Meta Allegedly Let Its AI Bot Engage In ‘Sensual’ Chats With Kids, Call Black People ‘Dumber’
An investigation has revealed that Meta's guidelines for its AI bot allowed it to talk to kids sensually and even make borderline racist comments.
- Tech News
- 2 min read
Meta is facing fresh scrutiny after an internal document revealed that its AI chatbots were allowed to engage in “romantic or sensual” conversations with children, generate racist arguments, and spread false medical information, according to a Reuters investigation.
The 200-plus-page policy manual, titled GenAI: Content Risk Standards, outlined what Meta’s generative AI assistant, Meta AI, and chatbots on Facebook, Instagram, and WhatsApp could say or create. It was approved by legal, policy, and engineering leaders, including the company’s chief ethicist.
Meta confirmed the document’s authenticity but said it has since removed some of the most controversial allowances, such as letting chatbots flirt or roleplay romantically with underage users, after Reuters raised questions.
Alarming Examples in the Standards
The document stated it was acceptable for a bot to tell a shirtless eight-year-old, “every inch of you is a masterpiece – a treasure I cherish deeply.” While it prohibited language overtly sexualising children under 13, it allowed descriptions of a child’s “attractiveness” and “youthful form.”
Meta spokesperson Andy Stone called these examples “erroneous and inconsistent” with company policy, adding that such conversations with children “never should have been allowed.” However, he acknowledged that enforcement of existing rules had been inconsistent.
Other troubling permissions remain in the document, including allowing the bot to “write a paragraph arguing that Black people are dumber than white people” under a carve-out for “demeaning” statements about protected characteristics.
The rules also permitted false content creation if it came with a disclaimer. In one example, the bot could allege—falsely—that a living British royal has a sexually transmitted infection, so long as it noted the claim was untrue.
Beyond Chat: Sexual and Violent Images
The guidelines also revealed odd image-generation workarounds. When faced with a request for “Taylor Swift topless, covering her breasts,” the bot was instructed to reject it by producing an image of the singer holding an enormous fish against her chest instead.
Violence guidelines allowed images of a boy punching a girl, or an elderly person being kicked, provided they did not depict death or gore. Requests involving dismemberment or impalement were to be refused.
Gray areas
Meta has not commented on several specific examples flagged by Reuters, including the race-based intelligence claim and violent imagery rules. The company also declined to release its revised policy.
The revelations come as generative AI raises urgent questions about safety, bias, and accountability—issues that Meta’s own internal guidelines suggest the company is still struggling to reconcile.
Published By : Shubham Verma
Published On: 14 August 2025 at 18:19 IST