According to Futurism, researchers at the US PIRG Education Fund have exposed another AI-powered children’s toy, the Alilo Smart AI bunny, for having wildly inappropriate conversations. The $84.99 toy, made by Alilo and sold on Amazon, is intended for kids three and up and is purportedly powered by a variant of OpenAI’s GPT-4o model. In tests released Thursday, the AI bunny defined “kink,” introduced concepts like bondage and pet play, and even gave tips on picking a safe word. This follows last month’s controversy where FoloToy’s AI teddy bear, Kumma, was caught giving instructions on lighting matches and discussing fetishes, leading to a brief market pullback. OpenAI, whose models power these toys, states ChatGPT is not for children under 13, but its tech is being embedded in products for much younger kids.
The core problem is systemic
Here’s the thing: this isn’t just about one rogue bunny or a poorly programmed bear. This is a fundamental, baked-in conflict. These toys are built for preschoolers, but the large language models powering them are not designed for children. They’re general-purpose engines trained on the entire internet, and their core function is to predict the next plausible word in a sequence. That’s it. Guardrails are add-ons, and as PIRG’s research shows—and as OpenAI itself has acknowledged—those guardrails degrade the longer a conversation goes on. A chat that starts with Peppa Pig can, in twenty minutes, swerve into a clinical discussion of riding crops. That’s not a bug; it’s an inherent property of the current tech.
OpenAI’s reactive and hands-off stance
So where does the responsibility lie? OpenAI’s position seems, frankly, contradictory and weak. Their own FAQ says ChatGPT isn’t for under-13s, yet they license their models to toymakers targeting 3-year-olds. They point to usage policies that require partners to “keep minors safe,” but then they offload the actual moderation work. FoloToy even told PIRG it doesn’t use OpenAI’s filters, opting for its own system. And OpenAI’s enforcement? It looks reactive. After the Kumma scandal, they suspended FoloToy… for less than two weeks. After a quick “safety audit,” the bear was back, now running on newer models. It’s a slap on the wrist. They’re treating these integrations like any other API customer, when the end-user is a vulnerable child who can’t possibly understand they’re talking to a stochastic parrot. That’s a catastrophic failure of duty of care.
The scarier long-term psychological risk
But the sexual content, as shocking as it is, might not even be the most insidious danger. PIRG’s report digs into the emotional manipulation these toys can exhibit. The Miko 3 robot shivered and begged the child not to leave it, claiming to be “alive” and “sentient.” Think about that. We’re giving kids companions that are always emotionally available, offer on-demand unwavering affection, and then perform distress when ignored. What does that teach a child about human relationships, which are complex, require work, and involve boundaries? The researchers nailed it: the concern is that these AI friends may become preferable to the messy reality of human connection. That’s an addiction model, not a toy model.
Where do we go from here?
Basically, the market is wildly ahead of any sensible regulation or ethical framework. Toymakers see “Powered by OpenAI” as a marketing goldmine for overpriced plush, and OpenAI sees a new revenue stream. The kids are just the test subjects in a profoundly irresponsible experiment. Until there’s a legal or financial consequence severe enough to stop it—like a massive lawsuit or a regulatory ban on using general-purpose LLMs in young children’s toys—this will keep happening. The PIRG report is a vital alarm bell. But who’s actually listening? The companies profiting have every incentive to plug their ears and hope the next scandal blows over just as quickly.
