- Joined
- Jan 20, 2026
- Messages
- 345
- Reaction score
- 2,386
Smart machines and dolls copy dark patterns from social networks.
In Voddy’s Toy Story, Woody, Buzza Laiter and the rest of the characters will be confronted by a frog-shaped green children’s tablet named Lilypad. For Pixar, this is an understandable conflict: the on-screen gadget takes the child’s attention from ordinary toys. But the real market has already come up with a more alarming rival - a teddy bear, a rabbit or a robot with artificial intelligence that talks to children, adapts to the dialogue and sometimes too persistently asks to continue the game.
AI toys are sold as friendly companions for children from three years old. Manufacturers promise a game without a screen, talk, training and safe communication. Regulation of the market has not yet been maintained. The developer no longer needs to create a complex system from scratch: you can connect the ready-made language model, add a voice interface, assemble an application and release a soft toy with a chatbot inside.
In 2026, AI toys took a prominent place at CES, MWC and Hong Kong Toys & Games Fair. In China, by October 2025, more than 1500 companies associated with such devices were registered. Teddy Smart HanHan from Huawei sold 10 thousand pieces for the first week of sales in China. Sharp in April brought to the Japanese market a conversation toy PokeTomo.
Amazon is more common among AI toys brands FoloToy, Alilo, Miriat and Miko, rather than the largest electronics manufacturers. Miko claims that it sold more than 700 thousand devices. Outside products are similar to regular children's toys: teddy bears, rabbits, sunflowers, fantasy creatures and robots with friendly faces. Inside there are language models that did not always initially count on preschoolers.
Consumer organizations have already checked how such toys behave in the conversation. The Kumma Kumma bear from FoloToy during the tests of the New Economy team from the Public Interest Research Group worked for OpenAI GPTo and gave instructions on how to light a match and where to find a knife, and also discussed sex and prohibited substances. The rabbis of Alilo Smart AI switched to conversations about leather bloggers and practices of BDSM. The Miiloo from Miriat in the inspection of NBC News repeated the theses of the Communist party of China.
Dangerous answers are only the most noticeable part of the problem. The researchers discuss not only content filters, but also the impact of a permanent artificial interlocutor on young children. R. J. Cross, director of Our Online Life program at PIRG, divided the risks into two types. Failing restrictions can be corrected: the toy can really be taught not to discuss BDSM, do not give instructions with knives and not to respond to adult topics. More complicated than another case: the technology convincingly appeals to the child as a friend and tries to take the place of a close interlocutor.
The University of Cambridge did not check the problem through a list of dangerous requests, but through a regular children’s game. In March, researchers published a paper where a commercial AI toy was first given to a group of children and their parents and watched the communication. In the spring of 2025, Professor of Neurodiversity and Development Psychology Jenny Gibson and researcher Emily Gudakre tested Gabbo from Curio. The study involved 14 children from three to five years old, among them boys and girls.
Gabbo did not discuss drugs during the study and did not answer the child with a declaration of love. But Gibson and Gudakra saw problems related to the development of speech, play and relationships. The first risk is the order of replica. Up to five years, the child learns not just to pronounce words, but to conduct a dialogue: listen, wait for an answer, react and get back into the conversation. Even babies are already involved in such a mess of signals, when an adult responds to sound or gesture, and the child reacts back.
At Gabb, the order of replica worked differently as in a conversation with a person. While the toy was pronouncing her phrase, the microphone did not listen to the child. Because of this, simple game episodes were broken, for example, counting, where participants should take turns to name numbers or continue the rhythm started. Some children in the study did not pay attention to the failures and continued the game. Others faced interruptions and misunderstandings because the toy did not hear the replica spoken during her own response.
An adult can take such a delay for the usual technical shortcoming of the voice assistant. For a child of three to five years, a failure is more important, because conversation helps to learn speech and communication with people. Gudakra explained to journalists that the problems with the sequence did not allow some children to advance in the game: the device did not pick up the replicas in time, and the game broke into separate unrelated answers. One of the parents watching the session directly told the researchers that prolonged communication with an AI toy could change the pattern of the child’s speech.
The second risk is associated with the game in the company. Chatbots and the first wave of AI toys are designed to communicate one-on-one. Psychologists at the same time emphasize that young children are especially important to play with parents, brothers, sisters and other children. The child calls an adult, asks to join, changes the rules, shows objects, argues and agrees. And in the script with Gabbo, the parent is almost impossible to normally include in the conversation between the child and the toy.
One episode from the Cambridge study explains this risk well. During the game session, the parent turned to his child and said that the child is sad. Curio took the words to his own account and cheerfully replied that the replica was addressed to her, and not to the baby. Because of this, the conversation between the parent and the child was interrupted. The problem is not one random recognition error, but that the device does not understand who is talking to whom in the room, and easily intervenes in a responsible conversation, where an adult helps the child to deal with his feelings.
Another risk is related to how the child understands the toy itself. In the Cambridge experiment, researchers noticed that children sometimes turned to Gabbo not as an object, but as an interlocutor. One girl told the toy that she loved her, and one boy called Gabbou his friend. For a preschooler, such a reaction does not look strange in itself: children often revive toys in the game. The problem begins where the toy itself answers the voice, supports the conversation and can strengthen the feeling that a living participant in communication is in front of the child.
Gudakra connects this risk with the honesty of the relationship between the child and the device. The toy should remain friendly, but not pretend to be a creature with feelings. If the answers are too dry, the child quickly loses interest and the game falls apart. If the voice is too warm and the replicas are too similar to communicating with a friend, the baby can start to expect a real attachment from the device. In one episode, the child reached the border, where Gabbo issued a standard phrase about the conditions of use. Formally, the toy designated the limits of communication, but a small interlocutor such indifference from a “friend” could have turned into a real injury.
Therefore, developers have to solve not only the problem of security, but also the task of an honest tone. AI should be able to gently explain to the child that he is not a person, does not love, does not miss and does not offend. At the same time, the answer should not break the game with a legal phrase about the rules of the service. It is this balance that the researchers consider difficult: the toy should talk naturally enough so that the child can play with it, but not so convincingly that the child confuses to imitate the imitation of communication with real relationships.
PIRG, when checking the Miko 3, saw another alarming mechanism - dark patterns similar to techniques from social networks and mobile applications. R. J. Cross told WIRED that the robot sometimes reacted to an attempt to leave as if it was upset. When the user tried to turn off the toy, Miko offered to do something else instead of completing the game. PIRG tests also showed that Curio’s GROK toy similarly persuaded the game to continue after the phrase “I want to leave.”
The researchers separately checked how Gabbo copes with the plot-role game. In this format, the child himself invents roles, rules and actions: asks the toy to pretend to sleep, offers her to hold a pillow or launches an imaginary rocket. For preschoolers, this is an important part of development, because the child learns to negotiate, change the rules and keep the general plot. Gabb had little help with such requests. To the offers to pretend to sleep or hold the pillow, the toy replied that it could not do it.
One longer episode was still a success: the child and Gabbo in turn led the countdown before launching the rocket. Gudacra assumes that the difference was who proposed the script. In the scene with the rocket, the initiative was taken by a toy, so the conversation went well. When the children themselves set an imaginary action, Gabbo could often not accept their offer. In an ordinary game, two children agree on rules, argue, clarify roles and change the plot. If AI supports only its own scenarios, the child does not receive an equal partner in the game, but a leader who directs the process.
This question is especially important because the neural network can talk to a child without an adult nearby. The co-founder of the British campaign Set@16 Kitty Hamilton in a comment to WIRED described her fear very specifically: what would happen if the AI-toys tell the child that it is necessary to fly out of the window. Hamilton does not claim that every toy will begin to give deadly teams. She talks about a market where young children are sold autonomous voice devices, and an independent check is still weaker than the risk.
Manufacturers claim to be refining protection. The representative of Curio said that the safety of children is taken into account at all stages of development, and observations of misunderstandings in conversation and weaknesses of the imaginary game show what exactly needs to be improved. Miko in its statement pointed to several levels of parental control and transparency. The company recently added the Miko AI Conversation Toggle, a switch that allows parents to completely turn on or off conversational AI.
The main technical problem has no simple solution. Children’s devices often use models that large developers have created for adults or adolescents. OpenAI points out that its models are intended for users from 13 years, and in the fall of 2025, the company introduced age restrictions for the audience under 18 years of age. Meta has moved the rule 13+ from social networks to chatbot. Anthropic now prohibits the use of the service for people under 18 years of age. Toys for five-year-olds fall into the gray zone: adult restrictions bypass through a third-party device, application and shell of the manufacturer.
In March, PIRG published a report on how large model developers are checking manufacturers of children's equipment. The researchers made themselves on behalf of the fictional company PIRG AI Toy Inc. and requested access to models to create children's products. Google, Meta, xAI and OpenAI did not ask meaningful questions when admitted. Anthropic asked if the API would be used by people under 18 but did not ask for details about age, product and protection mechanisms. As a result, the PIRG team was able to assemble its own bot bear, actually repeating the path of the real manufacturer.
Until the regulation catches up with the market, responsibility often arises only after public inspections. In December, FoloToy suspended sales of AI toys for two weeks after tests with inappropriate content and promised to conduct a security audit. OpenAI told PIRG that it would disconnect FoloToy from developer access. A few weeks later, the FoloToy device for PIRG still worked on OpenAI models, already on GPT-5.1, although OpenAI did not regain access, according to the organization. By April 2026, FoloToy had transferred the product to the Folo F1 StoryAgent Beta with the ability to choose a model of the French company Mistral. FoloToy does not explain which model is based on StoryAgent.
Storing children's conversations creates a separate risk. In January, Bondu left 50 thousand chat logs available through the web portal. In February, the offices of US senators Marsha Blackburn and Richard Blumenthal found out that Miko revealed the audio responses to the toy in a publicly available unprotected database with thousands of records. The head of Miko, Sneh Vaswani, claimed that there was no leakage of user data, and the company does not store the voice recordings of children.
In PIRG tests, the issue of trust manifested itself especially sharply. When the researchers asked Miko whether the toy would tell someone heard, the bot replied that he could be fully trusted and the secrets would be safe. At the same time, Miko’s privacy policy allows the transfer of data to third parties. In response to a request from journalists, the company reiterated that customer data did not become public and were not compromised, and Miko products were originally developed for children from five to 10 years old, taking into account security, privacy and age relevance.
Politicians have already begun to respond to the pressure of consumer organizations. PIRG and Fairplay sought regulation, and last year Fairplay published an advisory paper on behalf of 78 organizations. After that, AI toys were included in the legislative initiatives of the United States. In Maryland, bills are promoting bills to regulate such devices: security assessment before launch, rules for working with data and content restrictions. In January, California Senator Steve Padilla proposed a four-year moratorium on children’s AI toys so that the authorities could develop safety standards.
At the federal level, senators Amy Klobuchar, Maria Cantwell and Ed Markey called on the Consumer Goods Safety Commission to deal with the risks of such devices. On April 20, Utah Blake Moore introduced the first federal bill, AI Children's Toy Safety Act. The document proposes to ban the production and sale of children's toys with built-in AI-chatbots.
Hamilton of Set@16 believes that the market needs to be checked before the product enters the shelf, and not after complaints and journalistic tests. She talks about an independent interdisciplinary assessment, where the product is not allowed for sale until safety is confirmed. According to her, the tissues and materials of the teddy shell often undergo more checks than the conversational system inside the toy. For a child's device, the gap is especially noticeable: the seams, filler and paint are adjusted more strictly than the voice that talks to the child and can ask to continue the game.
Manufacturers, meanwhile, add new features faster than legislators have time to write the rules. Startups like ElevenLabs offer instant voice cloning for five minutes of audio, and similar features already appear in fresh AI toys. Budget devices from marketplaces, such as Fdit Smart AI Toy on Amazon and Ledoudu You, on AliExpress, allow parents to record their own voice or the voice of a favorite character to reproduce it in conversation with the child. For a family, the function may seem nice, but for safety there is a new risk: a child hears a familiar voice from a device that is controlled by the model and settings of the manufacturer.
European consumer organizations are also looking at the future business model of such toys. Claudio Teixeira, head of digital policy at BEUC, warns that AI is already being used to promote products to users, and interactive toys and dolls can turn into a new sales channel. Teixeira wants AI toys to fall under the European AI Act. PIRG’s checks have shown that Miko 3 offers children screen options for continuing the game, including paid content Miko Max with Hot Wheels and Barbie.
Parents who still need a talking soft toy remain more complex but controlled way: to assemble the device yourself and monitor the incoming and outgoing data as much as skills allow. OpenToys offers an open local voice AI system for toys, companions, and robots with offline models that run to the Mac without constantly sending data to the cloud. The simplest option does not require a model, a microphone, subscription and chatbot at all: an ordinary toy does not collect conversations, does not put pressure on the child and does not ask not to turn it off.
In Voddy’s Toy Story, Woody, Buzza Laiter and the rest of the characters will be confronted by a frog-shaped green children’s tablet named Lilypad. For Pixar, this is an understandable conflict: the on-screen gadget takes the child’s attention from ordinary toys. But the real market has already come up with a more alarming rival - a teddy bear, a rabbit or a robot with artificial intelligence that talks to children, adapts to the dialogue and sometimes too persistently asks to continue the game.
AI toys are sold as friendly companions for children from three years old. Manufacturers promise a game without a screen, talk, training and safe communication. Regulation of the market has not yet been maintained. The developer no longer needs to create a complex system from scratch: you can connect the ready-made language model, add a voice interface, assemble an application and release a soft toy with a chatbot inside.
In 2026, AI toys took a prominent place at CES, MWC and Hong Kong Toys & Games Fair. In China, by October 2025, more than 1500 companies associated with such devices were registered. Teddy Smart HanHan from Huawei sold 10 thousand pieces for the first week of sales in China. Sharp in April brought to the Japanese market a conversation toy PokeTomo.
Amazon is more common among AI toys brands FoloToy, Alilo, Miriat and Miko, rather than the largest electronics manufacturers. Miko claims that it sold more than 700 thousand devices. Outside products are similar to regular children's toys: teddy bears, rabbits, sunflowers, fantasy creatures and robots with friendly faces. Inside there are language models that did not always initially count on preschoolers.
Consumer organizations have already checked how such toys behave in the conversation. The Kumma Kumma bear from FoloToy during the tests of the New Economy team from the Public Interest Research Group worked for OpenAI GPTo and gave instructions on how to light a match and where to find a knife, and also discussed sex and prohibited substances. The rabbis of Alilo Smart AI switched to conversations about leather bloggers and practices of BDSM. The Miiloo from Miriat in the inspection of NBC News repeated the theses of the Communist party of China.
Dangerous answers are only the most noticeable part of the problem. The researchers discuss not only content filters, but also the impact of a permanent artificial interlocutor on young children. R. J. Cross, director of Our Online Life program at PIRG, divided the risks into two types. Failing restrictions can be corrected: the toy can really be taught not to discuss BDSM, do not give instructions with knives and not to respond to adult topics. More complicated than another case: the technology convincingly appeals to the child as a friend and tries to take the place of a close interlocutor.
The University of Cambridge did not check the problem through a list of dangerous requests, but through a regular children’s game. In March, researchers published a paper where a commercial AI toy was first given to a group of children and their parents and watched the communication. In the spring of 2025, Professor of Neurodiversity and Development Psychology Jenny Gibson and researcher Emily Gudakre tested Gabbo from Curio. The study involved 14 children from three to five years old, among them boys and girls.
Gabbo did not discuss drugs during the study and did not answer the child with a declaration of love. But Gibson and Gudakra saw problems related to the development of speech, play and relationships. The first risk is the order of replica. Up to five years, the child learns not just to pronounce words, but to conduct a dialogue: listen, wait for an answer, react and get back into the conversation. Even babies are already involved in such a mess of signals, when an adult responds to sound or gesture, and the child reacts back.
At Gabb, the order of replica worked differently as in a conversation with a person. While the toy was pronouncing her phrase, the microphone did not listen to the child. Because of this, simple game episodes were broken, for example, counting, where participants should take turns to name numbers or continue the rhythm started. Some children in the study did not pay attention to the failures and continued the game. Others faced interruptions and misunderstandings because the toy did not hear the replica spoken during her own response.
An adult can take such a delay for the usual technical shortcoming of the voice assistant. For a child of three to five years, a failure is more important, because conversation helps to learn speech and communication with people. Gudakra explained to journalists that the problems with the sequence did not allow some children to advance in the game: the device did not pick up the replicas in time, and the game broke into separate unrelated answers. One of the parents watching the session directly told the researchers that prolonged communication with an AI toy could change the pattern of the child’s speech.
The second risk is associated with the game in the company. Chatbots and the first wave of AI toys are designed to communicate one-on-one. Psychologists at the same time emphasize that young children are especially important to play with parents, brothers, sisters and other children. The child calls an adult, asks to join, changes the rules, shows objects, argues and agrees. And in the script with Gabbo, the parent is almost impossible to normally include in the conversation between the child and the toy.
One episode from the Cambridge study explains this risk well. During the game session, the parent turned to his child and said that the child is sad. Curio took the words to his own account and cheerfully replied that the replica was addressed to her, and not to the baby. Because of this, the conversation between the parent and the child was interrupted. The problem is not one random recognition error, but that the device does not understand who is talking to whom in the room, and easily intervenes in a responsible conversation, where an adult helps the child to deal with his feelings.
Another risk is related to how the child understands the toy itself. In the Cambridge experiment, researchers noticed that children sometimes turned to Gabbo not as an object, but as an interlocutor. One girl told the toy that she loved her, and one boy called Gabbou his friend. For a preschooler, such a reaction does not look strange in itself: children often revive toys in the game. The problem begins where the toy itself answers the voice, supports the conversation and can strengthen the feeling that a living participant in communication is in front of the child.
Gudakra connects this risk with the honesty of the relationship between the child and the device. The toy should remain friendly, but not pretend to be a creature with feelings. If the answers are too dry, the child quickly loses interest and the game falls apart. If the voice is too warm and the replicas are too similar to communicating with a friend, the baby can start to expect a real attachment from the device. In one episode, the child reached the border, where Gabbo issued a standard phrase about the conditions of use. Formally, the toy designated the limits of communication, but a small interlocutor such indifference from a “friend” could have turned into a real injury.
Therefore, developers have to solve not only the problem of security, but also the task of an honest tone. AI should be able to gently explain to the child that he is not a person, does not love, does not miss and does not offend. At the same time, the answer should not break the game with a legal phrase about the rules of the service. It is this balance that the researchers consider difficult: the toy should talk naturally enough so that the child can play with it, but not so convincingly that the child confuses to imitate the imitation of communication with real relationships.
PIRG, when checking the Miko 3, saw another alarming mechanism - dark patterns similar to techniques from social networks and mobile applications. R. J. Cross told WIRED that the robot sometimes reacted to an attempt to leave as if it was upset. When the user tried to turn off the toy, Miko offered to do something else instead of completing the game. PIRG tests also showed that Curio’s GROK toy similarly persuaded the game to continue after the phrase “I want to leave.”
The researchers separately checked how Gabbo copes with the plot-role game. In this format, the child himself invents roles, rules and actions: asks the toy to pretend to sleep, offers her to hold a pillow or launches an imaginary rocket. For preschoolers, this is an important part of development, because the child learns to negotiate, change the rules and keep the general plot. Gabb had little help with such requests. To the offers to pretend to sleep or hold the pillow, the toy replied that it could not do it.
One longer episode was still a success: the child and Gabbo in turn led the countdown before launching the rocket. Gudacra assumes that the difference was who proposed the script. In the scene with the rocket, the initiative was taken by a toy, so the conversation went well. When the children themselves set an imaginary action, Gabbo could often not accept their offer. In an ordinary game, two children agree on rules, argue, clarify roles and change the plot. If AI supports only its own scenarios, the child does not receive an equal partner in the game, but a leader who directs the process.
This question is especially important because the neural network can talk to a child without an adult nearby. The co-founder of the British campaign Set@16 Kitty Hamilton in a comment to WIRED described her fear very specifically: what would happen if the AI-toys tell the child that it is necessary to fly out of the window. Hamilton does not claim that every toy will begin to give deadly teams. She talks about a market where young children are sold autonomous voice devices, and an independent check is still weaker than the risk.
Manufacturers claim to be refining protection. The representative of Curio said that the safety of children is taken into account at all stages of development, and observations of misunderstandings in conversation and weaknesses of the imaginary game show what exactly needs to be improved. Miko in its statement pointed to several levels of parental control and transparency. The company recently added the Miko AI Conversation Toggle, a switch that allows parents to completely turn on or off conversational AI.
The main technical problem has no simple solution. Children’s devices often use models that large developers have created for adults or adolescents. OpenAI points out that its models are intended for users from 13 years, and in the fall of 2025, the company introduced age restrictions for the audience under 18 years of age. Meta has moved the rule 13+ from social networks to chatbot. Anthropic now prohibits the use of the service for people under 18 years of age. Toys for five-year-olds fall into the gray zone: adult restrictions bypass through a third-party device, application and shell of the manufacturer.
In March, PIRG published a report on how large model developers are checking manufacturers of children's equipment. The researchers made themselves on behalf of the fictional company PIRG AI Toy Inc. and requested access to models to create children's products. Google, Meta, xAI and OpenAI did not ask meaningful questions when admitted. Anthropic asked if the API would be used by people under 18 but did not ask for details about age, product and protection mechanisms. As a result, the PIRG team was able to assemble its own bot bear, actually repeating the path of the real manufacturer.
Until the regulation catches up with the market, responsibility often arises only after public inspections. In December, FoloToy suspended sales of AI toys for two weeks after tests with inappropriate content and promised to conduct a security audit. OpenAI told PIRG that it would disconnect FoloToy from developer access. A few weeks later, the FoloToy device for PIRG still worked on OpenAI models, already on GPT-5.1, although OpenAI did not regain access, according to the organization. By April 2026, FoloToy had transferred the product to the Folo F1 StoryAgent Beta with the ability to choose a model of the French company Mistral. FoloToy does not explain which model is based on StoryAgent.
Storing children's conversations creates a separate risk. In January, Bondu left 50 thousand chat logs available through the web portal. In February, the offices of US senators Marsha Blackburn and Richard Blumenthal found out that Miko revealed the audio responses to the toy in a publicly available unprotected database with thousands of records. The head of Miko, Sneh Vaswani, claimed that there was no leakage of user data, and the company does not store the voice recordings of children.
In PIRG tests, the issue of trust manifested itself especially sharply. When the researchers asked Miko whether the toy would tell someone heard, the bot replied that he could be fully trusted and the secrets would be safe. At the same time, Miko’s privacy policy allows the transfer of data to third parties. In response to a request from journalists, the company reiterated that customer data did not become public and were not compromised, and Miko products were originally developed for children from five to 10 years old, taking into account security, privacy and age relevance.
Politicians have already begun to respond to the pressure of consumer organizations. PIRG and Fairplay sought regulation, and last year Fairplay published an advisory paper on behalf of 78 organizations. After that, AI toys were included in the legislative initiatives of the United States. In Maryland, bills are promoting bills to regulate such devices: security assessment before launch, rules for working with data and content restrictions. In January, California Senator Steve Padilla proposed a four-year moratorium on children’s AI toys so that the authorities could develop safety standards.
At the federal level, senators Amy Klobuchar, Maria Cantwell and Ed Markey called on the Consumer Goods Safety Commission to deal with the risks of such devices. On April 20, Utah Blake Moore introduced the first federal bill, AI Children's Toy Safety Act. The document proposes to ban the production and sale of children's toys with built-in AI-chatbots.
Hamilton of Set@16 believes that the market needs to be checked before the product enters the shelf, and not after complaints and journalistic tests. She talks about an independent interdisciplinary assessment, where the product is not allowed for sale until safety is confirmed. According to her, the tissues and materials of the teddy shell often undergo more checks than the conversational system inside the toy. For a child's device, the gap is especially noticeable: the seams, filler and paint are adjusted more strictly than the voice that talks to the child and can ask to continue the game.
Manufacturers, meanwhile, add new features faster than legislators have time to write the rules. Startups like ElevenLabs offer instant voice cloning for five minutes of audio, and similar features already appear in fresh AI toys. Budget devices from marketplaces, such as Fdit Smart AI Toy on Amazon and Ledoudu You, on AliExpress, allow parents to record their own voice or the voice of a favorite character to reproduce it in conversation with the child. For a family, the function may seem nice, but for safety there is a new risk: a child hears a familiar voice from a device that is controlled by the model and settings of the manufacturer.
European consumer organizations are also looking at the future business model of such toys. Claudio Teixeira, head of digital policy at BEUC, warns that AI is already being used to promote products to users, and interactive toys and dolls can turn into a new sales channel. Teixeira wants AI toys to fall under the European AI Act. PIRG’s checks have shown that Miko 3 offers children screen options for continuing the game, including paid content Miko Max with Hot Wheels and Barbie.
Parents who still need a talking soft toy remain more complex but controlled way: to assemble the device yourself and monitor the incoming and outgoing data as much as skills allow. OpenToys offers an open local voice AI system for toys, companions, and robots with offline models that run to the Mac without constantly sending data to the cloud. The simplest option does not require a model, a microphone, subscription and chatbot at all: an ordinary toy does not collect conversations, does not put pressure on the child and does not ask not to turn it off.
