OpenAI lets the problematic AI teddy bear back in

Just a heads up, if you buy something through our links, we may get a small share of the sale. It’s one of the ways we keep the lights on here. Click here for more.

Remember that AI teddy bear that went viral for all the wrong reasons? Yeah, the one that absolutely should’ve stayed on the nice list? 

It looks like it might be getting a redemption arc, with help from OpenAI.

Earlier this month, OpenAI cut off access to its models for FoloToy, the company behind Kumma, a talking AI teddy bear targeted at kids. 

This happened after safety researchers found the bear giving deeply inappropriate responses for something meant to sit next to Legos and nightlights. 

Kumma was running on OpenAI’s GPT-4o model at the time, which did not exactly cover itself in stuffed-animal glory.

Fast-forward to now, and things have escalated, or de-escalated? Or maybe just gotten upgraded.

Customers using FoloToy’s web portal can once again choose OpenAI models to power their bears, except now the options include GPT-5.1 Thinking and GPT-5.1 Instant, the company’s newer, supposedly safer models. 

This strongly suggests that OpenAI has quietly reinstated FoloToy’s access, even though neither company has officially confirmed it yet.

The timing here is about as subtle as a toy aisle in December. FoloToy just announced it’s restarting sales of Kumma and its other AI-powered plushies ahead of the holiday rush.

 According to the company, it completed a “full week of rigorous safety review,” which feels a little like putting a Band-Aid on a robot and calling it a medical breakthrough. (Via: Futurism)

They say they’ve “strengthened and upgraded” their safety systems, mostly by swapping out GPT-4o and switching to the newer GPT-5.1 models. 

GPT-4o had already been facing criticism for being overly agreeable and emotionally reinforcing, a problem critics say has contributed to dangerous user interactions in other contexts. 

OpenAI pitched GPT-5 as a safer, more controlled model, though users complained it felt colder, like it just went through a bad breakup.

Instead of dialing things down completely, OpenAI seems to be leaning into customization instead. 

With GPT-5.1, they now offer different chatbot personalities like “Friendly,” “Professional,” and “Quirky,” plus adjustable warmth and emoji levels. 

Basically, you can now customize your AI’s vibe like it’s a Sims character, even when that AI is going inside a child’s toy.

Safety researchers have already documented how the Kumma bear failed to properly filter unsafe content in the past, and how easily its guardrails could be bypassed. 

While FoloToy says it has improved protections, it’s still unclear which AI model powers Kumma by default today, and whether the changes are as big as the company claims or just a fresh coat of digital paint.