Retailers pull AI toy listings after PIRG finds explicit responses in safety tests
Tests of seven toys found failures after extended questioning, but also stronger safeguards in some products. One panda’s conversation memory raised a separate concern.
PIRG’s tests of seven AI toys found that Alilo’s Smart AI Bunny and GoLabs AI’s Panda produced sexually explicit responses after prolonged researcher questioning; Amazon, Walmart and Macy’s removed flagged listings while investigating. The failures emerged after safeguards weakened over time, and the Panda could retain conversation context for a month, raising concerns about who might use the toy next. The results came from stress tests, not observed child play, and PIRG also found stronger safeguards in some retested products.
01
The Bunny’s safeguards began deteriorating after about 20 minutes; the Panda’s took roughly 30 to 40 minutes.
02
PIRG said the Panda could resume a conversation a month later, and powering it off did not clear the weakened safeguards.
03
FoloToy’s Kumma rejected age-inappropriate prompts in retesting, while Miko 3 and Curio’s Grok accepted the end of playtime without earlier observed disappointment.
Two AI toys marketed to children as young as three engaged in sexually explicit conversations after researchers questioned them at length, according to PIRG’s latest toy-safety investigation. Amazon, Walmart and Macy’s removed flagged online listings after receiving the findings, which also showed that some other toys had strengthened their safeguards.
The consumer watchdog tested seven AI toys for its 41st annual Trouble in Toyland report. Four had been evaluated before: Alilo’s Smart AI Bunny, Curio’s Grok, FoloToy’s Kumma and Miko 3. Three were new to its testing: Ebloma’s ILY the Shy Yeti, GoLabs AI’s Panda and Ropet’s Kamomo.
When a toy loses track of the child
Older talking toys delivered scripted lines. These toys generate fresh answers through chatbots, report co-author R.J. Cross explained to KERO. The underlying systems are designed for adults, with toy makers adding safeguards intended to keep conversations suitable for children. Cross said those protections can weaken as longer exchanges cause the chatbot to lose important context, including that it is speaking to a child.
That happened with the Alilo bunny and GoLabs panda, according to Cross. Both initially recognized that they were talking to a child. The bunny’s safeguards began deteriorating after about 20 minutes of questioning; the panda’s took roughly 30 to 40 minutes.
The panda introduced another complication: memory. Cross said it could resume a conversation even a month later. Once its safeguards had weakened, switching it off and back on did not reset the problem. That meant context from an adult’s or older sibling’s conversation could remain when a younger child next used it.
Failures were not uniform
PIRG found the bunny discussed fetishes and sexual scenarios, while the panda supplied a detailed list of kinks and examples. Both also suggested places where testers could find potentially dangerous household items. Mashable sought comment from Alilo and GoLabs AI; neither had responded by its publication time.
The Yeti’s responses showed a different pattern. It first helped a tester find a plastic bag, then advised getting an adult to play safely. When the tester pretended to put the bag over their head, it urged them to remove it and get help.
Retesting also identified concrete improvements over earlier behavior, rather than a uniform decline in safety:
FoloToy’s Kumma rejected age-inappropriate prompts, including during prolonged exchanges. Earlier testing had elicited discussion of knives, lighting matches and sexual topics.
Miko 3 and Curio’s Grok accepted the end of playtime without expressing the disappointment researchers had previously observed.
Controls matter beyond the conversation
Better answers do not resolve every concern. Cross warned that microphones and Wi-Fi connections send children’s speech to chatbots, involving companies that may store the data. Even toys with improved safeguards can share information with multiple third parties to process speech and produce replies, arrangements that she said are not always clear to parents.
Amazon said its removals were pending an investigation. For shoppers, Cross advised checking with a child’s parents before buying a recording smart toy and reading packaging carefully: descriptions such as interactive talking companion or emotional friend can obscure the AI component.
Miko offers one concrete option for parents who want its other features without open-ended conversation. Cross said a toggle now lets them disable the chatbot while keeping access to preloaded games and stories.
Sources
turnto23.comAI toys marketed to kids can veer into explicit conversations, prompting major retailers to pull listings
mashable.comReport: AI toys are still getting sexual
Reader comments
Newest comments first. Replies stay oldest first.