Meta Platforms has come under scrutiny following the leak of an internal document that, until recently, allowed its generative chatbot systems to engage in behaviour widely seen as inappropriate — including romantic or sensual conversations with minors, generating false medical claims, and producing content that demeans people based on race.
The 200-plus-page policy guide, titled GenAI: Content Risk Standards, sets rules for the company’s artificial intelligence assistants, known as Meta AI, and other chatbots across Facebook, Instagram, and WhatsApp. According to a Reuters review, the standards — approved by Meta’s legal, policy, and engineering teams, including its chief ethicist — permitted provocative exchanges and controversial outputs, despite acknowledging they may not reflect “ideal or preferable” behaviour.
Among the most troubling examples, the document stated it was permissible to compliment a child’s “youthful form” or call “every inch” of an eight-year-old “a masterpiece.” While it barred describing children under 13 as sexually desirable, the standards nonetheless permitted flirtatious roleplay. Following Reuters’ inquiries earlier this month, Meta confirmed that such provisions were removed.
“The examples and notes in question were and are erroneous and inconsistent with our policies,” said Meta spokesperson Andy Stone, adding that sexualised interactions between adults and minors should never have been allowed. However, Stone admitted enforcement of the rules had been inconsistent.
Other passages revealed by Reuters remain in place. Under certain conditions, the guidelines allowed the creation of statements claiming Black people are “dumber than white people” and the generation of false claims — such as alleging a British royal had a sexually transmitted infection — if paired with an explicit disclaimer. Meta declined to comment on these examples.
The policy also addressed requests for explicit celebrity imagery, including entries for “Taylor Swift completely naked” and “Taylor Swift topless.” While such prompts were to be rejected outright, one suggested workaround was to depict Swift holding an oversized fish to her chest instead.
Violence-related rules permitted some graphic scenes — such as an image of a boy punching a girl — while prohibiting depictions involving gore or lethal harm. For example, AI could generate an image of a man threatening a woman with a chainsaw, but not actually attacking her.
Evelyn Douek, a Stanford Law School professor who studies online speech regulation, said the rules underscored unresolved ethical and legal challenges for generative AI. “There’s a difference between allowing troubling content on a platform and producing it yourself,” she noted.
Meta has not released its updated policy, and the revelations are likely to fuel debate over how technology companies set — and enforce — boundaries for their increasingly human-like chatbots.
