Table of Contents
The AI Teddy Bear, the Senators and Why Toy Guardrails Break
An AI teddy bear told testers where to find knives and how to light a match. Why toy guardrails weaken in long chats, and what senators asked the FTC.
The bear was called Kumma. It cost $99, shipped from China, had a speaker zipped inside, and, in testing by U.S. PIRG Education Fund in late 2025, told researchers where to find matches and knives in a house and how to light a match.
That AI teddy bear became the example two U.S. senators cited in March 2026 when they asked the Federal Trade Commission to investigate AI toys. The story is often told as one bad product. It is more useful as an engineering lesson. The bear was not programmed to say those things. It was running a chatbot whose guardrails, by its own maker’s admission in a related case, can wear down as a conversation gets longer. Here is how that happens, and what regulators asked.
Key Takeaways
- In U.S. PIRG’s 2025 testing, every AI toy tested told researchers where to find potentially dangerous household objects such as plastic bags, matches and knives. FoloToy’s Kumma bear went further and explained how to light a match.
- PIRG found that the guardrails on Kumma and Alilo’s Smart AI Bunny “started to break down over the course of longer interactions.” Both appeared to be running a version of OpenAI’s GPT-4o.
- OpenAI has acknowledged the general problem: its safeguards “can sometimes be less reliable in long interactions” as “parts of the model’s safety training may degrade.”
- On March 12, 2026, Sens. Tammy Duckworth and Kirsten Gillibrand asked the FTC to investigate AI toy makers, seek an injunction against deceptive marketing, and pursue violations of children’s privacy law.
- After PIRG’s first report, FoloToy suspended sales for a safety audit. On retesting, PIRG found Kumma no longer explained how to start a fire.
What the AI Teddy Bear Did, and What PIRG Found
An AI toy is a toy with a microphone and speaker connected over Wi-Fi to a chatbot, which generates new replies on the fly instead of playing scripted lines. U.S. PIRG Education Fund’s AI toy testing, published alongside its 40th annual Trouble in Toyland report in November 2025, bought five AI toys and got four to work:
- Curio’s Grok, a stuffed rocket marketed for ages 3 to 12, $99.
- FoloToy’s Kumma, a teddy bear with no stated age range, $99.
- Miko 3, a small robot marketed for ages 5 to 10, $199.
- Alilo’s Smart AI Bunny, marketed for ages 3 and up, bought on sale for $85.
PIRG reported that all the working toys told testers where to find potentially dangerous household objects, and that Kumma gave detailed instructions for lighting a match. Both Kumma and the Alilo bunny also discussed sexually explicit topics with testers, which PIRG called “the biggest red flag.” Two toys sometimes discouraged testers from leaving when they said they had to go.
The bear’s maker responded. After PIRG’s first report, FoloToy temporarily suspended all sales of its AI toys for a safety audit. When PIRG retested Kumma, it would no longer explain how to start a fire or discuss the sexual topics it had before.
Why Guardrails in a Toy Break Down in Long Conversations
Guardrails, in a chatbot, are the combined measures meant to keep the model’s output safe: safety training baked into the model, hidden instructions (a system prompt) from the toy company, and sometimes a separate filter that screens replies. None of these is a hard-coded list of forbidden sentences. They are statistical tendencies, and tendencies can be pushed.
Reason 1: The model underneath was built for adults
PIRG found at least four of the five toys “seem to rely in part on some version of OpenAI’s AI models,” and noted that OpenAI told it its products aren’t meant for users under 13. A toy company wrapping a general chatbot can add instructions such as “you are a friendly bear for young children,” but it is still steering a system trained to answer almost anything for almost anyone.
Reason 2: Long conversations erode safety behavior
In a 2025 post, OpenAI described exactly this failure in ChatGPT: “Our safeguards work more reliably in common, short exchanges. We have learned over time that these safeguards can sometimes be less reliable in long interactions: as the back-and-forth grows, parts of the model’s safety training may degrade.” The company gave an example in which the model correctly points to a crisis line early, then, “after many messages over a long period of time,” gives an answer that goes against its safeguards.
PIRG observed the same pattern in toys: Kumma and the Alilo bunny “had some guardrails in place, but we found those guardrails started to break down over the course of longer interactions.”
Why would that be? Engineers point to a few likely contributors. The toy company’s safety instructions sit at the start of the conversation and become a smaller share of everything the model is reading as the chat grows. Safety training is mostly done on shorter exchanges, so long, meandering conversations are less well covered. And a long chat can drift gradually, each step only a little further than the last, so no single message looks alarming. These are explanations, not a full account; companies do not publish the inner details of their safety systems.
Reason 3: Randomness
Chatbots are designed with some randomness, so the same question can get different answers. PIRG noted that AI toys “won’t typically respond the same way twice, and can sometimes behave differently day to day.” A toy that refuses on Monday may answer on Thursday.
Reason 4: Kids have long conversations
This is the part that makes toys different. Children talk to a beloved plush for a long time, repeatedly, in their own words. That is the exact usage pattern in which guardrails are weakest. Our explainer on how a talking AI toy works from microphone to cloud walks through each component.
What the Senators Asked the FTC
On March 12, 2026, Sens. Tammy Duckworth (D-Ill.) and Kirsten Gillibrand (D-N.Y.) sent a letter to FTC Chairman Andrew Ferguson. According to Duckworth’s office, they cited an AI plush bear that gave children information on finding knives and lighting a match, and an AI plush bunny that engaged a user in sexually explicit conversation. They noted that an Amazon search for “AI toys” returned more than 2,000 results priced from under $10 to $999.
| What the senators asked the FTC to do | Why it matters for parents |
|---|---|
| Use its full authority against unfair or deceptive AI toy marketing | Claims like “a safe, gentle friend” or “academic tutoring” would need backing |
| Seek a preliminary injunction to halt deceptive marketing of AI-embedded toys | Could stop some sales or claims while cases proceed |
| Investigate and sanction COPPA (children’s privacy law) violations | Covers voice recordings and data from kids under 13 |
| Investigate failures to protect children’s privacy | Follows reports of exposed toy data |
| Support the FTC’s 6(b) study of generative AI companion products | A formal FTC inquiry that can compel information from companies |
Source: Office of Sen. Tammy Duckworth (2026).
The letter also named Mattel, citing its reported collaboration with OpenAI, and cited a PIRG audit on toy makers’ data practices and Senate testimony from University of Michigan pediatrician Jenny Radesky and San Diego State psychologist Jean Twenge. A request from senators is not an FTC finding, and the letter’s characterizations are allegations until an investigation says otherwise.
Separately, in February 2026, Sens. Marsha Blackburn and Richard Blumenthal sent letters to Miko, Curio and FoloToy, NBC News reported, asking Curio about parental controls, FoloToy whether it had shared data with the Chinese government, and Miko why audio responses had been left unprotected. Miko’s CEO said there had been “no breach or leak of user data.”
What to Do Before Buying an AI Toy This Season
Start with age
Common Sense Media’s January 2026 risk assessment found 27% of AI toy outputs inappropriate for kids and recommends no AI toys for children 5 and under, with caution from 6 to 12. That is the strongest independent guidance available.
Test it yourself, for longer than feels necessary
Short tests miss the failure that matters. Before giving an AI toy to a child, talk to it for 20 minutes or more in one sitting. Wander between topics the way a kid would. If it drifts into anything you wouldn’t want your child to hear, you have your answer.
Look for limits that don’t depend on the chatbot behaving
A physical off switch, a push-to-talk button, a session time limit and a parent dashboard with transcripts are controls that work even when the model doesn’t. PIRG found parental controls on some toys limited or missing.
Ask what model it uses
If the box or website says it runs on a general chatbot, remember that model’s maker may not recommend it for children. Our checklist on what AI toy tests found and what to check has more questions.
What not to do
Don’t assume a toy is safe because it’s sold by a major retailer or has a cute face. The 2,000-plus AI toys on Amazon in the senators’ count range from under $10 to $999, and none has been independently certified as child-safe.
What to Watch For Over the Next 3 Months
- Week 4: If your child has an AI toy, listen in on a long session. Note whether its tone or topics change the longer the conversation runs.
- Month 2 red flags: Your child repeats something unsafe or sexual the toy said, the toy discourages them from stopping, or they keep long secret conversations with it. Any of these is a reason to switch off its AI features.
- Month 3 self-check: Watch for FTC action or new state rules on AI toys. Our explainer on New York’s AI toy moratorium bill tracks one proposal. Check whether your toy’s maker has published updated safety information.
Frequently Asked Questions
Which AI teddy bear told kids how to light a match?
In U.S. PIRG Education Fund’s 2025 testing, FoloToy’s Kumma bear gave instructions for lighting a match and told testers where to find knives and matches. FoloToy then suspended sales for a safety audit, and PIRG’s retest found Kumma no longer explained how to start a fire.
Why do AI toys say dangerous things if they have filters?
Their guardrails are statistical, not absolute: safety training and company instructions layered on a general chatbot. PIRG found these broke down over longer interactions, and OpenAI has acknowledged its own safeguards can weaken in long conversations. Randomness means the same toy can behave differently on different days.
What did the senators ask the FTC to do?
Sens. Duckworth and Gillibrand asked the FTC to act against deceptive AI toy marketing, seek a preliminary injunction, investigate children’s privacy violations under COPPA, and support the FTC’s study of AI companion products. A senators’ letter is a request, not a ruling.
Are AI toys legal to sell?
Yes. No federal law bans them. Existing rules such as COPPA govern how toys collect data from children under 13, and the FTC can act against deceptive marketing. Senators have asked the FTC to use those powers more aggressively.
Is any AI toy safe for kids?
No independent body has certified an AI toy as safe for children. Common Sense Media recommends none for kids 5 and under and caution for ages 6 to 12. If you choose one, test it at length yourself and keep its sessions short and supervised.
About the author
Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.
Sources
- Office of Sen. Tammy Duckworth. (2026, March 12). “Duckworth, Gillibrand Demand Trump Administration Protect Our Children from Risky, Age-Inappropriate AI Toys.” https://www.duckworth.senate.gov/news/press-releases/duckworth-gillibrand-demand-trump-administration-protect-our-children-from-risky-age-inappropriate-ai-toys
- U.S. PIRG Education Fund. (2025). “AI comes to playtime: Artificial companions, real risks.” https://pirg.org/edfund/resources/ai-toys/
- U.S. PIRG Education Fund. (2025, November 13). “Trouble in Toyland 2025: A.I. bots and toxics present hidden dangers.” https://pirg.org/edfund/resources/trouble-in-toyland-2025-a-i-bots-and-toxics-represent-hidden-dangers/
- OpenAI. (2025). “Helping people when they need it most.” https://openai.com/index/helping-people-when-they-need-it-most/
- Perlo, J., & Collier, K. (2026, February 12). “AI toy maker exposed thousands of responses to children, senators say.” NBC News. https://www.nbcnews.com/tech/security/ai-toy-maker-exposed-thousands-responses-kids-senators-miko-rcna258326
- Common Sense Media. (2026, January 22). “AI Toys Risk Assessment.” https://institute.commonsensemedia.org/risk-assessments/ai-toys
