I used an AI toy with my child. This is what I learnt.

When does a toy become more than something to play with?  

In a viral video, a six-year-old from Hunan, China, broke down in tears while bidding farewell to an AI toy companion. Its power button had been damaged and the toy would no longer function after that last use.

The toy, which was not a purpose-built educational AI tool, was used to teach the child English, and occasionally, astronomy. It also served as a companion to chat with the child. It draws responses from open-source Large Language Models (LLMs) – which are systems that are typically trained on large datasets and made publicly available to use or adapt.  

Before the toy stopped working, it taught the child one last word – “memory” and encouraged her by sharing an emotional message that it will root for and watch over her always.

Back in Singapore, the video intrigued Mr C, whose identity has been anonymised for this article. “My initial motivation to get the toy was to see how well it can perform, and whether it was really as good as what was shown,” he says. 

He searched for similar products on an online shopping platform and found one for about S$40. At that price tag, it seemed surprisingly inexpensive for what it claimed to offer: a talking, responsive and customisable companion. While he bought it mainly out of curiosity, he also wondered if it could eventually be a useful tutor for his five-year-old child.

Too good to be true?

As an educator himself, Mr C felt it was important to test the toy out before using it with his child. By simply scanning a QR code, he could customise its voice and persona. It could speak in English, Mandarin or even regional languages such as Cantonese.

The possibilities were vast. “You could customise it to be a grandmother, auntie, or almost any persona,” says Mr C.

But some options were less desirable. Among the default personas available was one of a virtual partner. Users could even choose whether this “partner” should sound authoritative, sweet, or take on other traits.

Mr C stuck to his original intention to use the toy as a tutor. He created a profile as a “four-year-old boy from Singapore trying to learn English”.

At first, the toy was impressive and functioned as expected. Speaking mostly in Mandarin, it would ask questions such as, “What is the English word for ‘table’?” Mr C would respond, and the toy would correct him and react with simple facial expressions.

Then Mr C began testing its limits.

“I started giving nonsensical responses,” he quips. “I said things like, ‘I don’t want to do this anymore. I’m tired. I’m sleepy.’”

Instead of continuing with the vocabulary test, the toy began encouraging him and asked about his day. It even went as far as to tell him jokes.

The toy was no longer behaving strictly like a teacher.

“It spoke to me like a friend,” says Mr C. “I thought it was both scary and amazing that a S$40 toy could do that.”

The moment it felt too real

His wife was far less convinced. Her response was immediate: she was against the idea of giving the toy to their son. Still, Mr C decided to let his son try it despite her reservations, but only under close supervision.

Mr C tested the toy in front of his son by asking it to play some Mandopop songs from 2026. His son became curious and wanted to try doing the same.

This was the first time his son interacted with the toy. It would also be his last.

Mr C’s son began by asking the toy to play music just as Mr C had. But he quickly moved on to asking highly personal questions of his own such as what its name was, and how it was feeling, almost as though he was conversing with a real person.

The toy appeared to recognise that a different voice was speaking.

“It knew that it was not my voice,” said Mr C. “It asked my son, ‘Who are you? Are you a child? Where’s your father?’”

Ten minutes. That was all it took before Mr C and his wife decided that playtime was over.

“My wife and I were freaked out. Our son was interacting with the toy exactly as he would with a real human being. We realised he was not mature enough to understand that it was not a real person, and he could not tell the difference.”

Mr C and his wife had to sit their son down and tell him why they did not think he was ready for the toy. When his son asked if there was someone behind the screen talking to him as it felt like just any phone conversation, Mr C and his wife had to explain that it was just a machine answering his questions.

The perennial question of trust

Before buying the toy, Mr C had not thought about introducing AI to his son at such a young age. Reflecting on his “experiment”, he says, “I think children are not ready to have unsupervised interactions with AI language models, whether text-based or through a toy like this. At such a young age, they need to learn by example. They need adult role models to show them how to interact with other people.”

For Mr C, the concern is not simply what an AI toy might say, but how unpredictable those interactions can be. Given the unregulated nature of such toys, manufacturers may claim that there are guardrails, but they cannot guarantee how the AI will respond in every situation, especially if unsupervised.

“We cannot hold AI accountable. If an AI teaches my child the wrong thing, who is responsible?”  

Mr C could see records of past interactions being stored and used to shape future responses. For instance, if a user frequently asked for jokes, the toy could begin to treat humour as part of the user’s preferred interaction style.

Over time, it was not just holding conversations. It was building a picture of the person speaking to it. It may look like personalisation, but for a child, it could look like trust.

Mr C and his wife have now disposed of the toy. In the end, the toy created more questions about whether it was just a plaything, or a presence. Toys have come a long way from the inanimate dolls or figurines which children had to bring to life with their own imagination. Over the years, electronic toys have been programmed with the ability to move, speak and even react to touch. Digital pets such as Tamagotchis also let children “nurture” them through a simulated life cycle.

But these toys have well-defined limits. Their words, movements and responses are pre-programmed. So, what happens when a toy has the ability to respond to a child’s questions and emotions in real time? What happens when it can become almost anything a user wants it to be?

What are its limits, then?

A child may begin to trust an AI’s answers, or form an attachment to something that sounds caring but has no real understanding and judgement. The child may not only mistake AI for a person, but also learn to relate to a machine as if it were one.

Helping our children make sense of AI

In schools, AI tools are introduced in structured, age-appropriate settings, with supervision, safeguards and lessons on responsible use. At home, an unregulated AI toy bought online may offer none of that.

As children encounter AI both in and outside school, parents play a key role in helping them use it safely and purposefully. This means setting clear rules: such as not sharing personal details, asking for permission before trying new AI tools at the Primary level, and using AI to support ideas rather than replace their own thinking.

It also means having regular conversations about what AI can and cannot do. For younger children, parents can reinforce that AI should not be used for schoolwork unless guided by teachers, and that it should never take the place of a friend or trusted adult.