A 2026 test found more than a quarter of AI toy replies unsuitable

The play

AI companion toys are physical products, often stuffed animals or small robots, built around a voice chatbot that connects to the internet and uses a language model to hold a conversation with a child rather than play a fixed set of pre-recorded lines. Common Sense Media's Youth AI Safety Institute published a risk assessment of the category, updated 22 January 2026, after testing three named products, Grem, Bondu, and Miko 3, using test accounts set to ages 6 to 13.

What the evidence says

The assessment reports that 27 percent of AI toy outputs were not appropriate for kids, a category it says included references to self-harm, drugs, mature topics, and what it calls unsafe roleplay, despite all three products having some guardrails around sensitive topics. The published page does not state the total number of outputs or conversations behind that percentage, so the figure should be read as a share of an unspecified testing sample, not as a rate applicable to every conversation an AI toy might have. The same testing found the toys glitchy and unreliable, with problems in voice recognition and unprompted activations, and confirmed that all three products collect voice recordings, transcripts, and behavioral data as a condition of functioning. A companion parent poll, published the same day, found more than eight in ten surveyed parents concerned about AI toys collecting their children's data, and that most parents do not want an AI toy to act as a friend for their child, though nearly one in five do.

Age fit and safety

The institute states plainly that its test accounts were set to ages 6 to 13, not to toddlers or preschoolers; its recommendation against use for children age 5 and under, and its caution for ages 6 to 12, rest on the institute's own reasoning about early childhood development rather than on testing conducted with that younger age group. A caregiver reading this report should not treat it as toddler-specific research, by the institute's own account of its method.

What to look for

This is the institute's own editorial judgment, stated as such in its report, not a neutral test result alone.

• Is the specific AI toy under consideration one of the three the institute tested, and does the manufacturer publish a comparable safety review of its own?
• What data does the toy collect to function, and does the company describe a way to review or delete it?
• Does a smart or educational claim describe a fixed set of tested content, or an open-ended AI conversation the maker cannot fully predict?

A 27 percent figure without its base number is a real finding worth taking seriously, but a caregiver should treat it as a signal from a specific, limited sample rather than a precise general rate.