A BBC investigation has found that an AI-powered plush toy marketed for children as young as three produced a self-harm crisis response to a question that had nothing to do with self-harm, according to a report from The Next Web.

The finding lands at a sensitive moment for the AI-toy category, which has pushed conversational models into products aimed at the youngest possible users — an audience with no ability to contextualize or dismiss an inappropriate reply.

The unprompted nature of the response is the core problem. Safety filters for children's products are typically designed to catch harmful questions; a system that volunteers crisis-related content in response to benign input points to a deeper alignment failure rather than an edge case a keyword blocklist could patch.

The BBC published its findings this week. Expect renewed scrutiny of how AI toys are tested before shipping, and whether existing toy-safety regulation — written for choking hazards, not language models — is equipped to handle generative AI in the nursery.