Pediatricians Say AI Is a Tool, Not a Friend. What That Means for Kids Under 5

On 4 March 2026, a team of pediatricians, psychologists and informatics researchers at Children's Hospital of Philadelphia published a review in the journal Pediatrics that tried to answer the question a lot of families are now asking at the kitchen table: what does conversational AI actually do to a developing child? Their headline conclusion was blunt. It is critical to emphasize that AI is a tool, not a companion, said lead author Dr Robert Grundmeier.

That one sentence carries more practical weight for a family with a 4-year-old than most of the regulation news of the past year. It is not a call to ban anything, and the review is careful never to endorse or condemn a specific product. It is a claim about how a young mind reads a voice that answers back, and about what the adults in the room should be doing while it happens. Here is what the review found, what it deliberately does not say, and what it changes if you are weighing up a talking toy for a preschooler.

What did the new Pediatrics review actually find?

The review found that the benefits and risks of conversational AI are not the same at every age, and that grouping all children together is the main mistake parents and policymakers make. It splits childhood into three bands: early childhood (0 to 5), middle childhood (6 to 11) and adolescence (12 and up). Each band gets its own opportunity and its own specific failure mode.

The paper, Generative Artificial Intelligence: Implications for Families and Pediatricians by Grundmeier and colleagues, is what researchers call a state of the review article. It pulls together evidence scattered across early studies rather than reporting one new experiment. Two themes run through it, as summarised in the CHOP announcement: these tools are powerful but imperfect and need human oversight, and guided engagement matters more than access.

Age band	What the review says can help	What it flags as the risk
0 to 5 years	Interactive AI storytelling can support language development and vocabulary, and can improve interaction between family members	Children may not reliably tell AI apart from human interaction, so human contact must stay the priority
6 to 11 years	Personalised learning that can close gaps in reading and maths, plus creative expression through writing and art	Difficulty spotting AI-generated misinformation, and the temptation to let it do the homework
12 years and up	Digital literacy, career and college exploration, and some evidence that companionship can ease loneliness	Dependence that displaces face-to-face contact, and poor responses to mental health questions

Summarised from Grundmeier et al., Pediatrics, published online 4 March 2026.

Why does the phrase tool, not a companion matter most before age 6?

Because under 6, the boundary between a person and a device that talks is genuinely blurry, and the review argues that a child who treats a machine as a friend may build an inaccurate mental model of what friendship is. That is a developmental concern rather than a technical one, and no software update fixes it.

Grundmeier put the mechanism plainly in an interview with WHYY: when you are interacting with these systems, although they can appear empathic and can in many ways pretend to be human, they fundamentally are not. Young children are excellent at reading warmth in a voice and much worse at asking where the warmth comes from. The practical implication is not that a talking toy is dangerous, but that a parent should never let it stand in for the human conversation it is meant to sit alongside.

The scale of the question is no longer hypothetical. Roughly two thirds of American teenagers had used an AI chatbot by late 2025 according to the Pew Research Center, and the CHOP team cites a 2025 study finding 72 percent of American adolescents have used one as a companion. Younger children are next.

Can a talking toy actually help a young child's vocabulary?

Yes, and this is the part of the review that rarely makes the headlines. For children aged 0 to 5, the CHOP team lists interactive AI storytelling as a genuine opportunity for language development and vocabulary, and notes it can improve interaction between family members rather than replace it.

Grundmeier gave a concrete example: an exhausted parent whose child wants an original bedtime story at the end of a long day. You can use a tool to help generate a story, then read it to your child yourself. The mechanism that does the developmental work is still the shared attention between the two of you. If you have ever wondered how a talking toy turns a child's question into an answer, the honest version is that the technology handles the words and the adult handles the relationship.

There is a real limit worth naming. Nobody has yet run the long-term studies that would tell us how a child who grows up with a conversational device at age 4 turns out at age 14. Grundmeier described the current moment as a natural experiment that we are all participating in, and called for organised, rigorous research. Any company, including ours, that claims certainty here is overselling.

What is never-skilling, and does it apply to preschoolers?

Never-skilling is the risk that a child never acquires a skill at all because the tool did it for them from the start. It is the younger sibling of de-skilling, where someone loses an ability they once had, and the CHOP team considers it the more worrying of the two in early education.

The distinction matters when you are choosing a toy. A device that answers a question and stops has done a small amount of never-skilling. A device that answers a question and then asks the child what they think, or invites them to finish the story, has done the opposite. This is one of the few design choices you can actually evaluate before buying: does the product try to end the exchange, or extend it? Children who are already asking dozens of questions a day are doing exactly the cognitive work you do not want outsourced.

What has actually gone wrong with AI toys already on the market?

Independent testing has found real failures, and they are worth stating plainly rather than talking around. Testers at the PIRG Education Fund reported AI toys giving children information they should never have received, including where to find dangerous household objects and how to light a match.

Legislators noticed. The Children's Artificial Intelligence Toy Safety Act of 2026 was introduced in the Senate, and more than a hundred AI-related bills have been filed across more than thirty states since December. The American Academy of Pediatrics has been building its own guidance in parallel.

None of this is a validation of any product, ours included. What it does show is that these failures came from design decisions: no age-appropriate content boundary, no parental control over topics, no limit on what the underlying model was allowed to say. Those are avoidable choices, not inevitable properties of the category.

What should a parent check before a young child talks to a device?

Those six questions are also the ones we designed Ted, our interactive plush companion, around. Ted has a microphone and a speaker and no camera. Parents choose the language, the topics, the tone and the limits from the Ted&Co app. There is no subscription and no advertising, and no data is sold. The full list of EU and US certifications, along with our GDPR and COPPA position, is on our security and certifications page, and the practical questions parents ask most are answered in the FAQ.

What Ted cannot do is replace you. On that point the CHOP review and we are saying the same thing. For a child under 6, the most valuable thing in the room is still the adult.

Frequently asked questions

Is AI safe for a 4-year-old?

There is no blanket answer, and the CHOP review does not give one. It says that for children under 6, human interaction should stay the priority and that adults should review AI-generated content alongside the child. Safety depends far more on the specific product's design and on adult supervision than on the age of the child alone.

At what age can a child tell the difference between AI and a real person?

The review notes that children in early and middle childhood, so roughly up to 11, may struggle to distinguish AI from human interaction. There is no clean cut-off age. What helps is naming it out loud: telling a young child that the toy is a machine that talks, and repeating it, does more than waiting for them to work it out.

Do AI toys help children learn to speak?

The evidence is early but promising for one narrow use: interactive storytelling for children aged 0 to 5, which the review links to language development and vocabulary. The gains reported so far come from richer conversation around the toy, not from the child talking to it alone.

Should I let my child call an AI toy their friend?

The CHOP researchers advise against encouraging it. Their concern is that a child who frames a machine as a friend may build an incorrect mental model of social relationships. Calling it a toy, a helper or a storyteller costs nothing and keeps the category clear in the child's head.

What is the single most useful thing a parent can do?

Co-listen. The review's most repeated recommendation for young children is shared review of AI-generated content: be there, hear what it says, and talk about it afterwards. It is also the cheapest safeguard available, and it does not depend on any manufacturer keeping a promise.

Sources: Grundmeier et al., Generative Artificial Intelligence: Implications for Families and Pediatricians, Pediatrics (2026); Children's Hospital of Philadelphia; WHYY; Pew Research Center; PIRG Education Fund; US Congress; American Academy of Pediatrics.