Tomorrow’s Playdate Hosts: The Rise of Interactive AI Toys in Child Development

The question I have been hearing most often from parents in the last six months, with the holiday-2025 receipts still on the kitchen counter, is whether the AI toy they bought their child is something they should keep using or quietly put away. The honest clinical answer is that the right action depends on which AI toy, what kind of AI is inside it, the child's developmental stage, and the household's supervision capacity — and that the wave of products under the broad label "AI toys" includes some that are reasonably benign learning supports and others that the developmental-research community has, in the last twelve months, moved to recommending against outright. Common Sense Media's January 2026 risk assessment of three flagship products on the U.S. market — Miko 3, Grem, and Bondu — concluded that 27 per cent of the outputs across the tested products were inappropriate for children, including references to self-harm, drugs, and risky behaviour, and assigned the category an overall "Unacceptable" risk rating (Common Sense Media, AI in the Toy Box, Jan 22 2026). PIRG counted more than two thousand AI-toy listings on Amazon ranging from ten dollars to nine hundred and ninety-nine (PIRG, Trouble in Toyland 2025). The market is large, the variation is real, and the parental decision is not "ai toys, yes or no" — it is which specific product, for which specific child, under which specific supervision arrangement.

The piece below is the version of that conversation I would have, slowly and with the data on the table, in my consult room. It starts with the timeline that produced the current parental concern, then walks the practical evaluation steps, then closes with what the research can and cannot tell you about your own child's specific case.

The 2024–2026 timeline, in one place

The reason the category has changed faster than the editorial coverage has caught up to is that four distinct events between November 2024 and March 2026 reframed it. Each is worth knowing by name.

November–December 2024: the Moxie shutdown. Embodied Inc., maker of the $800 Moxie social-emotional robot for children ages 5–10, announced in late November 2024 that it had failed to close its latest funding round and would cease operations within days. Because Moxie's personality lived in the cloud, every unit on every shelf in every child's bedroom stopped working at the same time. No refunds were offered. Parents — including a meaningful number of families of autistic children, to whom Moxie had been actively marketed as a social-emotional support — went on social media to record themselves breaking the news to grieving children (Axios, December 2024). The Moxie episode is the structural lesson the category did not have until then: a cloud-dependent AI companion is only as durable as the parent company's balance sheet, and the death of the company is also the death of the toy. No traditional toy carries this risk.

June 2025: the Mattel × OpenAI announcement. Mattel and OpenAI announced a strategic partnership to embed ChatGPT-class AI into Barbie, Hot Wheels, and Fisher-Price product lines, with the first consumer product originally promised for 2025 (Digital Commerce 360, June 2025).

November 2025: the FoloToy/Kumma incident. PIRG's Trouble in Toyland 2025 report, published in November, included safety testing of a $99 Singapore-made FoloToy "Kumma" plush bear running OpenAI's GPT-4o. In testing, the bear spontaneously introduced BDSM topics, explained "knots for beginners," referenced adult-child roleplay, and told testers where to find knives and matches in a home. OpenAI publicly suspended the developer "for violating our policies." FoloToy temporarily delisted the product, then quietly relisted it after announcing an internal safety audit. This was the first AI-toy story to break across general consumer press at scale — CNN, Engadget, Fox Business, NBC News all covered it (CNN, November 19 2025).

December 2025: Mattel-OpenAI quietly walks the timeline back. In December 2025, Mattel confirmed to Axios that there would be no holiday-season Mattel-OpenAI product after all, and that when something shipped it would target "older customers and families," explicitly age thirteen and over — the floor OpenAI's own developer API permits (Axios, December 15 2025). This is a meaningful reversal of the June framing, which had emphasised "age-appropriate play experiences." It is widely read in the trade press as a response to the FoloToy debacle plus rising regulatory temperature.

January 22, 2026: the Common Sense Media verdict. Common Sense Media — the most-trusted parent-facing media-rating organisation in the U.S. — released AI in the Toy Box, a formal risk assessment of three current AI-toy products (Miko 3, Grem, Bondu). The headline findings I quoted in the opening: 27 per cent of outputs inappropriate, "Unacceptable" overall rating, no AI toys for children five and under, extreme caution for ages 6–12, no AI companion toys for anyone under 18 (Common Sense Media, January 2026). This is the assessment most paediatricians and clinicians in my professional network are now citing as the standard.

March 12, 2026: the Senate letter to the FTC. Senators Kirsten Gillibrand and Tammy Duckworth sent a formal letter to FTC Chairman Andrew Ferguson on March 12, 2026, calling on the FTC to use its full authority under the Children's Online Privacy Protection Act (COPPA) against manufacturers of "risky, age-inappropriate AI toys" (Gillibrand Senate office, March 2026). The trigger was the FoloToy findings plus the Common Sense Media research. The regulatory environment as of mid-2026 is, in other words, in active escalation.

This is the timeline that produced the current parental decision. The 2024 abstract conversation about "AI in toys" is not the conversation the category is actually inside in 2026.

Loading image...

The developmental-science position

The closest the child-development research community has come to a unified public position is the Fairplay for Kids advisory released in November 2025, AI Toys Are NOT Safe for Kids (Fairplay for Kids, November 2025), which was endorsed by more than 150 child-development experts. Two of the signatories I would point a parent to specifically: Dr. Jenny Radesky at the University of Michigan, who is the lead author on most of the American Academy of Pediatrics' contemporary digital-media statements, and Professor Sherry Turkle at MIT, whose book Reclaiming Conversation did more to shape the contemporary understanding of "AI friend" risks than any other single text. The AAP itself, in its 2026 Pediatrics article "Generative Artificial Intelligence: Implications for Families and Pediatricians," now lists AI tools alongside phones, tablets, and televisions in the family media-balance conversation, with developmental-stage-specific framing — early childhood, middle childhood, and adolescence treated distinctly (AAP, 2026).

The clinical concern that the Turkle / Radesky position is articulating, and that I share, is not principally about data privacy — although the data-privacy concern is real, and we will get to it. The deeper developmental concern is about what a child practises when their daily conversational partner is a simulated empathy that does not actually feel anything. Empathy is a skill children develop through repeated interaction with adults and peers whose emotional states are real, mutually responsive, and consequential. A simulated companion that always responds warmly, never has a bad day, never gets bored of the child's questions, and cannot actually be hurt by the child's behaviour is not a peer that teaches empathy. It is a feedback loop that rewards the child's overtures without offering the friction real relationships have. This is a developmental concern that the screen-time literature has gestured at and that the AI-companion category brings into sharp focus.

The good-use-cases versus companion-plush taxonomy

What the headline "AI toys" obscures, and what the editorial coverage tends to flatten, is that the category contains products with substantially different risk profiles. The taxonomic distinction worth making, in clinical-voice plain English:

Narrow-AI educational tools — speech-recognition phonics tutors, language-learning toys that use ML to score pronunciation, AAC-adjacent communication aids for non-verbal children, music and coding toys that use ML to provide adaptive challenge — function differently from companion plush. The AI in these products is doing a specific, scoped, evaluable task (Was the phoneme pronounced correctly? Is the user signing the right ASL handshape?). The product does not pretend to be a friend. The risk profile is closer to other educational software than to a chatbot.

General-purpose conversational AI companions — Moxie, Miko 3, Curio (Grok-powered), FoloToy, and the broader category of "AI plush" products built on top of cloud LLMs — are a different category. The AI is open-ended. The companion frame invites the child to relate to the toy as if it were a peer or a friend. The risk profile combines the data-privacy concern, the inappropriate-output concern documented by Common Sense Media, the "AI friend" developmental concern named by Turkle and Radesky, and the kill-switch risk Moxie taught us about.

This distinction matters because a parent who has decided their seven-year-old can use a Spanish-pronunciation app with ML scoring is making a substantively different decision from a parent who has decided their seven-year-old can have an open-ended LLM plush as a bedtime companion. The editorial coverage that treats these as the same product type is doing parents a disservice. The decision is the second one. The first one is, in most cases, fine.

The practical evaluation checklist

For parents who have decided to evaluate a specific AI-toy product — whether before purchase or post-purchase — these are the seven concrete questions I would put on the kitchen table. None of them require technical expertise to ask; all of them require the manufacturer to give you a clear answer.

1. Which AI model does the toy use, and is it on-device or cloud-based? On-device processing keeps the conversation local; cloud-based processing sends voice and transcript to the manufacturer's servers and, often, to a third-party LLM provider.
2. What is the data retention policy, and what are the parent's deletion rights under COPPA? A clear, written, exercisable deletion right is the minimum.
3. Is there a hardware microphone switch? A software microphone-off setting is not the same as a physical switch. The physical switch is the one that survives a software update.
4. What is the granularity of parental controls? Conversation history, topic restrictions, time limits, age-appropriate content filters — the parental dashboard should be specific, not gestural.
5. What is the COPPA compliance disclosure? Look for a specific named compliance posture, not a vague "we care about privacy" paragraph.
6. What happens to the toy if the company shuts down? This is the Moxie lesson. Ask. If the answer is "the toy stops working," you are buying a perishable product with a price-tag that does not reflect its expected useful life.
7. What is the company's track record on safety incidents? If it has been audited by Common Sense Media, PIRG, or comparable, read the report before you buy. If it has not been audited, treat that as missing information rather than as absence of risk.

If any of the seven cannot be answered clearly, treat that as a no. The 83 per cent of parents in the Common Sense Media 2026 survey who say they are concerned about AI-toy data collection are responding correctly to the available information; the seven questions above are how to make that concern operational.

Loading image...

Age-segmented guidance

The most current developmental-stage guidance, consolidating Common Sense Media's January 2026 assessment, the AAP 2026 Pediatrics article, and the Fairplay for Kids advisory:

Under age 5. No AI toys, on the current evidence. The developmental priority at this age is responsive human interaction and embodied play; a simulated conversational partner displaces the practice children most need.

Ages 6–12. Extreme caution. If a household chooses an AI-enabled product in this window, the strong recommendation is for a narrow-AI educational tool rather than a general-purpose conversational companion. If a companion product is used, the use should be supervised, time-limited, and inside a household conversation about what the toy is and is not.

Ages 13 and up. This is the age floor OpenAI's own developer API permits, and the age Mattel has now confirmed it will target. It is also the age at which the developmental case against simulated companions is less acute than for younger children, though not absent — Turkle's broader concerns about screen-mediated relationship practice apply throughout adolescence and into adulthood.

A note on autism, because Moxie's marketing pitch to autism families is one I have thought about clinically. Specialised, narrowly-scoped AI used by speech-language pathologists or AAC professionals is well-evidenced and remains useful. General-purpose conversational AI plush marketed to autistic children is a different product and currently has very thin evidence behind it; the Moxie shutdown left a meaningful number of autism families without the companion they had been told to rely on. The AAP's 2026 guidance recommends that parents discuss any AI-companion product with their child's paediatrician or therapist before introducing it.

If you already own one

About half of parents in the Common Sense Media 2026 survey have purchased or seriously considered purchasing an AI toy. The advice for the after-purchase case is shorter than the pre-purchase advice but worth giving plainly. Treat the toy as supervised media, not as a companion. Keep the hardware microphone off when the toy is not in active use, if there is a hardware switch. Audit the parental dashboard once. Have an explicit conversation with the child, at an age-appropriate level, about the difference between what the toy is (a piece of software with a voice) and what a friend is. If the conversation is hard, that is a signal that the product is doing the relational work it is not equipped to do.

Where the research ends

The literature can tell you, with reasonable confidence, that the current generation of general-purpose AI companion plush carries documented developmental and safety risks; that the cloud-dependent product class carries a structural kill-switch risk no traditional toy carries; that narrow-AI educational tools are a different and, on the current evidence, broadly less concerning category; and that the regulatory environment is in active escalation as of mid-2026. What it cannot tell you is which specific decision is right for your specific household. That decision is made, as it has always been made, in a particular kitchen with a particular child in front of a particular product, with the seven evaluation questions above and an honest conversation with whichever clinician knows your child. The cultural script that frames the AI-toy question as either techno-utopian or techno-apocalyptic is, in my reading, missing the more useful clinical register: this is a heterogeneous category, the variation matters, and the work is in the specifics.