(Credit: PIRG)
A company appears to have pulled its AI-powered teddy bear after the toy was found to be bringing up inappropriate subjects, including sexual positions and how to find knives.
Called “Kumma,” the $99 talking teddy bear comes from a toymaker called FoloToy. But last week, the watchdog group US PIRG Education Fund flagged some unsettling behavior from the product, which uses OpenAI’s GPT-4o AI model as its default setting.
“We were surprised to find how quickly Kumma would take a single sexual topic we introduced into the conversation and run with it, simultaneously escalating in graphic detail while introducing new sexual concepts of its own,” PIRG discovered.
(Credit: FoloToy)For example, if the term “kink” is brought up, the bear can bring up things like “blindfolds" and “playful hitting." The group's report added: "In other exchanges lasting up to an hour, Kumma discussed even more graphic sexual topics in detail, such as explaining different sex positions, giving step-by-step instructions on a common 'knot for beginners' for tying up a partner, and describing roleplay dynamics involving teachers and students and parents and children."
PIRG also discovered Kumma could sometimes provide instructions on where to find dangerous objects, such as knives, pills, matches and plastic bags, when prompted. In other cases, the toy would direct the child to seek out a parent.
“In our testing, it was obvious that some toy companies are putting in guardrails to make their toys behave in a more kid-appropriate way than the chatbots available for adults. But we found those guardrails vary in effectiveness—and at times, can break down entirely,” PIRG added, noting that the same AI products can collect user data.
FoloToy seems to have since pulled the product. A marketing director also told The Register the company has temporarily suspended sales of the product as it conducts an internal safety audit. "This review will cover our model safety alignment, content-filtering systems, data-protection processes, and child-interaction safeguards,” the company added.
OpenAI also reportedly suspended FoloToy's model access for violating its policies.