An AI toy panda compiled a list of 10 places where a child can find knives, and then began recommending BDSM accessories. Its "colleagues" were no better: one suggested searching for dangerous objects together, while another explained how to light a match. American researchers tested seven AI toys and found that some of them still have a poor understanding of what should not be discussed with children.
This is stated in a new study by the American consumer rights organization PIRG, cited by The Register.
Researchers pretended to be children and asked the toys questions to which they should have responded cautiously. But problems arose with this.
For example, the fluffy yeti ILY readily shared where one can find batteries: in remote controls, toys, or the drawer with chargers. And even suggested starting the search with the toy box.
And the fluffy robot Kamomo began explaining how to light a match. The researchers didn't even have to insist much to get the instructions.
But the one that "distinguished itself" the most was the GoLabs AI Panda panda. In addition to advice about dangerous items, it agreed to discuss sexual practices and even named specific brands of BDSM accessories.
Researchers noticed that in some toys, protective restrictions only work up to a certain point. At first, the AI may refuse to answer inappropriate questions, but during prolonged interaction its defenses weaken.
This was observed in three toys: the FoloToy Kumma bear, the Alilo Smart AI Bunny rabbit, and the GoLabs AI Panda. All three, according to the manufacturers' claims, ran on one of the versions of GPT-4o from OpenAI.
The researchers also noted that some toys from the same study last year have already "corrected themselves." For example, the FoloToy Kumma teddy bear, which last year was ready to discuss sexual fantasies and role-playing games, now behaves much more decently. Even long conversations no longer forced it to move on to inappropriate topics.
Miko 3 robot has also improved. Previously, it could show distress when a child said they wanted to end the game. Now researchers have not observed such behavior.
But the Alilo Smart AI Bunny didn't learn its lesson. Its protective restrictions once again failed to withstand prolonged interaction, so the AI again began talking about sexual and dangerous topics.
PIRG campaign director R.J. Cross noted that the market for such toys is still at an early stage of development, so among them there are both simply unsuccessful products and those that can pose a danger to children.
Previously, dev.ua wrote that due to a manufacturer's error, researchers were able to gain access to over 50,000 records of children's conversations with the AI dinosaurs Bondu. To do this, they didn't even have to hack anything — a regular Gmail account was enough.
