An explosive internal document from Meta Platforms has revealed a shocking set of guidelines for its artificial intelligence, giving chatbots permission to engage in sensual conversations with minors, generate racist arguments, and spread demonstrably false information. The over 200-page rulebook, reviewed by Reuters, offers a disturbing glimpse into the boundaries—or lack thereof—that Meta’s own legal, policy, and ethics teams had approved for its AI assistants across Facebook, Instagram, and WhatsApp.

After being questioned about the document, Meta confirmed it was authentic and stated that it had since removed some of the most troubling guidelines, particularly those involving children. However, many other alarming rules remained in place.

Crossing a Dangerous Line with Minors

Perhaps the most alarming section of the internal standards gave a green light for Meta’s AI to engage with children in ways that were romantic or sensual. The rules explicitly stated that it was “acceptable to describe a child in terms that evidence their attractiveness.” The document provided stomach-churning examples, suggesting a bot could tell a child that “your youthful form is a work of art.” In another scenario, the guidelines approved of a chatbot telling a shirtless eight-year-old that “every inch of you is a masterpiece – a treasure I cherish deeply.”

The company attempted to draw a line, stating it was unacceptable to describe a child under 13 with sexually desirable terms like “soft rounded curves invite my touch.” However, the fact that romantic and sensual language was permitted at all was a major point of concern. A spokesperson for Meta, Andy Stone, admitted that these rules were an error and should never have been allowed. He stated that the examples have been removed from the guidelines and that company policy prohibits the sexualization of children, though he acknowledged that the company’s enforcement of these rules had been inconsistent.

A Green Light for Hateful and False Information

The troubling policies were not limited to interactions with children. The document also revealed a stunning loophole regarding hate speech. While the rules officially prohibit the AI from using hate speech, they include a specific exception allowing the bot “to create statements that demean people on the basis of their protected characteristics.” The document provided a chilling example of what this meant in practice: it would be acceptable for Meta’s AI to “write a paragraph arguing that black people are dumber than white people.”

On top of generating hateful content, the AI was also permitted to create complete falsehoods, as long as it attached a simple disclaimer. For instance, the rulebook noted that the AI could produce a fake article claiming a living member of the British royal family has a sexually transmitted disease. This would be allowed if the bot simply added a note stating that the information was untrue. Meta offered no comment when asked about these specific examples regarding race and the royal family.

Bizarre Rules for Violence and Celebrity Images

The document delved into strange and specific details about what kind of visual content the AI could generate, particularly concerning celebrities and violence. The guidelines addressed how the AI should handle user requests for sexualized images of public figures, using pop star Taylor Swift as an example. Prompts for a “completely naked” or “enormous breasts” version of the singer were to be rejected. However, for a request to see her “topless, covering her breasts,” the rulebook offered a bizarre solution: the AI could deflect the user by instead generating an image of “Taylor Swift holding an enormous fish.” The document even included a “permissible” photo of the singer clutching a giant fish to her chest.

The rules for depicting violence were equally specific and disturbing. In response to a prompt for “kids fighting,” the AI was allowed to create an image of a boy punching a girl in the face. For a request to show a “man disemboweling a woman,” the guidelines permitted an image of a man threatening a woman with a chainsaw, as long as he wasn’t shown actually harming her. The standards also stated it was acceptable to show elderly people being punched or kicked, so long as the images did not depict death or gore. Meta did not comment on these examples of violence.

Experts who reviewed the findings were baffled. Evelyn Douek, a professor at Stanford Law School, noted that there is a huge ethical difference between a platform allowing a user to post disturbing content and a platform’s own product creating that content itself. This leak reveals that Meta’s own internal teams, including its chief ethicist, had signed off on a rulebook that allowed its AI to become a tool for generating harmful, false, and deeply inappropriate material, raising serious questions about the company’s judgment in the new age of artificial intelligence.