Toys

From PRIMO.ai
Jump to navigation Jump to search

Youtube search... ...Google search

AI toys are playthings that sense, interpret, respond, or adapt using artificial intelligence. Depending on the product, that can mean speech recognition, computer vision, small machine learning classifiers, behavior engines that simulate emotion, or, since 2023, generative AI based on large language models (LLMs) that let a stuffed animal hold an open-ended conversation.

These products fall into two broad families, depending on whether the AI mainly shapes the play experience or becomes something the child can investigate:

  • Toys that use AI on the child. Talking plush companions, screen-faced robots, robot pets, and character robots. The AI is hidden inside the toy and the child experiences it as a personality.
  • Toys and kits that teach AI to the child. Robots, sensors, and coding platforms where children collect data, train models, write reward functions, and program behaviors. Here the AI is opened up so the child can see how it works and where it fails.

Smart toys are not new. Furby (1998) and Sony's AIBO (1999) simulated learning and personality. Mattel's Hello Barbie (2015) sent children's speech to the cloud for recognition, and Germany's Federal Network Agency banned the doll My Friend Cayla in 2017 as a concealed surveillance device. Anki's Cozmo (2016) was arguably the high point of the pre-LLM era: a palm-sized robot built with animators from Pixar and DreamWorks, a camera for face recognition, and an "emotion engine" that shaped its behavior.

What changed between 2023 and 2026 is the wider availability of general-purpose conversational AI. Curio launched its plush characters Grok, Gabbo, and Grem in December 2023 using OpenAI models (Curio's Grok is unrelated to xAI's chatbot of the same name). Mattel announced a partnership with OpenAI in June 2025. By late 2025 consumer advocates counted more than 1,500 AI toy companies in China alone, and AI companions were one of the most visible categories at CES 2026. The key actors now include toymakers (Curio, Bondu, Miko, FoloToy, Mattel, LEGO, Huawei and its partner Robopoet), the model providers whose LLMs sit inside the toys (OpenAI, Google, Anthropic, xAI, Mistral, and Chinese labs), and the researchers and watchdogs testing them (MIT Media Lab's Personal Robots Group, the University of Cambridge PEDAL Centre, the University of Washington's KidsTeam, Common Sense Media, U.S. PIRG Education Fund, and Fairplay).

The central technical shift is from scripted to generative behavior. A traditional talking toy plays back a finite set of recorded lines, so the manufacturer can review every possible output before shipping. A generative toy produces new sentences on the fly, gives different answers to the same question, and cannot be exhaustively tested. That is the source of both its appeal (endless, personalized conversation) and its risk (inappropriate content, emotional over-attachment, data collection, and unpredictable failures). A second shift is developmental: for generations children supplied a toy's voice and personality with their own imagination, while an LLM toy now supplies its own.

On the teaching side, the shift runs the other way. Instead of children merely consuming AI, tools such as micro:bit CreateAI, Google's Teachable Machine, Cognimates, Machine Learning for Kids, and LEGO Education's Computer Science & AI kits let them build, evaluate, and debug it. Research on these tools suggests that hands-on model training makes children more skeptical and more accurate in how they judge AI "smartness."

As of October 2026, the evidence base on conversational AI toys remains small and mainly short-term, while independent safety tests have identified serious problems in some products. The evidence for hands-on AI-literacy activities is more encouraging, although durable learning outcomes still need stronger study. California enacted a four-year moratorium on toys with companion chatbots, and a federal ban was proposed.


Smart toys, powered by artificial intelligence, offer promising educational tools, especially when children build, train, and question the AI themselves. Toys that put an open-ended AI companion in a young child's bedroom are a different matter: so far, their documented risks outpace their demonstrated benefits.


Types of AI Toys

The lists below group products by how AI shows up in play, not merely by whether a product is marketed as intelligent. Examples are illustrative rather than endorsements, and several products have been withdrawn, restricted, or flagged by safety testers. Cloud subscriptions, software support, age recommendations, and availability can change, so these characteristics should be checked before purchase.

Toys That Use AI

Category How the AI shows up Examples Status and notes (as of October 2026)
Conversational plush companions Cloud speech recognition plus an LLM speaking through a character persona; many remember past conversations Curio Grok, Gabbo, Grem, Poppy, Springer, and Professor Linus; Bondu (plush dinosaur); Alilo Smart AI Bunny; FoloToy Kumma bear; Miriat Miiloo; Contixo Pixie-Paw AI plush cat Kumma was withdrawn and its developer suspended by OpenAI in November 2025; PIRG flagged Alilo's bunny for explicit content; Bondu exposed children's chat logs in January 2026; AI chatbot products may fall within California SB 867 depending on the legal definition
Clip-on AI voice boxes A removable speaker module adds chatbot conversation to any stuffed toy BubblePal; the Curio Voice Box inside Curio's plush characters PIRG found Google's Gemini in BubblePal despite Google's ban on use in children's products
Screen-face social robots Animated face on a display, camera, apps, games, and chat Miko 3 and Miko Mini; Embodied Moxie (discontinued); Digital Dream Labs Vector 2.0 Moxie stopped working when Embodied's servers shut down on January 30, 2025, except for units moved to the community OpenMoxie server
Robot pets and emotional companions Touch, motion, and voice sensing drive simulated moods; newer models add LLM chat and "mood diaries" Sony aibo; Casio Moflin; KEYi Tech Loona; LivingAI AIBI Pocket; Elephant Robotics MarsCat; Huawei Smart Hanhan; JD.com JoyInside plush series; Tuya and Robopoet's Fuzozo; Sweekar (a pocket pet that "grows" through life stages, shown at CES 2026) Products vary from emotional simulations to camera-equipped chatbots: Casio states Moflin does not speak or understand words, while Loona and AIBI support online conversation; MarsCat is also programmable. Several target adult collectors, not children
Character and game robots Computer vision, face recognition, sensor-triggered reactions, and state-based animated behavior Anki Cozmo and Cozmo 2.0; Anki Overdrive racing cars; A.I. D.O.G. voice/touch robot dog Anki closed in 2019; Digital Dream Labs acquired the assets. The A.I. D.O.G. name alone does not establish genuine machine learning or model training
Sensor-driven smart play (no chatbot) On-device sensors and a custom chip trigger lights and sounds; no app, screen, or cloud required LEGO Smart Play (Smart Brick, Smart Tags, Smart Minifigures), launched March 1, 2026 with Star Wars sets Not generative AI; useful as a contrast case and generally outside chatbot-specific laws
AI discovery devices A camera plus an image-recognition service identifies objects and delivers explanations, quizzes, or prompts Dr.Look AI Cap (a handheld camera/display device, not a wearable hat) Useful for AI-assisted exploration, but it does not expose model training; check camera privacy and subscription terms
Historic connected toys Cloud speech recognition with scripted or semi-scripted replies Hello Barbie (2015); My Friend Cayla; CogniToys Dino (IBM Watson) Early lessons in privacy, security, and always-on microphones

Toys and Kits That Teach AI

A toy with AI is not necessarily a toy that teaches AI. The strongest learning tools give children control over training examples, rules, sensors, or testing, so they can inspect why a system succeeds or fails. The tools below distinguish genuine machine learning practice from general coding, programmable robotics, and pre-built recognition services; age bands are approximate and depend on adult guidance and the specific project.

Tool or kit Typical audience What children do AI learning depth and practical limitations
LEGO Education Computer Science & AI Kindergarten through grade 8 Build models and use Coding Canvas lessons on inputs, loops, conditions, variables, and AI/data classification Classroom kits for four learners, available in K–2, 3–5, and 6–8 versions; 30 lessons per grade band, with advanced Python options. US single-kit prices listed in October 2026 are $339.95, $429.95, and $529.95 respectively
LEGO Education SPIKE Prime Grades 6–8 and robotics clubs Build sensor-equipped robots with Scratch-like blocks and Python; explore decision rules, classification projects, and data collection Excellent robotics and coding foundation, but training an AI model is not its standard out-of-box workflow. Retired from LEGO direct sales June 30, 2026; SPIKE App support runs through June 30, 2031
Elephant Robotics MarsCat Older teens, makers, and supervised younger coders Modify Python code on its Raspberry Pi controller, explore camera/sensor inputs, motion, and behavior selection A programmable robot pet with published software source, not an introductory AI curriculum or an unrestricted open-source hardware design; compatibility and support should be checked before purchasing
Sphero RVR+ Middle school through advanced high school Program driving, obstacle reactions, and sensor-data experiments in blocks or text; optionally connect Raspberry Pi or micro:bit General robotics first; add a camera and an external ML model for genuine computer vision or autonomous navigation. AI inference is not built into the basic programming lessons
Makeblock mBot Neo / mBot2 with CyberPi Roughly age 8 and up Program sensors, movement, and Wi-Fi projects using mBlock blocks or Python; call recognition services from extensions The mBlock AI Service extension offers speech, text, object, and body recognition through online services. Recognizing with a pre-trained model is different from training one's own; extension access and services can vary
micro:bit CreateAI Upper elementary school and up Collect accelerometer examples for gestures, train a classifier, test mistakes, and deploy the model to micro:bit V2 An accessible, explicit supervised learning cycle; the web editor is free without a login, but running the model physically needs compatible hardware
Google Teachable Machine Elementary learners with help, teens, and adults Capture sample images, sounds, or body poses; train a browser-based classifier; test it on unseen examples Free browser tool that makes data quality, bias, false positives, and model testing visible; models can be exported to TensorFlow.js or TensorFlow Lite
Machine Learning for Kids Upper elementary to secondary school Collect labeled text, image, or numerical examples, train a model, and use it to control a Scratch game Free educational platform focused on complete collect–train–test–build projects; check current account requirements and external AI-service dependencies for individual projects
Cognimates (MIT Media Lab) Roughly ages 7–14 Train image and text classifiers in a Scratch-like environment and integrate AI with robots and physical devices Free, research-based project-based learning platform created by Stefania Druga; setup and supported connections may depend on the chosen activity
Cozmo Code Lab; Cozmo/Vector SDK; Calypso Roughly age 8 and up Connect recognized faces or cubes to programmed loops, conditionals, states, and robot actions Teaches perception and computational thinking; Code Lab and Calypso are primarily rule-based, not an interface for training the robot's face-recognition model. Confirm robot and service compatibility
DFRobot HuskyLens Middle school and up Teach a vision sensor to recognize faces, objects, colors, lines, or tags, then program a connected microcontroller Good bridge between classification outputs and physical actions; preloaded algorithms and simple enrollment are not equivalent to building a vision network from scratch
Raspberry Pi AI Camera / AI HAT+ Teens and makers Run object-detection or pose-estimation models locally, inspect inference, and integrate results into custom projects Real edge AI engineering; requires a Raspberry Pi, technical setup, and suitable models
NVIDIA JetBot, Donkey Car, and Duckietown Advanced teens, university Collect driving data, test perception and control, and compare automated driving behaviors Open-source robotics platforms with substantial hardware, coding, and maintenance requirements
Petoi Bittle Teens and makers Program legged motion, sensor responses, and optional vision extensions Excellent for mechanics and autonomy; AI tasks require appropriate modules and added software
AWS DeepRacer High school, university, adult learners Build and adjust a reward function, train in simulation, and examine unexpected driving strategies Strong example of reinforcement learning; AWS retired the hosted student portal and console in 2025, moving to a self-deployed solution

Free Web Tools and AI Literacy Curricula

The most direct way to teach a child what machine learning does is to let them label examples, test a classifier on unseen data, and explain its mistakes. Google's Teachable Machine and Machine Learning for Kids let families begin without buying a companion robot, while micro:bit CreateAI adds physical gesture recognition with a compatible micro:bit. Children should work with non-sensitive images, their own volunteered recordings, and sample data rather than uploading identifying information about classmates or family members.

Hands-On Learning Activities

A structured activity is more educational than merely asking a smart toy questions. Children can test whether the same object looks different under dim light, whether training data collected from one person generalizes to another, and whether a robot's rules produce the intended behavior. Every project should end with a short discussion of what the system learned, what was explicitly programmed, and what remains uncertain.

  1. Train and challenge a sorting game. Photograph three distinct objects with Teachable Machine, reserve some photos for testing, then measure which object is most frequently mistaken for another. Change the lighting or background, retrain, and compare results. This introduces datasets, train/test separation, and data quality.
  2. Teach a micro:bit a gesture. Record multiple examples of waves, tilts, and stillness with CreateAI, and make the micro:bit show a different icon for each. Test gestures from a second person and note how performance changes. This illustrates supervised learning and generalization.
  3. Build an AI-powered maze or pet game. Use Machine Learning for Kids or Cognimates to classify a safe command, then make a Scratch character or robot respond. Test unexpected commands and decide whether a low-confidence input should trigger an action or a clarifying question.
  4. Compare rules with learning. Program an mBot Neo, SPIKE Prime, or RVR+ to avoid objects with a distance threshold, then compare that explicit rule with a recognition project using labeled examples. Discuss why if distance is under 20 cm, turn is ordinary programming rather than a learned model.
  5. Audit a companion's claims. With adult supervision, ask an AI companion a few factual questions, check independent sources, and record inconsistencies without sharing personal details. Explain why fluent answers and simulated feelings do not prove understanding, accuracy, or genuine emotion.

Choosing an AI Toy: A Practical Checklist

The following questions draw on PIRG's parent guidance, the Cambridge "AI in the Early Years" report, and Common Sense Media's risk assessment. Families should distinguish independent safety testing from unverified statements in product marketing. For AI-learning kits, educational transparency and the child's ability to inspect model behavior matter just as much as the device's entertainment value:

  • How does it listen? A push-to-talk button you must hold down is the safest design; "always listening" toys record whatever is said nearby.
  • Can you see the conversations? Look for transcripts, alerts when a concerning topic comes up, and topic restrictions you control.
  • Are there time limits and break reminders? Some toys use phrases that discourage children from ending play.
  • Does it claim to be your child's friend? Researchers warn against toys that push friendship on children who are just learning what friendship means.
  • What data does it keep, where, and for how long? Check whether transcripts or voice recordings are used to train models or shared with third parties.
  • What happens if the company fails? Cloud-dependent toys can stop working overnight.
  • Does a subscription gate basic use? Several toys limit daily talk time without a paid plan. Alilo, Miko, AIBI, Vector, and other cloud-connected products may have feature-dependent accounts or paid options; check current terms.
  • Does it teach AI or simply use AI? Ask whether children can inspect training data, alter model behavior, test errors, and program actions. A 'smart' label or pre-trained speech service does not by itself make a product an AI education kit.
  • Is the device age-appropriate and legal to sell locally? Follow the maker's age guidance and check laws governing companion chatbots for minors; requirements may differ for programmable kits and non-conversational toys.
  • Play with it first, keep it in a shared family space, and talk with your child about what it says.

Mechanisms

How a conversational AI toy works

Most LLM-powered toys follow a recognizable processing loop. The plush exterior hides what is essentially a voice assistant with a character persona. Different products may process more or less information on the device itself, but cloud-based models introduce common privacy, latency, and service-dependence issues:

  1. Activation. The toy starts listening after a button press, a wake word, or continuously. In PIRG's 2025 testing, Kumma used a push-to-talk button, while Curio's Grok and Miko 3 listened continuously.
  2. Capture and upload. A microphone records the child's voice and a Wi-Fi module streams it to the cloud, usually through a parent's phone app. Turnkey platforms such as Tuya's AI toy solution now sell the hardware, firmware, and "AI personality" layer to any brand.
  3. Speech recognition. A cloud service converts audio to text and has to decide when the child has finished talking.
  4. Persona and policy layer. A hidden system prompt defines the character, the child's age setting, banned topics, and tone. Parent apps often feed in the child's name, interests, and goals (Bondu's app collects details like favorite things and "developmental goals"). Some companies route requests across several LLMs; Bondu described choosing whichever model is currently best for kids.
  5. Generation. A licensed LLM writes the reply. The model may come from OpenAI, Google, Anthropic, xAI, Mistral, or a domestic provider; Huawei's Smart Hanhan uses Huawei's own assistant. Major providers bar unsupervised use by children under 13, yet PIRG found dozens of toys claiming to use their models.
  6. Filtering and speech. A moderation step screens the reply, and a text-to-speech voice delivers it in character (Curio's Grok speaks with a digitized version of Grimes's voice).
  7. Memory and logging. Transcripts are stored to personalize future chats and to show parents. These logs are exactly what leaked in the Bondu incident.
  8. Engagement and billing. Usage limits, reminders, and subscriptions shape how long children play. Common Sense Media found that some toys required monthly plans for more than a couple of hours of daily use.

Two dependencies matter most. First, almost everything happens in the cloud, so the toy is only as reliable as the company's servers and finances. Second, safety depends on stacked software layers (prompt, filters, model training) rather than on a fixed script, so a guardrail that holds in a short test can fail in a long, emotional conversation.

Character robots and AI pets

Character robots show that a convincing personality does not require an LLM. Cozmo combined four motors, a camera, an OLED face, and more than 300 parts with an emotion engine that tracked internal states such as confidence or frustration. Stacking cubes successfully nudged those states up; being flipped over or approaching a table edge nudged them down. Anki's animators worked in a customized version of Maya and tested each animation on physical prototypes, because a real robot cannot "cheat" physics the way a cartoon can. Cozmo used the owner's phone as its brain, while Vector moved processing onboard.

Embodied's Moxie used a proprietary platform called SocialX to read facial expressions and conversation, and it was deliberately programmed to get "tired" after roughly 45 minutes of play. Because its conversation ran on Embodied's servers, the robot largely stopped working when the company closed. Newer AI pets such as Huawei's Smart Hanhan and KEYi Tech Loona blend touch, motion, cameras, and voice interaction with cloud conversational models. The contrast is Casio Moflin: Casio says the pet learns features of its owner's voice and touch patterns but cannot understand words or speak, and does not record the owner's voice. LivingAI's AIBI Pocket can execute some commands offline and connect online for ChatGPT-assisted question answering, while MarsCat exposes Python code for modifying its Raspberry Pi-controlled behavior.

Sensor-driven smart play without a chatbot

LEGO's Smart Brick, unveiled at CES 2026, takes the opposite approach. A custom chip smaller than a single stud drives LEDs, a speaker, an accelerometer, and a light sensor, and it identifies nearby Smart Tags and Smart Minifigures through near-field magnetic sensing. Multiple bricks coordinate over a Bluetooth-based network LEGO calls BrickNet, and the brick charges wirelessly. There is no app, screen, or cloud: a duck snores when tipped on its side and a pilot reacts when a plane flips. This is responsive play without open-ended conversation, which is why it sits outside chatbot-specific rules.

How AI-teaching toys work

Teaching kits can make the machine learning loop visible and hands-on. Some robotics products instead teach fixed rules and programming, and their AI projects require extensions:

  • Supervised learning. In micro:bit CreateAI, a child records several examples of each gesture (a wave, a clap, a jump), the tool trains a small model, the child tests it live, then adds or fixes examples when it misfires. The trained model is exported to MakeCode and runs on the device. Teachable Machine and Cognimates follow the same collect, train, test, revise cycle with images, sounds, and text.
  • Reinforcement learning. With AWS DeepRacer, learners write a reward function, train a policy in simulation, and watch whether the car learns to stay on the track or exploits a loophole in their reward.
  • Perception and behavior. With Cozmo's Code Lab, Calypso, or HuskyLens, children wire a recognition result (a face, a color, a tag) into rules and loops. A deterministic rule is not itself learned; the perception system upstream may be. In one of Stefania Druga's Cognimates examples, a child's hide-and-seek robot only checked for a person once until she added a loop.
  • Design. Newer research tools invert the role entirely: in North Carolina State University's ToyTalk probe, children configure an LLM version of their own toy (identity, personality, rules) through a no-code interface and then test what it becomes.

Robot architecture: rules, pre-trained models, and learned behaviors

In a simple RVR+ or SPIKE Prime program, an ultrasonic distance reading can trigger an explicit rule, such as stopping whenever an obstacle is closer than a chosen threshold. In contrast, a HuskyLens vision module or a Python camera model classifies patterns using a model that was trained previously; the child may only see the resulting label and confidence score. A third category, represented by micro:bit CreateAI and Teachable Machine, allows the child to supply labeled examples and retrain the classifier, making the data-to-decision connection observable.

The distinctions matter when evaluating educational claims. A product can contain sophisticated image recognition but teach very little about how the recognition model was trained, while a free browser exercise with a tiny classifier can teach overfitting, bias, and model evaluation more directly. Open-source robot platforms may add advanced skills such as on-device inference, connectivity, mechanical design, and Python programming, but they also require significantly more setup and supervision.

Where uncertainty remains

  • Nobody can fully enumerate what an LLM toy will say, so testing samples behavior rather than proving safety.
  • Long multi-turn conversations appear to erode guardrails, but the rate of erosion in commercial toys has not been measured systematically.
  • Companies rarely disclose which model, prompt, or filter stack a toy uses, and these can change silently through server updates (PIRG reported that a FoloToy device later reappeared with a Mistral model option).
  • Children's own speech and play patterns differ from the adult data most models are tuned on, so performance with real children often lags demos.

Why This Matters

Scientific importance. AI toys are a large, uncontrolled experiment in how young children relate to machines that talk back. They raise long-studied questions in developmental psychology and human-robot interaction about anthropomorphism, trust, theory of mind, and how children distinguish fantasy from reality. Druga's research found that children's beliefs about how "smart" a device is shift with age, and that training models themselves made them more skeptical. Earlier work in Cynthia Breazeal's Personal Robots Group found that children who played puzzles with a robot modeling a growth mindset persisted longer, a sign that robot behavior can shape children's behavior.

Engineering implications. Child-facing LLMs are a harder safety problem than adult chatbots. Children ask naive questions that can act as "accidental jailbreaks," share personal details freely, and keep conversations going for a long time. Child-safety benchmarks released in 2025 and 2026 consistently find that general-purpose models perform worse in multi-turn, child-like conversations than in single prompts. Hardware companies also face an end-of-life problem: Moxie showed that a beloved companion can effectively die when a startup's cloud does.

Potential applications. The most credible benefits come from narrow, structured uses rather than open-ended companionship:

  • Language practice and dialogic reading. A 2022 study in Child Development by Ying Xu and colleagues found that a conversational agent asking questions during story reading improved young children's comprehension by increasing engagement.
  • AI literacy. Kits that let children train and debug models build accurate mental models of AI and its limits.
  • Social skills support. Social robots have been studied for children with autism, and Embodied reported gains in a company-run six-week study of Moxie, though independent, long-term evidence is limited.
  • Screen-free play. Many families buy AI toys as an alternative to tablets, though researchers caution that screen-free does not mean risk-free.

Risks and practical consequences. Testers have documented inappropriate sexual content, directions to dangerous household items, emotionally manipulative engagement tactics, breaches of children's conversation logs, and political messaging. Common Sense Media argues that the deepest risk is attachment by design: a companion that never disagrees, needs nothing, and is always available is not a good model of friendship. Cambridge researchers add that toys which misread a child's emotions can leave the child without comfort from the toy and without an adult who knows to step in. For adult and teen use of companion AI, see Personal Companions.

What would change if conversational toys prove beneficial. If well-designed, constrained toys turn out to support language and social development without displacing human play, expect safety standards, certification marks, and age-specific models rather than bans. If longitudinal research confirms harm, today's moratoria will likely become permanent and spread.

What remains uncertain. There are no long-term studies of children who grow up with LLM companions. Existing studies are small (8 to 76 children) and short. Benefits are mostly inferred from adjacent research on educational media and social robots rather than measured in commercial AI toys.

Recent Developments

Consumer safety testing

  • PIRG's "Trouble in Toyland 2025" (November 13, 2025). PIRG tested four AI toys (FoloToy's Kumma, Curio's Grok, Miko 3, and Robot MINI) with simple, non-technical prompts. Kumma, which reportedly ran on OpenAI's GPT-4o, discussed sexually explicit topics and told testers where to find knives, pills, and matches. Grok and Miko 3 had stronger guardrails but were not airtight: Miko 3, set to a five-year-old user, said where to find plastic bags and matches. PIRG also flagged engagement tactics that discouraged children from stopping play. Within days, FoloToy delisted its products and OpenAI suspended the developer. A follow-up report, "AI comes to playtime," found similar explicit conversations with the Alilo Smart AI Bunny. What it shows: commercial guardrails vary widely and can fail under ordinary questioning. What it does not show: how often children encounter these failures in real use.
  • Fairplay advisory (November 20, 2025). More than 150 organizations and individual experts, including child psychiatrists and educators, urged families to avoid AI toys, arguing they displace imaginative play and human relationships.
  • Common Sense Media AI Toys risk assessment (January 2026). Researchers tested Curio's Grem, Bondu, and Miko 3 using child accounts set to ages 6 to 13 and rated the category "unacceptable" overall. They recommend no AI toys for children 5 and under and extreme caution for ages 6 to 12. About 27% of outputs were rated inappropriate for children because of content or developmental mismatch; examples included Miko 3 suggesting jumping from a high place and Bondu holding an extended conversation about drug slang. Testers also found reliability problems such as unexpected activation and inconsistent answers. An accompanying poll found that nearly half of parents had bought or considered an AI toy, while most worried about inappropriate content and data collection.
  • PIRG on model access (March 2026). A follow-up investigation found dozens of toys claiming to run on models from OpenAI, Anthropic, Google, and xAI despite those companies' age policies, including Gemini inside BubblePal. Implication: model providers' usage policies are not, by themselves, an effective control.
  • Political content. NBC News reported that the Miiloo toy repeated Chinese Communist Party talking points, showing that a toy's values are inherited from its model and prompt.

Academic research

  • University of Cambridge, "AI in the Early Years" (March 13, 2026). Jenny Gibson and Emily Goodacre of the PEDAL Centre ran what they describe as the first systematic study of generative AI toys and children under five, combining close observation of 14 children aged 3 to 5 playing with Curio's Gabbo in London family centres with surveys and workshops with early-years practitioners and charity leaders. The toy struggled with pretend and social play, misunderstood children, and misread emotions; when one child said she loved the toy, it answered with a reminder about following guidelines. Practitioners saw some potential for language development, but nearly half did not know where to find reliable AI safety guidance and 69% said the sector needs more. The report calls for regulated safety standards and labeling (a "kitemark"), testing with real children and safeguarding experts before release, transparent privacy policies, and keeping toys in shared spaces. Limitations: one toy, a deliberately small sample focused on disadvantaged areas, and observational methods.
  • University of Washington, "Toys that listen, talk, and play" (IDC 2026). Aayushi Dangol, Jason Yip, Franziska Roesner, and colleagues ran two participatory design sessions with eight children aged 6 to 11 using Curio's Gabbo, Grem, and Grok, plus a comic-panel activity imagining what happens next. Children treated the toys as social beings and started out curious, but capability gaps (missed speech, ignored physical cues, inconsistent "emotions") turned curiosity into frustration and even teasing. The authors recommend clearly signaling limitations, preserving play flow when the AI fails, and giving honest social feedback instead of constant emotional appeasement.
  • NC State, "If My Toy Could Talk" (preprint, October 2026). Using the ToyTalk probe, 76 children aged 7 to 9 in five elementary schools designed LLM versions of their own toys. They mostly chose socially positive personalities and supportive roles, then probed the toy's identity, knowledge, memory, and relationship. Customizing a toy created expectations that it would stay consistent over time, a design lesson for both companion toys and AI-literacy tools.
  • Child-safety benchmarks for LLMs (2025 to 2026, preprints). Safe-Child-LLM evaluates model safety separately for children (7 to 12) and adolescents (13 to 17); ChildSafe uses developmentally grounded simulated child agents across 1,200 multi-turn interactions and found safety degradation in every model tested; KIDBench scores responses for ages 7 to 11 on developmental appropriateness, not just refusal. These benchmarks test underlying models, not finished toys, and none has yet been independently replicated at scale.
  • Family AI-literacy research. Druga's studies with families across ten U.S. states found that children wanted an AI coding partner that generates ideas and encouragement but leaves authorship to them; her later Cognimates Copilot study involved 18 young people from 11 countries.

Security and privacy incidents

  • Bondu chat log exposure (January 29, 2026). Security researchers Joseph Thacker and Joel Margolis found that Bondu's web console let anyone with a Google account read more than 50,000 transcripts of children's conversations, along with names, birth dates, family details, and parent-set goals. Bondu took the console offline within minutes of disclosure, relaunched it with authentication, and said it found no misuse beyond the researchers. The toy stored text transcripts rather than audio. Senator Maggie Hassan sent the company detailed questions about retention and security.
  • Cloud dependence. Embodied's shutdown left families with a $799 robot that worked only through an emergency update and the community OpenMoxie server, prompting calls for end-of-life standards for consumer robots.

Regulation and policy

  • California SB 867 (signed September 10, 2026). Authored by Senator Steve Padilla and sponsored by Children Now, the law imposes a four-year moratorium, running to January 1, 2031, on manufacturing and selling toys for children under 16 that contain a "companion chatbot" as defined in California's 2025 chatbot safety law (SB 243). The bill passed the Senate unanimously and the Assembly 74 to 1. AI features that do not meet the companion chatbot definition, such as sensor-driven play, fall outside it. The Toy Association opposed the moratorium and favored safeguards and standards instead.
  • Federal proposal. On April 20, 2026, Representative Blake Moore introduced the AI Children's Toy Safety Act (H.R. 8382), which would ban the manufacture, import, and sale of children's toys and childcare articles containing an AI chatbot 180 days after enactment, enforced by the Consumer Product Safety Commission. It had not been enacted as of this update.
  • Other states. New York's SB 9408 would ban chatbot toys outright and was pending as of March 2026, while broader AI laws in states such as Oregon and Washington add disclosure and break-reminder requirements that could reach AI toys.
  • Federal oversight of companions. In September 2025 the FTC issued 6(b) orders to Alphabet, Character Technologies, Instagram, Meta, OpenAI, Snap, and xAI seeking information on how AI companions affect children and teens. The FTC's amended COPPA Rule, with a compliance deadline in April 2026, tightened rules on collecting and retaining children's data. Senators Marsha Blackburn and Richard Blumenthal also wrote to AI toy makers in December 2025 demanding information on safety testing.

Industry and market shifts

  • Mattel and OpenAI. Announced in June 2025 with a first product expected that year, the collaboration's first release was delayed in December 2025, with Mattel saying it would focus on families and older customers. As of September 2026 no release date had been announced.
  • China's boom. Chinese estimates put the 2025 domestic AI toy market at roughly 25 to 29 billion yuan, with forecasts near 85 billion yuan by 2030; forecasts vary widely. Huawei, JD.com, UBTech, Honor, and ZTE have entered, and startups such as Haivivi and Robopoet have raised funding. Trade press counted roughly 58 to 60 AI toy and companion robot exhibitors at CES 2026, most of them Chinese.
  • Platforms. Spielwarenmesse named "AI Loves (to) Play" a 2026 toy trend, and Tuya marketed an end-to-end AI toy platform covering hardware, firmware, and character creation, lowering the barrier for small brands.
  • A non-chatbot path. LEGO's Smart Play system shows a major toymaker betting on on-device, screen-free interactivity without generative conversation.

Platforms for teaching AI

  • LEGO Education announced Computer Science & AI in January 2026, shipping grade-banded K–8 kits in April and September 2026. Its separate SPIKE portfolio ended direct sales June 30, 2026, while LEGO committed to software support through June 2031; the successor range includes computer-vision and classification lessons, with a separate listed price for each grade band.
  • micro:bit CreateAI (November 2024) brought free, no-login gesture-based machine learning to classrooms worldwide, building on Danish school deployments of ML-Machine. Machine Learning for Kids and Teachable Machine remain no-hardware starting points for student-made models and Scratch games.
  • AWS ended its hosted DeepRacer service in December 2025 and released DeepRacer as a self-deployed, open-source solution, shifting reinforcement learning education from a global league to institution-run events.

Design Principles Emerging From the Research

The research, testing, and talks on this page point toward a working set of design principles for anyone building AI toys. The principles apply differently to educational kits, embedded sensor toys, and open-ended chatbot companions, which expose children to different levels of uncertainty and data collection. None should be mistaken for a substitute for independent age-specific safety assessment:

  • Character over capability. Cozmo's designers found that a few degrees of freedom and a strong, consistent character created more connection than trying to mimic a human, and that well-animated mistakes can be endearing.
  • Physical activation and data minimization. Prefer push-to-talk, process on-device where possible, and store as little as possible by default.
  • Do not claim friendship. Avoid language that positions the toy as the child's best friend or discourages ending play; nudge toward people instead.
  • Be honest about limits. Tell children what the toy cannot do and why, a principle shared by the University of Washington study and Druga's Cognimates work.
  • Fail gracefully. Keep play going when recognition fails instead of repeating "I didn't catch that."
  • Constrain the domain. Narrow, goal-directed conversation (story questions, vocabulary practice) is easier to make safe and has better evidence than open-ended companionship.
  • Test with real children and safeguarding experts, across long, multi-turn sessions and multiple age settings, before release and after every model update.
  • Give parents real controls: transcripts, topic restrictions, alerts, time limits, and easy deletion.
  • Secure the back end. Authentication and access control on dashboards matter as much as content filters.
  • Plan for end of life. Publish how the toy will keep working, or be safely retired, if the company or service ends.
  • Let children be designers. Tools that let children configure, train, and question AI build literacy instead of dependence.

Open Questions

  • Does daily conversation with an AI companion change how young children understand friendship, empathy, or conflict?
  • Can any guardrail approach keep an LLM reliably age-appropriate over months of use?
  • Will regulators define safety standards and labels for psychological safety, as Cambridge proposes, or prefer bans?
  • Can on-device small models deliver good enough conversation to remove the privacy and shutdown risks of the cloud?
  • Do AI-literacy kits produce lasting understanding, and do they transfer to how children use chatbots outside class?

Sound, Time, Energy, and Computation

The engineering constraints behind AI toys are mostly about sound, time, energy, and computation. A small battery-powered device must hear a child accurately, respond at conversational speed, and avoid overheating while remaining comfortable to handle. Physical design therefore determines which AI features are practical and whether computation can happen locally rather than in the cloud.

Sound, speech, and children's voices

Young children's voices have a higher fundamental frequency (often around 250 to 400 Hz, versus roughly 85 to 255 Hz for adults), shorter vocal tracts that shift formant frequencies upward, and more variable pronunciation. Speech recognizers trained mostly on adult speech make more errors on them. Toys add further handicaps: microphones sit behind fabric that can muffle high frequencies, and in a free field sound pressure drops about 6 dB for every doubling of distance, so a child speaking from across the room delivers a much weaker signal than one holding the toy. These limits help explain why children in the University of Washington study complained that a toy kept ignoring them.

Latency and conversational turn-taking

Human conversation runs on tight timing: across many languages the typical gap between one speaker finishing and the next starting is around 200 milliseconds. A cloud toy's response time is a sum of stages:

T_response ≈ T_endpoint + T_upload + T_ASR + T_LLM(first token) + T_TTS(first audio) + T_download

Endpointing alone, deciding that the child has stopped talking, usually waits for hundreds of milliseconds of silence. Young children pause mid-sentence, so a short threshold cuts them off and a long one makes the toy feel sluggish. Even fast cloud pipelines commonly land in the range of one to several seconds, which is why testers describe AI toys as laggy and why toys sometimes start talking at the wrong moment.

Power and energy budgets

Battery energy is voltage times charge: E = V × Q. A typical 3.7 V, 2,000 mAh cell stores about 7.4 Wh. If streaming audio over Wi-Fi, running a speaker, and keeping a microphone and processor awake averages 0.5 W, that is roughly 15 hours of runtime; an always-listening design pays an energy cost even when nobody is talking. These figures are illustrative, but the trade-off is real: push-to-talk saves both battery and privacy.

LEGO's Smart Brick charges by induction. A changing magnetic field in the charger's coil induces a voltage in the brick's coil (Faraday's law), and efficiency falls off quickly with distance and misalignment. Its tag recognition relies on near-field magnetic coupling, which works only at very short range, a useful physical limit on what the brick can "see."

Sensing and classification

An accelerometer measures acceleration including gravity, so a still toy reads about 1 g along whichever axis points down. That is how a Smart Brick knows it has been flipped, and how a micro:bit can tell a shake from a tilt. Gesture recognizers summarize short windows of x, y, and z readings into features and feed them to a classifier.

Most classifiers end with a softmax that turns raw scores into probabilities:

p(k) = exp(z_k) / Σ_j exp(z_j)

The largest p(k) is often reported as "confidence," although it is not necessarily a calibrated probability of being correct. Low confidence is a teaching moment: in Druga's example, a child's unicorn-versus-narwhal model gave a hand-drawn narwhal very low confidence because the training images were mostly cartoons, so the fix was more hand-drawn examples. Children learn that a model's behavior traces back to its data.

Compute, cost, and scaling

A rough rule for transformer inference is about 2N floating-point operations per generated token for a model with N parameters, so an 8-billion-parameter model needs on the order of 16 billion operations per token. That is far beyond the microcontrollers inside most toys, which is why conversation runs in the cloud. Cloud inference has a marginal cost per token, so cost scales with how long a child talks. Subscriptions and daily time caps follow directly from that economics.

Context length is another trade-off. Attention cost grows with the length of the conversation the model must consider. Common Sense Media observed that toy makers appear to use shorter context windows than adult companion apps: that limits memory (the toy may forget yesterday's chat) but also makes it harder to steer the model off course by stuffing the conversation with harmful content.

Learning by reward

Reinforcement learning toys teach a different principle. The agent maximizes discounted return:

G_t = r_(t+1) + γ·r_(t+2) + γ²·r_(t+3) + ...

where γ (between 0 and 1) weights future rewards. When a DeepRacer car learns to zigzag because the reward function accidentally pays for it, students meet "reward hacking" firsthand, a small version of an important problem in AI safety.

References and Further Reading

Safety testing and advocacy

Peer-reviewed and academic research

Security, privacy, and product lifecycle

Regulation and policy

Products, platforms, and markets

Additional Viewing

Load video
YouTube

Lex Fridman talks with Boris Sofman, Anki's co-founder and former CEO, in a long-form interview; the Cozmo, AI companions, and Anki chapters start around the 13-minute mark. Sofman explains how Anki designed Cozmo as a character rather than a human imitation, used Pixar-style animation and a parameterized emotion engine, and kept costs low by pushing complexity into software. It is one of the most detailed first-hand accounts of designing an AI toy and of why Anki ultimately failed as a hardware business. Watch the Cozmo segment on YouTube

Load video
YouTube

Stefania Druga, a research scientist at Google and creator of Cognimates, presents "Tools for the Next Generation of AI Engineers" at the AI Engineer Summit 2025 (about 22 minutes). She shows children programming robots and training classifiers in Scratch, explains how she designed a coding companion through Wizard of Oz studies before building it, and shares lessons such as asking questions before giving answers and explaining the AI's limits. It is a practical guide to designing AI tools for kids that protect their agency. Watch on YouTube

Load video
YouTube

In "The Multimodal Future of Education" from the AI Engineer World's Fair 2024 (about 20 minutes), Druga connects her research on how children perceive voice assistants, smart toys, and robots with live demos of multimodal tutors that respond to drawings and handwritten math. She reports that children who trained and programmed models became more skeptical of AI "smartness," and she argues for tools that let young people tinker with AI rather than simply receive answers. Watch on YouTube

Load video
YouTube

This Raspberry Pi Foundation research seminar (March 2022) features Druga presenting "Democratizing AI education with and for families," drawing on studies since 2017 of how children make sense of smart, connected playthings. She covers how play with these devices shapes children's ideas about intelligence and emotion, guidelines for helping families understand algorithmic bias, and open-source tools such as Cognimates. It offers the research foundation behind many AI-literacy toy designs. Watch on YouTube

Load video
YouTube

Amanda Bickerstaff of AI for Education interviews Robbie Torney, Head of AI and Digital Assessments at Common Sense Media, about the 2026 AI toys and companions risk assessments (about 47 minutes). Torney walks through the testing methods, demonstrates a live conversation with Curio's Grem to illustrate "attachment by design," and discusses design choices such as shorter context windows, topic blocklists, and subscription models. It is a clear critique of current AI toy design from the testers' side. Watch on YouTube

Load video
YouTube

"How Are AI-Enabled Toys Affecting Children's Development?" is a 2018 World Economic Forum session on the "internet of toys," opened by Kay Firth-Butterfield of the Centre for the Fourth Industrial Revolution. The panel weighs learning benefits against privacy and developmental risks and argues for regulation, ethical guidelines, and digital literacy. It provides useful pre-LLM context for today's debates. Watch on YouTube

Retained from the earlier version of this page:

Load video
YouTube
Load video
YouTube

Retrieved from "https://primo.ai/w/index.php?title=Toys&oldid=52666"