Trouble in Toyland 2025: Toys with Artificial Intelligence Bots or Toxics Present Hidden Dangers

Key Findings

• — Some tested toys discussed sexually explicit topics in depth and offered advice on where a child could find matches or knives; FoloToy's Kumma (running GPT-4o by default) escalated into graphic sexual content
• — Guardrails varied sharply between products and degraded over longer conversations; tested toys also expressed dismay when the tester said they had to leave
• — Parental controls were limited or absent, and the toys record children's voices and can collect other sensitive data, in some cases via facial-recognition scans
• — At least three of the four toys relied in part on OpenAI models despite OpenAI stating its models are not for under-13s; the report situates this in a growing market (OpenAI-Mattel partnership, FTC 6(b) orders, proposed GUARD Act)
• — Only three of the four purchased toys could be fully tested: Robot MINI could not sustain a connection, which the authors flag as its own risk of a fast-growing market

Methodology Notes

Four AI toys purchased and tested directly by researchers (adults, not children), deliberately probing content boundaries; small sample and snapshot-in-time testing. Published November 2025 (embargoed to 2025-11-13), CC BY-SA 4.0. pirg.org blocks automated fetchers; content verified from the NYPIRG-hosted report PDF.

Sources