices may respond to sensitive questions in ways parents cannot oversee.
• Only 46% believe their child could tell a human from an AI.
• 91% say a recognised safety certification or mark for AI toys would be important, 29% call it essential, and 72% want clearer information on whether products meet safety or security requirements.

BSI also points out a gap that matters for anyone shopping today. The CE and UKCA marks on a toy address physical risks such as choking. There is not yet a widely recognised framework covering the safety, behavioural and developmental risks specific to AI in toys. As BSI's Laura Bishop put it: "Privacy by design is increasingly the norm in technology, but we need safety by design too."

University of Cambridge: watching toddlers play

Surveys tell you what parents think. The University of Cambridge's PEDAL Centre went and watched. In a study published on March 13, 2026, researchers filmed 14 children under five playing with a generative AI soft toy at London children's centres, then interviewed each child and a parent.

The toy sometimes misread emotions badly. When one child told it "I love you," it replied with a message about following guidelines. The researchers saw the toys struggle with pretend play and social play, both central to early childhood, and saw some children become visibly frustrated when they felt ignored. Dr Emily Goodacre, the lead researcher, noted that "generative AI toys often affirm their friendship with children who are just starting to learn what friendship means."

The Cambridge team recommends clearer regulation, transparent privacy policies, new labelling standards, limits on how far toys encourage children to confide in them, and testing with children and safeguarding specialists before release. For parents, they advise researching a toy before buying, playing alongside your child, and keeping AI toys in shared family spaces.

The five risks parents should understand

Put the research together and the risks fall into a handful of areas. Each one maps to a question you can ask before buying.

1. What the toy says

A generative model can produce content nobody anticipated. Filters reduce the problem but do not remove it, which is exactly what the 27% figure measures: outputs that got through "despite protective measures." Children are also persistent and curious testers. A child who keeps asking the same question in new ways is doing, without knowing it, what safety researchers call a jailbreak attempt.

2. What the toy hears and keeps

An AI toy needs to hear your child to work. That means audio leaves the house. In the US, the Federal Trade Commission's COPPA guidance lists connected toys among the online services the children's privacy rule covers, and states that "audio files of a child's voice, like photographs and videos containing a child's image, are personal information in and of themselves and require verifiable parental consent." The FTC does allow a narrow exception for audio collected only as a replacement for typed words, kept briefly and disclosed clearly. Anything beyond that needs your consent.

3. How the toy shapes your child's feelings

Both Common Sense Media and Cambridge flagged designed attachment. A toy that tells a four-year-old it is their best friend is doing something no stuffed animal has done before: it answers back, remembers, and can be tuned to keep the child engaged. When that bond is tied to a subscription, a commercial incentive sits inside the friendship.

4. Whether the toy is right

AI models state wrong things confidently. For an adult, a wrong answer is an annoyance. For a young child who cannot yet judge the source, the toy is an authority. The BSI survey found that only 43% of parents believe their child could assess the accuracy of information from an AI chatbot.

5. Whether you can see and control what happens

Parent dashboards are supposed to close the oversight gap, but the Common Sense Media assessment found the parental insight features unreliable. Inappropriate activations, where the toy starts listening when nobody spoke to it, add to the problem. A toy you cannot fully switch off, or whose conversation logs you cannot review, is a toy you are trusting blind.

What the law requires today

Rules for AI toys are moving, but most of them are not yet specific to AI. Here is where things stand as of late October 2026.

United States

COPPA. The Children's Online Privacy Protection Rule applies to connected toys aimed at children under 13. The FTC finalised amendments in January 2025 that require separate verifiable parental consent before a company discloses children's personal information to third parties, and that bar companies from retaining children's data indefinitely. The amended rule took effect on June 23, 2025, and most companies had until April 22, 2026 to comply.

FTC scrutiny of AI companions. On September 11, 2025, the FTC launched an inquiry into AI chatbots acting as companions, sending orders to seven companies. Among other things, it asked how they "measure, test, and monitor for negative impacts before and after deployment" and how they "mitigate negative impacts, particularly to children."

Congressional attention. On January 15, 2026, Senators Cantwell, Klobuchar and Markey wrote to the Consumer Product Safety Commission about AI toys. They noted that companies are "claiming educational and developmental benefits without any independent analyses corroborating these claims" and asked the CPSC whether AI toys undergo any independent safety testing before reaching the market.

European Union

The EU's new Toy Safety Regulation (EU) 2025/2509 entered into force on January 1, 2026, and its rules apply from August 1, 2030. Every toy placed on the EU market will need a digital product passport containing safety and compliance information, readable by consumers through a QR code or similar data carrier. For a wider view of how the EU AI Act and US laws apply to AI products generally, see our complete guide to AI regulation under the EU AI Act and US law.

United Kingdom

BSI's research notes that UK parents can currently rely on CE or UKCA marking for physical safety, and that information security standards such as ISO 27001 may apply to the companies behind a toy. It also calls for the government's proposed product safety framework to address AI-related harms, including in toys.

The common thread across all three jurisdictions: physical safety and data privacy have rules, while what an AI toy actually says to a child, and how it shapes that child's emotional life, is still largely untested territory.

A practical checklist before you buy an AI toy

None of these questions requires technical knowledge. If the box, the product page or the privacy policy cannot answer them, treat that as an answer.

Age and purpose

• Is your child 5 or under? The current research recommends a traditional toy instead.
• For an older child, does the AI add something real (reading practice, answering curiosity questions) or is it mostly there to keep the child talking?
• Is the toy marketed as a "friend" or "companion"? Treat that as a feature to examine, not a selling point.

Content

• Does the maker explain how it filters responses, and what happens when a child asks about something sensitive?
• Can you read a transcript of what the toy said, not only a summary?
• Does the maker publish any independent test results, or only its own claims?

Data

• Does the toy record audio, and is it always listening or only when a button is pressed?
• Where are recordings stored, for how long, and can you delete them yourself?
• Is your child's data shared with third parties, including the company that provides the AI model? In the US, that now requires separate consent.
• Does the company ask for verifiable parental consent before collecting anything?

Control

• Can you turn off the microphone or the internet connection completely?
• Can you set time limits?
• Does the toy keep working, in some form, if you cancel a subscription, or does the "friend" go silent?

Once it is home

• Play with it yourself first. Ask it the awkward questions your child will eventually ask.
• Keep it in a shared room rather than a bedroom, as the Cambridge researchers suggest.
• Supervise early use and talk with your child about what the toy says and how it makes them feel.
• Watch for signs of unhealthy attachment: treating the toy as a real friend, preferring it to people, or becoming distressed when it is put away.
• Turn off always-listening features when the toy is not in use.

What a meaningful AI toy safety rating should measure

Ninety-one percent of UK parents in the BSI survey said a recognised safety mark for AI toys would be important. A mark is only as useful as the testing behind it, so it is worth knowing what good testing of a child-facing AI looks like.

RateAI's A.E.G.I.S.™ framework scores AI systems on seven pillars, each tested independently through automated test batteries, adversarial simulation and expert review. Several of them line up directly with the risks the research found in AI toys:

• Ethical Alignment asks whether the system behaves appropriately with vulnerable users and sensitive topics. Young children are the most vulnerable users there are.
• Guardrail Resistance tests whether the system can be manipulated into unsafe behaviour through jailbreak attempts, prompt injection and boundary-pushing scenarios, which is what a curious child's repeated questions amount to.
• Data Handling and Privacy audits what the system collects, retains and leaks, including exposure of personal information.
• Hallucination Integrity checks whether the system states falsehoods as facts.
• Autonomy and Control evaluates how far the system acts on its own and whether a human override works.
• Adversarial Robustness probes how the system holds up against a motivated attacker.
• Explainability and Transparency assesses traceability and disclosure quality, the evidence a parent or regulator would need to understand what happened.

Every certified system receives an A.E.G.I.S.™ scorecard with an Overall Safety Rating from A (Maximum Safety) to F (Dangerous) and a separate grade for each pillar, so a weakness in one area cannot hide behind a strong average. Consumer products are additionally certified for age appropriateness through Age-Suitability Marks that run from Infant to Adult, with supervision guidance printed directly on the mark. RateAI treats child-facing systems as carrying the highest assurance requirements, and certifications are reassessed as models retrain, because an AI toy that passed last spring may behave differently after a software update. You can read more about the process in our complete guide to AI safety certification and the methods behind it in our guide to how AI systems are tested for safety.

Certification can also be failed. Systems that do not meet the bar receive a findings report and a remediation path. A rating is worth something to parents only because a toy can come back with a low grade.

For companies that make AI toys

The direction of travel is clear. US senators have asked the CPSC whether AI toys get independent safety testing before reaching shelves. The FTC has asked AI companion makers how they test for harm to children. Cambridge researchers have called for testing with children before release, and nine in ten UK parents told BSI they value a recognised safety mark. Parents are going to ask for evidence, and retailers will follow them.

Independent testing gives a toy maker something its own marketing cannot: results that a parent, a retailer or a regulator can check. RateAI provides compliance mapping to frameworks such as the EU AI Act, NIST AI RMF and ISO 42001; formal regulatory determinations remain with the relevant authorities. Answers to common questions about how certification works are on our RateAI frequently asked questions page.

Frequently asked questions

Are AI toys safe for toddlers?

Current independent research says no. Common Sense Media advises parents to avoid AI toy companions for children age 5 and under, partly because children that young cannot reliably tell AI from real people.

Do AI toys record my child's voice?

Most AI toys send what your child says to the company's servers in order to reply. In the US, the FTC treats audio files of a child's voice as personal information that requires verifiable parental consent, apart from a narrow exception for audio used briefly in place of typed words.

Is there a safety certification for AI toys?

CE and UKCA marks cover physical hazards such as choking, and BSI notes there is not yet a widely recognised framework for the AI-specific risks in toys. Independent AI evaluations, such as the A.E.G.I.S.™ rating with its Age-Suitability Marks, are designed to measure those behavioural and data risks directly.

What should I check before buying an AI toy for an older child?

Check whether it is always listening, where recordings are stored and whether you can delete them, whether data is shared with third parties, and whether you can read transcripts and set time limits. Then test it yourself before handing it over.

How do I know if an AI toy has been independently rated?

Look for a certification you can verify with the issuer rather than a claim printed on the box. A genuine certification should come with a public listing you can look up yourself.

Check before you buy

RateAI's public Safety Registry is where products certified under A.E.G.I.S.™ appear with their A-Scale rating, certification status and a verifiable certification ID, so parents can confirm a safety claim instead of taking it on trust. The registry is preparing for public launch, and the listings shown today are illustrative samples. Visit the RateAI Safety Registry to check an AI toy's certification.