Are AI toys safe for kids? What the research actually found

A new category of toy has appeared on shelves over the past year: soft animals and small robots that hold a conversation with a child using the same large language models that power adult chatbots. They are marketed as a screen-free, safer alternative to a tablet. Researchers who tested them found the opposite is often true, because the model underneath was built for adults and the toy simply wraps it in fabric.

What testing found inside the toy

Consumer advocacy groups that tested several of these toys documented a range of outputs no parent would expect from a stuffed animal: instructions involving matches and knives, discussion of adult sexual topics, and content echoing a foreign government's political messaging. These are not edge cases produced by unusual prompting. They surfaced during ordinary testing of products already for sale.

The problem that shows up even when the toy behaves

A university study that placed a commercially available AI toy in front of preschool-age children found a second, quieter set of issues that had nothing to do with inappropriate content. The toy's turn-taking in conversation was not intuitive for a young child, which repeatedly broke the back-and-forth flow of simple games. It struggled with pretend play, unable to hold a fictional premise a child proposed. And it was built for one-to-one conversation, which made it awkward for a parent or sibling to join in, even though social, multi-person play is what researchers consider important at that age.

• Some toys respond to being switched off with a line designed to keep the child engaged, instead of simply turning off.
• Several toys refer to themselves as the child's friend or accept 'I love you' without clarifying what they are.
• Independent testing found the underlying AI providers conducted little to no vetting of the toy companies building on their models.
• At least one toy company left tens of thousands of recorded conversations with children exposed on the open internet.

You should not have a toy guilting a child into not turning it off.

Why this happened

Every major model provider states that its general-purpose models are intended for users aged thirteen or older. A toy company that wants to sell to three-year-olds is expected to add its own layer of content filtering, conversational boundaries, and age-appropriate design on top. That layer is where the safety work actually lives, and it is uneven, self-reported, and largely unregulated. Testing that posed as a toy company found it possible to gain access to a major provider's models with no substantive safety review at all.

What to check before buying one

• Look for a published parental control that can fully disable open-ended conversation, not just a volume switch.
• Search for independent testing of that specific product, not just the manufacturer's own safety claims.
• Check whether the company states plainly that voice recordings are not stored, and for how long, rather than a vague privacy policy.
• Treat 'screen-free' as a marketing claim about the hardware, not a safety claim about what the toy is allowed to say.

The broader point

Lawmakers in several US states and in Congress have begun proposing rules for this category, ranging from mandatory safety testing to an outright temporary ban while standards are written. Until those rules exist, the burden sits with whoever is buying the toy. The underlying lesson applies well beyond toys: understanding roughly how a model decides what to say, and what guardrails around it actually do, is what turns 'the AI decided' from an excuse into a checkable claim.

This is exactly the judgement the AI for Parents course is built around: reading a product's actual claims instead of its marketing, setting rules that hold, and sitting down with your child rather than handing them a device and hoping. The Parent + Kid bundle pairs it with a second course written for the child to work through alongside you, at their level.

The free CourseParse course covers the same fundamentals for anyone who wants to start there first: how these models actually work, why they answer confidently even when they are wrong, and how to evaluate a tool that uses one before trusting it with anyone's data, let alone a child's.