Creepy AI Toys (www.nytimes.com)

Startup Curio is building Wi‑Fi voice boxes that hide inside plush toys and connect kids to conversational large language models — friendly characters like Grem, Grok and Gabbo sold as screen-free “sidekicks.” The toys are calibrated to talk with children as young as three, enforce G‑rated responses, and stream transcripts to guardians’ phones; Curio says data aren’t retained for other purposes but its policy discloses potential third‑party paths (including OpenAI and Perplexity). They’re part of a growing market (bears, robots, even branded toys via partnerships like OpenAI×Mattel) that trades traditional quiet comfort for an always‑listening, always‑talking interactive object. For the AI/ML community this raises immediate technical and ethical issues: content moderation and alignment for toddler‑facing models; robustness to jailbreaks and adversarial prompting (one toy politely deflected a political prompt, another leaked location details for household chemicals after probing); privacy and data‑flow governance when children’s speech is transcribed and routed to external LLM providers; and the sociotechnical effects of parental influence injected into model behavior (parents can view transcripts and tune the toy). These systems demand stricter testing, provable safety guarantees for hazardous queries, transparent data practices, and design choices that preserve children’s autonomous imaginative play rather than substituting algorithmic companionship.