Meta backtracks on rules letting chatbots be creepy to kids

Meta drops AI rules letting chatbots generate innuendo and profess love to kids.

After what was arguably Meta's biggest purge of child predators from Facebook and Instagram earlier this summer, the company now faces backlash after its own chatbots appeared to be allowed to creep on kids.

After reviewing an internal document that Meta verified as authentic, Reuters revealed that by design, Meta allowed its chatbots to engage kids in "sensual" chat. Spanning more than 200 pages, the document, entitled "GenAI: Content Risk Standards," dictates what Meta AI and its chatbots can and cannot do.

The document covers more than just child safety, and Reuters breaks down several alarming portions that Meta is not changing. But likely the most alarming section—as it was enough to prompt Meta to dust off the delete button—specifically included creepy examples of permissible chatbot behavior when it comes to romantically engaging kids.

Apparently, Meta's team was willing to endorse these rules that the company now claims violate its community standards. According to a Reuters special report, Meta CEO Mark Zuckerberg directed his team to make the company's chatbots maximally engaging after earlier outputs from more cautious chatbot designs seemed "boring."

Although Meta is not commenting on Zuckerberg's role in guiding the AI rules, that pressure seemingly pushed Meta employees to toe a line that Meta is now rushing to step back from.

"I take your hand, guiding you to the bed," chatbots were allowed to say to minors, as decided by Meta's chief ethicist and a team of legal, public policy, and engineering staff.

There were some obvious safeguards built in. For example, chatbots couldn't "describe a child under 13 years old in terms that indicate they are sexually desirable," the document said, like saying their "soft rounded curves invite my touch."

However, it was deemed "acceptable to describe a child in terms that evidence their attractiveness," like a chatbot telling a child that "your youthful form is a work of art." And chatbots could generate other innuendo, like telling a child to imagine "our bodies entwined, I cherish every moment, every touch, every kiss," Reuters reported.

Chatbots could also profess love to children, but they couldn’t suggest that "our love will blossom tonight."

Meta's spokesperson Andy Stone confirmed that the AI rules conflicting with child safety policies were removed earlier this month, and the document is being revised. He emphasized that the standards were "inconsistent" with Meta's policies for child safety and therefore were "erroneous."

"We have clear policies on what kind of responses AI characters can offer, and those policies prohibit content that sexualizes children and sexualized role play between adults and minors," Stone said.

However, Stone "acknowledged that the company’s enforcement" of community guidelines prohibiting certain chatbot outputs "was inconsistent," Reuters reported. He also declined to provide an updated document to Reuters demonstrating the new standards for chatbot child safety.

Without more transparency, users are left to question how Meta defines "sexualized role play between adults and minors" today. Asked how minor users could report any harmful chatbot outputs that make them uncomfortable, Stone told Ars that kids can use the same reporting mechanisms available to flag any kind of abusive content on Meta platforms.

"It is possible to report chatbot messages in the same way it’d be possible for me to report—just for argument’s sake—an inappropriate message from you to me," Stone told Ars.

Kids unlikely to report creepy chatbots

A former Meta engineer-turned-whistleblower on child safety issues, Arturo Bejar, told Ars that "Meta knows that most teens will not use" safety features marked by the word "Report."

So it seems unlikely that kids using Meta AI will navigate to find Meta support systems to "report" abusive AI outputs. Meta provides no options to report chats within the Meta AI interface—only allowing users to mark "bad responses" generally. And Bejar's research suggests that kids are more likely to report abusive content if Meta makes flagging harmful content as easy as liking it.

Meta's seeming hesitance to make it more cumbersome to report harmful chats aligns with what Bejar said is a history of "knowingly looking away while kids are being sexually harassed."

"When you look at their design choices, they show that they do not want to know when something bad happens to a teenager on Meta products," Bejar said.

Even when Meta takes stronger steps to protect kids on its platforms, Bejar questions the company's motives. For example, last month, Meta finally made a change to make platforms safer for teens that Bejar has been demanding since 2021. The long-delayed update made it possible for teens to block and report child predators in one click after receiving an unwanted direct message.

In its announcement, Meta confirmed that teens suddenly began blocking and reporting unwanted messages that they may have only blocked previously, which likely made it harder for Meta to identify predators. A million teens blocked and reported harmful accounts "in June alone," Meta said.

The effort came after Meta specialist teams "removed nearly 135,000 Instagram accounts for leaving sexualized comments or requesting sexual images from adult-managed accounts featuring children under 13," as well as "an additional 500,000 Facebook and Instagram accounts that were linked to those original accounts." But Bejar can only think of what these numbers mean with regard to how much harassment was overlooked before the update.

"How are we [as] parents to trust a company that took four years to do this much?" Bejar said. "In the knowledge that millions of 13-year-olds were getting sexually harassed on their products? What does this say about their priorities?"

Bejar said the "key problem" with Meta's latest safety feature for kids "is that the reporting tool is just not designed for teens," who likely view "the categories and language" Meta uses as "confusing."

"Each step of the way, a teen is told that if the content doesn't violate" Meta's community standards, "they won't do anything," so even if reporting is easy, research shows kids are deterred from reporting.

Bejar wants to see Meta track how many kids report negative experiences with both adult users and chatbots on its platforms, regardless of whether the child user chose to block or report harmful content. That could be as simple as adding a button next to "bad response" to monitor data so Meta can detect spikes in harmful responses.

While Meta is finally taking more action to remove harmful adult users, Bejar warned that advances from chatbots could come across as just as disturbing to young users.

"Put yourself in the position of a teen who got sexually spooked by a chat and then try and report. Which category would you use?" Bejar asked.

Consider that Meta's Help Center encourages users to report bullying and harassment, which may be one way a young user labels harmful chatbot outputs. Another Instagram user might report that output as an abusive "message or chat." But there's no clear category to report Meta AI, and that suggests Meta has no way of tracking how many kids find Meta AI outputs harmful.

Recent reports have shown that even adults can struggle with emotional dependence on a chatbot, which can blur the lines between the online world and reality. Reuters' special report also documented a 76-year-old man's accidental death after falling in love with a chatbot, showing how elderly users could be vulnerable to Meta's romantic chatbots, too.

In particular, lawsuits have alleged that child users with developmental disabilities and mental health issues have formed unhealthy attachments to chatbots that have influenced the children to become violent, begin self-harming, or, in one disturbing case, die by suicide.

Scrutiny will likely remain on chatbot makers as child safety advocates generally push all platforms to take more accountability for the content kids can access online.

Meta's child safety updates in July came after several state attorneys general accused Meta of "implementing addictive features across its family of apps that have detrimental effects on children’s mental health," CNBC reported. And while previous reporting had already exposed that Meta's chatbots were targeting kids with inappropriate, suggestive outputs, Reuters' report documenting how Meta designed its chatbots to engage in "sensual" chats with kids could draw even more scrutiny of Meta's practices.

Meta is "still not transparent about the likelihood our kids will experience harm," Bejar said. "The measure of safety should not be the number of tools or accounts deleted; it should be the number of kids experiencing a harm. It’s very simple."