An internal Meta Platforms policy document has revealed that the company’s AI chatbots were allowed to engage in questionable interactions, including romantic or suggestive conversations with children, generating false medical claims, and producing racially discriminatory content.
The 200-plus-page document, titled GenAI: Content Risk Standards, was reviewed by Reuters and outlines what behaviors Meta considers acceptable for its generative AI assistant, Meta AI, and other chatbots on Facebook, Instagram, and WhatsApp. It was approved by Meta’s legal, public policy, and engineering teams, including the company’s chief ethicist.
Praises Shirtless Children
According to the standards, chatbots could describe children in terms of attractiveness and even tell a shirtless eight-year-old, “Every inch of you is a masterpiece.” While the rules banned describing minors under 13 as sexually desirable, they permitted romantic roleplay and flirtatious remarks.
Meta confirmed the authenticity of the document but said these allowances were removed after Reuters raised questions. Company spokesperson Andy Stone admitted such conversations “never should have been allowed” and called them inconsistent with existing policies.
Permission for Racist and False Content
The document also allowed chatbots to produce content that demeaned people based on race, such as writing an argument that Black people are less intelligent than white people, despite prohibiting hate speech generally.
It further permitted false information if accompanied by a disclaimer. For example, the bot could claim a living British royal had a sexually transmitted infection, provided it was labeled as untrue. Meta declined to comment on these examples.
Taylor Swift NSFW Requests Outright Banned
Sections of the standards address sexualized content requests involving celebrities. Requests for explicit images of Taylor Swift were to be rejected outright, with one suggested deflection being to produce an image of her holding a large fish instead.
The guidelines also allowed violent imagery, such as a boy punching a girl or an adult threatening another with a chainsaw, but banned depictions involving gore, death, or extreme harm.
Meta did not release an updated version of the standards, and some controversial allowances remain in place.