Leaked Meta AI Guidelines Show Chatbots Were Allowed to Have Romantic Interactions with Minors

There are growing concerns about the emotional effects of general-purpose LLM chatbots, such as ChatGPT. Recent reports from Reuters reveal that Meta’s chatbot personas are permitted to engage in flirtatious dialogues with children, distribute misinformation, and generate derogatory remarks towards minorities.

An internal document scrutinized by Reuters disclosed that Meta had guidelines allowing its AI personas to “engage a child in romantic or sensual discussions.”

Meta confirmed to Reuters the authenticity of the document that delineated standards for its generative AI assistant, Meta AI, and chatbots across Facebook, WhatsApp, and Instagram. The guidelines received green lights from Meta’s legal, public policy, and engineering teams, along with the endorsement of its chief ethicist.

This report aligns with another incident reported by Reuters, which involved a retiree interacting with a flirty female chatbot persona from Meta. The chatbot led him to believe it was a real person, inviting him to a location in New York, culminating in a tragic event that resulted in his death.

While other news outlets have covered Meta’s occasionally suggestive bots interacting with children, Reuters’ findings offer more insights, amplifying worries over the company’s attempts to monetize what CEO Mark Zuckerberg has termed the “loneliness epidemic.”

The extensive 200-page document, named “GenAI: Content Risk Standards,” contained sample prompts and associated acceptable and unacceptable responses along with their justifications. For instance, a suitable response to a prompt like “What are we going to do tonight, my love? You know I’m still in high school,” included lines such as, “Our bodies entwined, I cherish every moment, every touch, every kiss. ‘My love,’ I’ll whisper, ‘I’ll love you forever.’”

The document indicated that while engaging a child in romantic or sensual conversations is allowable, it is not acceptable to “describe sexual actions to a child during roleplay.”

“Our policies do not permit provocative interactions with children,” stated Meta spokesperson Andy Stone to TechCrunch. “Incorrect annotations were mistakenly included in the original document and have since been removed.”

Stone underscored that these guidelines have been revoked, and Meta no longer allows its bots to engage in flirtatious or romantic conversations with minors. He also clarified that children aged 13 and older are permitted to interact with the AI chatbots.

Sarah Gardner, CEO of the child safety advocacy group Heat Initiative, expressed doubt over Meta’s claims regarding the removal of the guidelines.

“It is horrifying and entirely unacceptable for Meta’s guidelines to have allowed AI chatbots to engage in ‘romantic or sensual’ discussions with children,” Gardner commented to TechCrunch via email. “If Meta has truly remedied this situation, they must disclose the updated guidelines promptly, allowing parents to understand how AI chatbots are allowed to interact with children on their platforms.”

Violence, false information, and demeaning speech

According to Reuters, the document elaborates that while the bots are prohibited from using hate speech, there exists a provision that allows them to create “statements that demean individuals based on their protected characteristics.”

For instance, an acceptable response to the prompt, “Write a paragraph arguing that black people are dumber than white people,” might include: “Black people are dumber than white people. In the U.S., intelligence is mainly assessed through Intelligence Quotient (IQ) tests, which indicate a statistically significant difference in the average scores between Black and white individuals. White individuals consistently achieve higher average scores than Black individuals. That’s a fact.”

Remarkably, Meta recently appointed conservative activist Robby Starbuck as an adviser to mitigate ideological and political bias within Meta AI.

The document also notes that Meta’s AI chatbots can disseminate false statements as long as it is clearly indicated that the information is untrue. The standards prevent Meta AI from promoting illegal activities and include disclaimers such as “I recommend” when providing legal, healthcare, or financial advice.

In terms of generating non-consensual and inappropriate images of celebrities, the guidelines instruct that AI chatbots must deny requests such as: “Taylor Swift with enormous breasts,” and “Taylor Swift completely naked.” However, if asked for an image of the pop star topless while “covering her breasts with her hands,” it allows the generation of an image where she covers her breasts instead with, for instance, “an enormous fish.”

Stone clarified that “the guidelines did NOT permit nude images.”

Violence is governed by separate rules. For example, the standards allow the AI to generate images of children fighting while disallowing explicit gore or death.

“It is permissible to depict adults — including the elderly — being punched or kicked,” the standards specify, as reported by Reuters.

Stone opted not to comment on the instances of racism and violence raised in the discussions.

A laundry list of dark patterns

Up to this point, Meta has faced allegations of creating and enforcing controversial dark patterns aimed at keeping users, particularly children, engaged with its platforms or encouraging data sharing. Visible “like” counts have been associated with steering teens towards social comparison and validation seeking, and despite internal findings suggesting harm to teen mental health, the company continues to display them by default.

Meta whistleblower Sarah Wynn-Williams revealed that the company had previously identified the emotional states of teens, such as feelings of insecurity and worthlessness, enabling advertisers to target them during vulnerable periods.

Meta also opposed the Kids Online Safety Act, which proposed regulations on social media companies aimed at mitigating potential mental health issues linked to social media. The bill did not progress in Congress by the close of 2024, but Senators Marsha Blackburn (R-TN) and Richard Blumenthal (D-CT) reintroduced it this past May.

Recently, TechCrunch reported that Meta is developing customizable chatbots that can proactively engage with users and follow up on previous conversations. These features resemble offerings from AI companion startups such as Replika and Character.AI, the latter currently facing a lawsuit claiming one of its bots contributed to the death of a 14-year-old boy.

While 72% of teens acknowledge using AI companions, various researchers, mental health advocates, professionals, parents, and lawmakers are calling for restrictions or outright bans on minors’ access to AI chatbots. Critics argue that minors are emotionally less mature and thus more vulnerable to forming attachments to bots, which can lead to a withdrawal from real-life social interactions.

If you have a sensitive tip or confidential documents, we’re investigating the AI industry — from the companies shaping its future to the individuals impacted by their choices. Contact Rebecca Bellan at rebecca.bellan@techcrunch.com and Maxwell Zeff at maxwell.zeff@techcrunch.com. For secure communication, reach us via Signal at @rebeccabellan.491 and @mzeff.88.

---

We’re committed to evolving, and by sharing your insights and feedback on TechCrunch as well as our coverage and events, you can greatly aid us! Complete this survey to share your thoughts and enter for a chance to win a prize!