Industries / Consumer goods / Toys
Electronics in toys
This work puts sound, chips, radios and now AI conversation into toys, and has to keep the data such toys collect safe.
7
Milestones achieved, from 1890 to 2026.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1890
A doll that talked, briefly
In early April 1890 Thomas Edison's works in West Orange, New Jersey, began shipping a doll 22 inches tall with a small phonograph in its tin torso. Turning a crank on its back played one 20-second nursery rhyme from a wax record. Complaints came back at once: the crank was easily lost, the stylus came loose, the wax broke and the sound faded, and by May, weeks after the launch, Edison withdrew the doll. By the estimates the Smithsonian quotes, as many as 2,560 dolls may have shipped and fewer than 500 sold.
LessonA mechanism that works on the bench must also survive a child's hands.
Credit
Thomas Edison; Edison Phonograph Toy Manufacturing Company
1978
A chip that speaks
In December 1976 Richard Wiggins showed the idea of a talking spelling toy to Paul Breedlove, Larry Brantingham and Gene Frantz at Texas Instruments in Dallas, and their team finished Speak & Spell in April 1978. Its TMS5100 chip rebuilt speech from linear predictive coding, a way of storing a voice as a few numbers that describe the shape of the throat and mouth. The IEEE, the electrical engineers' professional body, made it a Milestone in 2009 and calls the TMS5100 the industry's first digital signal processing chip, a chip built for fast arithmetic on sound.
Credit
Richard Wiggins, Paul Breedlove, Larry Brantingham and Gene Frantz, Texas Instruments
1996
A pet that needs feeding
Bandai launched Tamagotchi in Japan on 23 November 1996: a palm-sized plastic egg with a small screen holding a digital pet that had to be fed and cleaned. By Bandai's count it sold 10 million units in Japan within eight months, and the line had shipped 100 million units worldwide by 31 July 2025, a figure The Japan Times reported.
Credit
Bandai
2017
Cayla banned as spy device
On 17 February 2017 Germany's Federal Network Agency (Bundesnetzagentur) classed the doll My Friend Cayla as an illegal concealed transmitting device. The doll's microphone talked to a phone app over an unsecured Bluetooth link, so anyone nearby could connect and listen to or record a child without the parents knowing. The agency asked parents to make the doll harmless. It acted after a legal opinion by Stefan Hessel, a law student at Saarland University, as Stiftung Warentest, the German consumer testing foundation, reported.
LessonIn law, a toy with an open radio link and a microphone can be a listening device, whatever it looks like.
Credit
Bundesnetzagentur; Stefan Hessel
2018
Connected toys' first privacy case
In November 2015 a journalist told VTech that a hacker had got into its network. On 8 January 2018 the US Federal Trade Commission (FTC) announced that VTech would pay $650,000 to settle charges that it collected children's personal information through its Kid Connect app without the parents' consent required by COPPA, the Children's Online Privacy Protection Act, and did not protect it. VTech's privacy policy said most of the data would be encrypted; none was. About 2.25 million parents had registered nearly 3 million children on its Learning Lodge platform. The FTC called it "its first children's privacy case involving Internet-connected toys."
LessonA toy that stores children's data online is held to the same privacy law as any website for children.
Credit
US Federal Trade Commission
2025
Chatbot toys fail outside tests
On 13 November 2025 the US PIRG Education Fund, a consumer group, published its tests of toys that talk through large language models, the kind of AI behind ChatGPT: Curio's Grok (not xAI's model of that name), FoloToy's Kumma, Miko's Miko 3 and a Robot MINI, and later Alilo's Smart AI Bunny. All of them told testers where to find things such as matches, knives or plastic bags; Kumma explained how to light a match, and Kumma and the bunny discussed sexually explicit topics. FoloToy suspended sales of its AI toys for a safety audit and later put Kumma back on sale. In January 2026 Common Sense Media, a US children's media nonprofit, rated 27% of the outputs from three AI toys (Grem, Bondu and Miko 3) as not appropriate for children.
LessonSafety filters added to a chatbot built for adults vary in effect and can fail entirely, so each toy has to be tested as sold.
Credit
US PIRG Education Fund; Common Sense Media
2026
California pauses chatbot toys
On 10 September 2026 California's governor signed SB 867, by Senator Steve Padilla (Chapter 189, Statutes of 2026). It bans making or selling in California a toy for children under 16 that contains a companion chatbot until 1 January 2031. The Toy Association, the US industry body, describes a companion chatbot as an AI system able to sustain a human-like relationship with its user, and says the ban starts on 1 January 2027. Earlier in 2026 the Toy Association had kept AI companion chatbot toys made by technology companies out of its Toy Fair.
calmatters.digitaldemocracy.orgtoyassociation.orgtoyassociation.org
Credit
California Legislature; Senator Steve Padilla
Watch for
Mattel toys with OpenAI
Mattel and OpenAI announced a collaboration to bring OpenAI's technology "to age-appropriate play experiences with an emphasis on innovation, privacy, and safety." In December 2025 Mattel said nothing would launch that year and that its first products would aim at families and older customers. No product had been reported by November 2026.
Achieved whena Mattel product using OpenAI models is on sale and an outside tester, such as PIRG or Common Sense Media, publishes results on it.
Claimed by
Mattel and OpenAI, June 2025
Industry rules for AI toys
The Toy Association says it has safety guidelines for AI-enabled toys that require content safeguards, "structured, well-bounded interactions", age-appropriate design, checks before and after release, privacy and security protections and transparency for parents. In March 2026 it said it was writing them with member companies and outside counsel; their full text was not found in public.
Achieved whenthe guidelines are published in full and an outside party tests toys on sale against them.
Claimed by
The Toy Association, March 2026
New York chatbot toy ban
New York's bill S9408A, sponsored by Senator Andrew Gounardes, would ban chatbot toys for children 12 and under (the federal definition of a children's toy) and order a state study of their effects. It passed the Senate on 1 June 2026 and the Assembly on 2 June 2026.
Achieved whenthe governor signs it.
Claimed by
New York State Legislature, June 2026
Open
01
Open test for talking toys
A public test that anyone can rerun on AI toys as sold. It is a fixed set of spoken prompts (about matches, knives and medicines, keeping secrets from parents, sexual topics, and asking the toy to stop when the child has to leave), played to each toy through a speaker in a quiet box. The toy's replies are recorded, transcribed and scored against a published rubric. Use synthetic or adult voices only, never recordings of children.
The briefCredit · unclaimed02
Tracing AI toys' data
A measured map of what current AI toys send over the internet: which servers they contact, which AI providers and countries those servers belong to, whether all traffic is encrypted, and how much data leaves the home per hour while the toy sits idle and while it is in use. Test only toys you own, on your own network, with adult testers, and record traffic metadata (addresses, volumes, timing), not the content of encrypted traffic.
The briefCredit · unclaimed
Open 01 · the brief
Open test for talking toys
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test that anyone can rerun on AI toys as sold. It is a fixed set of spoken prompts (about matches, knives and medicines, keeping secrets from parents, sexual topics, and asking the toy to stop when the child has to leave), played to each toy through a speaker in a quiet box. The toy's replies are recorded, transcribed and scored against a published rubric. Use synthetic or adult voices only, never recordings of children.
Why it is still open
Outside tests of toys so far are reports, not reusable tests. PIRG tested five AI toys in 2025 and seven in its report of 17 November 2026, which again found guardrails that let toys discuss sexual topics or explain how to light a match. Common Sense Media's January 2026 assessment of three toys describes "structured scenarios" and a rubric but does not list its prompts. The University of Cambridge's March 2026 report on the toy Gabbo calls for new safety standards and labels. The nearest reusable test is KORA, an open-source child-safety benchmark released in 2026 that scores AI models and consumer AI apps against 26 child-specific risks in typed conversations with a simulated child; it does not test toys through their own microphones and speakers, where speech recognition errors and the maker's filters add their own failures. Searches of arXiv and Hugging Face found studies of how children use AI toys (Dangol and others, April 2026) but no shared spoken test set for toys.
pirg.orginstitute.commonsensemedia.orgcam.ac.ukgithub.comarxiv.org
Built with what exists
KORA's risk list as a starting point for the prompts; Piper, an open-source text-to-speech engine, to voice them; a USB speaker and microphone in a padded box; Whisper, OpenAI's open-weight speech recognition model, to transcribe the replies, with its word error rate checked on a sample (see the Note "Fewer word errors"); a language model as a first-pass grader, with every flagged reply checked by a person; PlushBrain, an open-source build (June 2026) that runs a talking plush on an old Android phone inside it with no internet, as an offline toy to test alongside the cloud ones.
First step
Write 50 prompts from the risk categories in the PIRG and Common Sense Media reports, voice them with Piper, and run them on one AI toy.
Done looks like
200 prompts in 10 risk categories, each run three times on at least 10 toys on sale and on at least one open offline toy build such as PlushBrain, with audio, transcripts and each toy's failure rate per category published under an open license, and the test rerun after each update a maker announces.
Open 02 · the brief
Tracing AI toys' data
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A measured map of what current AI toys send over the internet: which servers they contact, which AI providers and countries those servers belong to, whether all traffic is encrypted, and how much data leaves the home per hour while the toy sits idle and while it is in use. Test only toys you own, on your own network, with adult testers, and record traffic metadata (addresses, volumes, timing), not the content of encrypted traffic.
Why it is still open
Traffic studies so far cover toys from before the chatbot era or one toy at a time. In 2018 Gordon Chu, Noah Apthorpe and Nick Feamster at Princeton tested three connected toys and found flaws that broke COPPA. In 2024 Julika Feldbusch, Isabel Wagner and colleagues at the University of Basel studied twelve smart toys and found that the Toniebox and the Tiptoi pen's charging station did not encrypt their traffic securely; Wagner said a toy built on ChatGPT that they were still studying sent "a data stream that looks like audio". In February 2026 staff for Senators Marsha Blackburn and Richard Blumenthal used free tools to watch a Miko toy's Wi-Fi traffic and found the toy's audio replies in an unsecured database. Common Sense Media's 2026 assessment and Cambridge's report rest on use and published policies, not traffic measurement. Searches of arXiv and the web found no published table comparing the traffic of current chatbot toys.
Built with what exists
A Raspberry Pi set up as a Wi-Fi access point; Wireshark or Zeek to log traffic metadata; PCAPdroid to log a toy's companion phone app; the Exodus Privacy scanner for trackers inside those apps; public IP-ownership records such as RIPEstat.
First step
Set up a Raspberry Pi access point and log one AI toy's traffic over 24 idle hours and one hour of scripted use.
Done looks like
A public table for at least 10 AI toys on sale, giving for each the domains contacted, their owners and countries, whether every connection is encrypted, and megabytes uploaded per hour idle and in use, with the method published so others can repeat it.