In August 2026, a research team at Harvard did something no toy company had done publicly: they built their own conversational teddy, sat children down with it, and started recording what happened. The lab is the Harvard Child-Centered AI Lab, led by Ying Xu of the Harvard Graduate School of Education, and the toy is not for sale. It exists only to be studied.
That detail matters more than it sounds. Until now, almost everything parents have read about talking toys came from two places: marketing copy written by the people selling them, and safety tests run on products already in shops. A purpose-built research toy sits in between. Here is what the study is actually measuring, what an earlier British report already found, and what any of it changes for a parent deciding what to put under the Christmas tree.
What exactly did Harvard announce?
Harvard's Child-Centered AI Lab has built an AI-enabled toy and begun testing it with children in a controlled lab setting. This is phase one of a two-phase study. The lab phase compares how children play with the conversational toy against how the same children play with ordinary, non-talking toys.
Xu announced the launch in August 2026, and the reporting was picked up by EdTech Innovation Hub. A planned second phase would move the toy into participating families' homes for longer observation, which is where the interesting data usually lives. The lab has not published a start date or a sample size for that phase, so treat any claim about home results as premature.
Alongside the experiment, Xu's team with co-authors Lauren Girouard-Hallam and Zhonghao Shi produced a literature review, "Artificial Intelligence in Toys: Implications for Child Development and Play", pulling together what the existing evidence already says. Separate work from the same broad research community, including the paper "Toys that listen, talk, and play", has been looking at how children make sense of a toy that answers back.
Why build a toy instead of buying one off the shelf?
Because a commercial toy is a moving target. Its model gets updated, its safety filters change overnight, and the company controls what the researchers can see. A research toy can be frozen, logged, and varied on purpose.
That lets the team isolate the thing everyone actually argues about. Does the conversation itself change how a child plays, or is any difference just novelty? Researchers can hold the plush shell constant and switch the talking on and off. They can vary how much the toy leads versus follows. None of that is possible with a product that is simultaneously being sold, patched and defended.
It also removes a conflict of interest. Independent evidence about talking toys has been thin precisely because the companies with the best data have the least incentive to publish it.
What did the Cambridge report already find?
It found real problems, and they are worth reading before any marketing. The University of Cambridge's AI in the Early Years project published the first systematic look at how generative toys affect children under five, and the results were not flattering.
Researchers logged repeated moments where the toy misread a child's emotional display, or shut down exactly the kinds of play that matter most at that age. Social play and pretend play are not decoration; they are how a preschooler rehearses other minds. A toy that keeps answering literally when a child is inviting it into a story is quietly getting in the way. The report called for tighter regulation and a recognisable safety mark, and argued that designers should be working alongside child development specialists rather than shipping first.
We think that criticism is fair and we do not think it validates any product, including ours. What it does say is that most of the failures described are design choices, not laws of physics. A toy that pauses instead of interrupting, that hands the story back to the child, that has an adult deciding what it will and will not discuss, is a different object from one tuned to maximise engagement.
What do parents themselves actually say?
Parents are curious and unconvinced at the same time. Common Sense Media, with fieldwork by SSRS, surveyed a nationally representative sample of 1,004 US parents of children aged 0 to 8 in December 2025, and the pattern in the answers is remarkably consistent.
Read the two blue bars against the rest. Only 22 percent want a toy that acts as a companion or emotional support, while 70 percent want one that helps their child learn. Parents are not asking for a friend. They are asking for a tool that behaves itself. The full figures are in the AI in the Toy Box white paper and summarised by SSRS.
The 83 percent worried about data collection is the number we take most seriously, and it is the reason what happens to a child's voice data deserves a plain-language answer from every company selling one of these, ours included.
What does none of this research tell us yet?
It does not tell us what happens over months. Every study described above is short. A lab session lasts an hour. Even the eight-day home trials that exist in the literature are a blink compared with how long a child keeps a favourite toy.
It also does not separate the toy from the household. A child who talks to a plush bear for twenty minutes while a parent cooks dinner is in a different situation from a child parked with one for two hours. Research that treats "AI toy" as a single exposure will keep producing muddy results, which is one of the points raised in a recent Journal of Medical Internet Research overview of where policy and evidence currently stand.
And it does not yet compare designs. Grouping a screen-free plush with no camera together with a screened robot that stores biometric data is like reviewing bicycles and motorbikes as one category.
What can a parent do while the studies run?
Do not wait for a verdict that is years away. Use the questions the researchers are using.
Five checks before any talking toy enters the house
- Is there a camera? If yes, ask why a listening toy needs eyes.
- Where do the recordings go, how long are they kept, and can you delete them in one tap?
- Can an adult set the topics, the language and the tone, or does the toy decide?
- Is there a subscription? A toy that stops working when a payment lapses is a service, not a toy.
- Does it hand play back to your child, or does it try to hold their attention?
Then watch the play itself for a week. If your child is narrating, arguing, correcting and inventing, the toy is doing its job. If your child has gone quiet and is only listening, something has gone wrong, whatever the box promised.
Where does Ted sit in all this?
Ted is a plush companion that talks with children from age 3 and up, in more than 60 languages, with no camera anywhere on it and no subscription. Parents set the language, the topics and the limits from the Ted&Co app, which is also where you decide whether Ted listens automatically or only when a child presses to talk.
We built it that way because the criticisms above were predictable. A toy without a camera cannot leak footage of a bedroom. A toy without a subscription has no commercial reason to keep a child engaged for longer. Our certifications and our data commitments, including GDPR and COPPA, are laid out on the security page, and you can see the product itself on the Ted smart teddy page. None of that makes Ted exempt from the Harvard and Cambridge findings. It means we would rather be measured against them than avoid them, and we will report on the results when they land.
Frequently asked questions
Is the Harvard AI toy something I can buy?
No. The Harvard Child-Centered AI Lab built it as a research instrument, not a product. It exists so that researchers can control every variable in the experiment, and there is no indication it will ever be sold.
Has any study proven that AI toys harm young children?
No study has shown long-term harm, and no study has shown long-term benefit either. The Cambridge report documented specific failures during play sessions, such as misreading emotions and blocking pretend play, which is a different and narrower claim.
Are talking toys safe for a 3-year-old?
It depends entirely on the design. The features that matter are whether the toy has a camera, what it does with voice recordings, whether an adult controls the topics, and whether it encourages a child to keep playing or to keep listening.
Why do so many parents worry about data collection?
Because 83 percent of surveyed parents said so, and because past testing of toys on sale found privacy policies allowing third-party sharing and long retention of biometric data. Reading the policy before buying is still the single most useful thing a parent can do.
Should I just avoid the category until the research is finished?
That is a reasonable choice, and some guidance for under-fives points that way. If you do buy one, treat it as a shared activity rather than a substitute for adult conversation, since back-and-forth exchanges with a person remain the best-evidenced driver of language growth.