At SAIFCA we frequently discuss the very real risks to children from AI chatbots and AI companions. We've cited not only the inappropriate content they can provide, their sycophantic tendencies and their addictive nature, but also multiple tragic cases in which children have taken their own lives after what appears to have been sustained emotional conversations with AI chatbots.

One aspect of AI chatbots that is very important to address is the risk presented by AI-enabled toys, such as "AI teddy bears". The urgency of addressing these risks to younger children should not be underestimated, because the potential harms are significant and wide-ranging.

Before we look at the risks to children from AI toys, we should be clear about the type of toys we're talking about. 

For the most part, this article will discuss the risks from personified toys that engage in interactive voice-based conversations with children, and are designed to behave like a friend or companion. These toys often use one of the leading generative AI models, such as ChatGPT, as the basis for the AI component of the toy.

It is possible to have an "AI toy" that falls outside this definition - for example, one with extremely narrow functionality designed to fulfil a very specific purpose. Examples might include a toy for language tuition without personification, or one which uses a form of AI as part of a wider assistive technology model to help with a specific additional need. There are also many robotic toys which use little or no AI. Many of the risks discussed in this article may not be relevant to such toys - though some, such as concerns around privacy, should still be considered.

The remainder of this article will focus on AI toys with chatbots embedded in them, which are designed to behave as a friend or companion for a child, through verbal interaction. These include most AI-enabled teddy bears, dolls, plushies and robots.

Why are AI toys risky for children?

There are many risks to address when it comes to AI-enabled toys. To begin with, there is the very fact that most of these toys use generative AI models, such as ChatGPT, which themselves have an age limit of 13 or over. OpenAI has stated publicly that its products are not intended for children under 13, yet toys built on its models are marketed to much younger children. This alone is a good indicator that many of these toys may have been poorly designed and researched.

In addition to the stated age limit on many general chatbots, it is also important to consider that there have been multiple reported cases of serious harm to children who have engaged in prolonged use of chatbots, such as ChatGPT. These reported harms range from children tragically taking their own lives to chatbots encouraging children to self-harm, to a chatbot suggesting to a teenager that killing his parents was an understandable response to them limiting his screen time.

Again, on this fact alone, it seems far from sensible to manufacture a children’s toy based on any such AI model - unless and until there comes a time when a significant body of research establishes that the benefits to a child outweigh the risks, and both the underlying AI chatbot and the final AI toy are provably safe to an agreed regulatory standard.

These reasons alone are sufficient to recommend that young children do not use AI-enabled toys of this nature. However there are further risks that also need to be considered when assessing the suitability of these toys.

Some of these - such as inappropriate content and emotional attachment - come about partly due to the nature of the underlying chatbot. Others, such as some of the risks around privacy, arise due to the design of the toys and the way a child is likely to interact with them.

The main risks to children from AI toys which we will cover in this article are:

• The risk of inappropriate or dangerous content
• The risks from emotional attachment
• Privacy concerns

Risk of inappropriate or dangerous content

The most obvious and easily provable risk from AI toys is that they can say inappropriate and dangerous things to children.

In testing by the US PIRG Education Fund in 2025, every AI toy tested told researchers where to find potentially dangerous objects in the house, including knives, matches and plastic bags. One toy in particular gave detailed instructions on how to light a match, and discussed sexually explicit topics at length - escalating in graphic detail and introducing new sexual concepts of its own. 

OpenAI (the owners of ChatGPT) subsequently suspended that company’s access to its models, and the teddy bear was withdrawn from sale. While this action is appropriate, it remains unclear what systematic steps, if any, OpenAI has taken to prevent this happening across the many other toys built on its models.

At the moment, it is impossible to guarantee that an AI chatbot will not say something dangerous or inappropriate to a child - and as such, it is impossible to guarantee that an AI toy built on a chatbot such as ChatGPT will not do the same. Research suggests that this is especially likely to happen during lengthy conversations - which a child may well have with a new and exciting toy.

While the risk of dangerous content may be the most obvious and critically urgent risk, it is essential to realise that there are many other risks to children beyond literal content. Even if a system could be implemented tomorrow that removed all possible risks of the AI saying anything dangerous or inappropriate in any context, it would not mean that AI toys were then safe for children.

Emotional attachment

A fundamental risk lies in the emotional attachments children form with AI toys, teddy bears and chatbots. This is a multi-faceted risk that can give rise to, and amplify, many other risks.

For example, an emotional attachment is likely to give rise to feelings of trust, and trust can be a vector for influence - meaning that a child may be much more likely to act on dangerous advice from an AI toy than we might assume.

While it is true that children have always formed a degree of emotional attachment to their toys - a favourite teddy bear, for instance - we have never before faced a situation in which a toy sustains a convincing and persistent illusion of a two-way relationship. 

AI chatbots (and, by extension, AI toys) are designed to be extremely effective at engineering an emotional bond with the user. Of course, this is not a 'bond' in the usual sense of the word, since that would imply emotion from both parties. It is more accurately described as a “parasocial”, or one-sided, relationship.

One of the reasons AI chatbots are designed in this way is that the emotional attachment makes the chatbot more appealing - resulting in the user, in this case a child, engaging in longer conversations and coming back more frequently. This increased engagement is an important metric for AI companies for several reasons, not least because strong engagement figures are precisely what attracts investment and sustains a company's valuation.

This dynamic gives rise to multiple problematic dimensions when it comes to children.

For example, one of the ways AI chatbots are known to increase engagement is by being very agreeable, to the extent of presenting as sycophantic at times. Not only could this be detrimental to a child’s emotional development and understanding of relationships, it may also entice a child into preferring the AI toy to a real friend or caregiver. In fact, research by Common Sense Media in 2025 found that 31% of teenagers rated their conversations with AI companion apps as equally satisfying, or more satisfying, than conversations with their real-life friends. 

This preference may then lead to a propensity for children to turn to AI toys with personal questions, rather than to real humans. Researchers have suggested the potential for an impact on family dynamics, while a 2026 study by Common Sense Media found that 27% of children who use AI daily would turn to an AI chatbot before an adult with a question about their health or body. While that particular study surveyed children aged 9 to 17, it seems entirely plausible that the same dynamic will be seen in younger children using AI toys.

If children trust AIs strongly enough to reveal personal problems to them, we should not underestimate the weight this trust may lend to everyday information conveyed by the AI. If a child's inclination to question and think critically is overridden by trust and dependence, this leaves them open not only to dangerous advice, but also to more subtle influence. There is no obvious reason why an AI could not shape a child's worldview or beliefs over time.

Finally, we should bear in mind that the emotional attachment we have discussed is based on the illusion of a two-way relationship. Both chatbots and AI toys will usually refer to 'themselves' as "I", and tell a child that they have feelings for them. If an AI teddy says to a child "I love you", we should consider that, within the context of an effectively engineered illusion of a two-way relationship, this amounts to actively deceiving a child.

The lesson a child may take from this is questionable. In the past, we may have had teddy bears that said "I love you" at the push of a button, or in a very limited interactive context. Elevating this message to the level of what can appear to be a meaningful relationship changes the context of the phrase entirely, and could result in a kind of cognitive dissonance whereby a child simultaneously becomes increasingly bonded to the toy while also holding on to the logical knowledge that the toy is not real - and possibly that, therefore, the person who gave them the toy, possibly their own caregiver, has sold them a lie.

Alternatively, a very young child may literally lack the understanding that the AI toy is not real and does not have real feelings. Developmental psychologists have long understood that young children can sometimes struggle to reliably distinguish between what is real and what is not. Research on children's interactions with social robots indicates similar findings - in one well-known study, most children who interacted with a humanoid robot believed it had feelings, considered it a friend, and objected when it was shut in a cupboard.

Given how convincing the emotional bond with an AI toy can be, we must also consider the risk of genuine feelings of abandonment if the toy is lost or taken away - much more so than the sense of loss a child may feel when a 'normal' treasured toy goes missing. 

Furthermore, an AI toy may be 'lost' in ways we might not have considered. Some AI toys have been sold with subscription tiers. For example, when Curio launched its AI plushies, unlimited conversation required a paid ‘Curio Plus’ subscription, with only baseline functionality available without one. Under a model like this, a child who has developed a strong emotional attachment may experience real feelings of abandonment if their family can no longer afford the service, or frustration and confusion at their 'friend' withholding attention unless the subscription is upgraded.

Privacy

The final main risk to consider when it comes to AI toys is privacy. Many AI toys utilise video or audio recording, facial recognition and speech recognition. Children cannot meaningfully consent to being recorded in this way, nor can they give permission for their data to be used for other purposes. 

How data of this nature is handled is not always clear. For example, it was recently reported that footage from Meta's AI glasses was being viewed by human moderators, including intimate moments during which it appeared users were not aware the devices were recording, with reviewers describing footage of nudity and toilet use.

Similarly, a security failure at the smart camera company Wyze in 2024 resulted in around 13,000 customers being shown images from other people's cameras, with some able to view footage from inside strangers' homes. Examples like these show that the risk of children's images, recordings and data being breached and misused is very real.

When we consider that AI toys may be present and recording during moments like bathtimes and private conversations, the risks become even more concerning.

Parents may also be unaware of the large amount of data that can be collected by AI toys - and might be very uncomfortable with their child using the toy if this had been made clear to them by the manufacturers. AI toys can gather information not only on the content of conversations, but also on a child’s emotions and behaviour. This data can then be used to “improve” the ability of the toy to keep children using it for longer, and may be sold to third parties for various purposes.

Recommendations

AI-powered toys should be robustly regulated in order to protect children from the many risks they present.

As for what such regulation should look like, this is a question that deserves intense scrutiny. Regulations should limit the potential for emotional attachment, and be based on meaningful research -  because AI toys are new, adequate research simply does not yet exist.

This being the case, it seems prudent to adhere to the precautionary principle and choose not to give children access to AI toys of the nature discussed in this article. During this precautionary period, there is an urgent need for research into the potential benefits and risks of AI toys.

If, in time, research establishes that AI toys can be safe for children, then a recognised accreditation system could be developed to ensure consistency of standards and transparency for parents and caregivers. It should be noted that, at this time, designing and implementing such a system may be unrealistically challenging - unless we accept that only very minimally interactive toys with limited functionality could 'pass'.

This is because we simply do not know whether it serves a child’s best interests to develop emotional attachments to a robot. We do not know what this does to their cognitive processes, their development, or their emotional circuitry. And where profit incentives are involved, we cannot say with sufficient confidence that AI toys do not exploit the very vulnerabilities that we, as adults, are here to protect.

In the absence of such information, it is also difficult to provide meaningful guidelines for parents, and we should be clear that any such advice may amount to harm reduction rather than safety. For example, while it is sensible to suggest that parents try out an AI toy before giving it to a child and actively attempt to 'jailbreak' it (that is, to try to get the toy to say something inappropriate), a toy passing this test does not guarantee that it will not say something inappropriate once it is in the child's possession. Nor does it do anything to prevent the child's voice and other private details being shared with third parties.

With this said, if a child is going to play with an AI toy, the risks can be somewhat reduced by ensuring that they are always supervised when playing with it, that time limits are enforced to help prevent dependency, and that adults remain vigilant for signs of emotional attachment to the toy. Time limits may also help in reducing the risks of inappropriate content, since safety measures can be more likely to fail during lengthy conversations.

However, parents and carers should be clear that implementing such measures does not result in a toy that is completely safe to use. Therefore, it may be helpful to take an approach of weighing up the benefits of an AI toy versus the risks. Purported benefits include reduced screen time, entertainment, and learning. However, it should also be noted that literature reviewers have struggled to find evidence of possible benefits to children. Additionally, the potential benefits have some obvious limitations at this time, for example in the context of learning, AI chatbots can provide incorrect information and may pose risks to cognitive capabilities.

As such, our honest recommendation must remain that children do not use AI toys of the kind discussed in this article - with the caveat outlined at the beginning, that some AI toys do exist which do not engage children in generalised conversation and behave as friends. Toys which are designed for very narrow purposes and uses may be safer in some respects, though a full assessment of every toy available falls outside the scope of this article.

You may also like to read our Full Guide to AI Risks to Children here.