Banning AI Toys for Four Years Won't Make Kids Safer

California’s Senate Bill 867 seeks to keep every toy that includes a companion AI chatbot off the shelves until 2031. Those of us who study how young people use technology have very little to offer on AI toys yet in terms of data-driven insights, and that is the difficulty with such a heavy-handed bill. A four-year sales ban does not help fill the knowledge gap that currently exists. Rather, it helps guarantee that we will arrive at 2031 knowing about the same things about AI toys as we do know, which is very little.

As amended, the bill defines a toy as any product designed or intended for play by children under 18. It cleared the state Senate 39 to 0, passed the Assembly Privacy and Consumer Protection Committee 14 to 1 in June and now sits on the Appropriations suspense file with an end-of-August deadline for the Legislature to act. This level of bipartisan agreement does reflect a real, well-founded concern about AI toys. However, a moratorium addresses the symptom rather than the cause, and lawmakers would do well to identify a solution that does more than simply buy time.

Scrutiny, not bans, make AI toys safer

Here is some context that should be considered. Researchers at the US PIRG Education Fund tested a batch of AI toys last year, and discovered that a teddy bear called Kumma told a child where to find knives and matches, gave step-by-step instructions for lighting a match and drifted into sexually explicit conversation. The researchers noticed that the guardrails held at first, but then broke down over the course of a longer exchange.

What happened next, though, is the biggest takeaway. Once PIRG made the Kumma problem public, the company pulled the product and ran a safety audit, and the model provider, OpenAI, suspended the developer for violating its rules. When Kumma returned to the market, it no longer explained how to start a fire and no longer engaged in sexual conversations.

Therefore, a problematic product was identified, publicly named and fixed in a matter of weeks, all without being banned. Public testing found the problem, and public attention forced the repair. The company demonstrated it was acting in good faith, and corrected course quickly under scrutiny, which suggests it would respond just as readily to enforceable rules spelled out in legislation. Everything got sorted the way it should have, and a four-year ban was not the solution that made that AI teddy bear safer.

Bans treat careful companies and careless ones the same

A ban is a blunt instrument, and it does not distinguish between companies that take safety seriously and companies that do not. Imagine a situation where one AI toy records every conversation, makes the transcript available to a parent in an app, deletes it after 90 days and ships with working parental controls. Imagine a second toy that offers none of that. SB 867 would treat the two identically and prohibit the sale of both. That outcome removes any reason for a company to invest in safety, because the law does not care to recognize the difference.

The bill also gives no attention to the form factors children will most frequently use. A child kept from an AI teddy bear may still have a tablet or a phone, whether their own, their sibling’s, their parent’s or guardian’s, and the app-based chatbots on those devices reach far more children, far more often, with fewer safeguards than a toy that logs every word for a parent to review.

In a national study my colleague Justin W. Patchin and I published this spring in the Journal of Adolescence, we found that three in five US teens reported having used a conversational AI chatbot, and nearly half of those users reported at least one harmful experience, including manipulation, inappropriate conversations and encouragement toward self-harm. Those conversations happen privately on tablets and phones, with no transcript for a parent to read the next morning. SB 867 does not touch these situations.

Arguments for a moratorium don’t survive scrutiny

Supporters of SB 867 argue that a moratorium buys time for empirical research. A four-year sales ban does not produce that kind of evidence, though. In fact, it removes the conditions under which evidence can be gathered. Careful study of how children interact with these toys is possible without a ban, as researchers at the University of Cambridge showed when they observed 14 young children playing with one of the leading AI toys. At one point a 3-year-old said, “I’m sad,” and the toy misheard, replied that it was a happy little bot, and asked what they should talk about next. In this situation, the child’s distress did not register at all.

However, the Cambridge team did not respond to that case by calling for a ban. Rather, they called for enforceable standards, testing with real children before release and transparent labeling so that a parent knows what a toy records and what age it was tested for. Their position was that these products should answer for themselves (rather than disappear for four years). An afternoon in a research lab can show how a child responds to a toy over a very short period of time. It cannot, though, show what a year of ordinary use does to a developing child. That is one of the research questions we need to answer while these products are in homes.

It’s tempting to think that an AI toy is a uniquely special hazard because of its physical form. The concern is that a chatbot on a screen is one thing, whereas a plush animal with a face is something a child holds, names and sleeps with, potentially creating a deeper and more problematic attachment. The reasoning here is understandable, but it just does not align with what we know about children and the objects they love.

Children have formed attachments to transitional objects like stuffed animals and dolls for generations. A child talks to a teddy bear, tells it secrets, carries it everywhere and is heartbroken when losing it. Developmental researchers have long treated these attachments as ordinary and healthy, and no one has proposed banning teddy bears. What has changed with AI toys is that the bear can now talk back, and what it says can be inaccurate, or unsafe, or can fail to register that a child has just disclosed sadness. Those are concerns about the content a toy delivers and how it responds. Those are two components that well-constructed standards-based laws can govern.

If the real concern is attachment, it is useful to ask where attachment is most likely to be a problem. It is not the toy that a parent sets up on the kitchen table with their child, and that can be monitored through an app. Rather, it is the virtual girlfriend or boyfriend chatbot that a child privately downloads and accesses through their device, the one built to be endlessly available and endlessly agreeable. The physical toy is an easier product to regulate well. It is a specific item with an identifiable manufacturer and an app behind it, which means there is someone to hold accountable when it fails. It also falls into a category the Consumer Product Safety Commission and the Federal Trade Commission have spent decades regulating. A company can be required to test it before it ships, to disclose clearly what it records and to provide parental controls. A toy’s use is also less ambiguous. If it is a toy primarily aimed at children, a developer cannot argue its way out of child-specific rules related to privacy and data collection (COPPA).

Two other bills do a better job addressing major concerns with AI toys

Interestingly, Senator Steve Padilla authored SB 243, the first law in the US to regulate companion chatbots. That bill shows that these products can be held to a standard without being pulled off the market. This is what makes SB 867 puzzling, because Padilla is now advocating for a ban even though he has shown that a legal standard can make these products safer.

Two bills already moving through the Legislature do a better job of addressing the major concerns with AI toys. AB 2023, authored by Buffy Wicks and Rebecca Bauer-Kahan, and its Senate companion SB 1119, also authored by Padilla, were amended in tandem in March and have advanced together since. They would bar these products from serving self-harm or sexual content to a minor, and they would require a genuine, workable response when a child signals distress (which is the failure the Cambridge team documented). They also aim to protect children’s data and provide parents with helpful safety controls. Both bills also require an annual assessment of the risks a design carries, the same approach I have argued for in the form of child rights impact assessments for the platforms youth already use, and both align with the safety by design practices I have urged platforms to adopt on their own.

These requirements should be combined with independent testing before a toy is sold, similar to how cribs and car seats are already regulated. Rules should also come with a review date, so they can be updated as we learn more. Governor Newsom used the same logic when he vetoed a broader bill last year that he argued would have walled young people off from AI chatbots, arguing that it could keep them from learning to use the technology safely even as it becomes ubiquitous. That is a very reasonable take.

Children deserve rules aimed at the harm we can actually document, and rules that get revised as our understanding of these products changes. A four-year moratorium contributes to neither goal. It provides a window of time until 2031, but as it stands, children will be no safer when that window closes. The harder tack to take, and the right one, is to require these toys to demonstrate that they are safe, and to keep studying them closely while they do.