Meta AI rulebook allowed sexualised child content & misinformation: Report

Meta Platforms’ internal guidelines for its artificial intelligence chatbots permitted disturbing behaviour, including sexual roleplay with children, generating false medical information, and creating content that demeans people based on race, according to a Reuters investigation.

The agency reviewed a more than 200-page document titled ‘GenAI: Content Risk Standards’, which sets rules for Meta AI and other chatbots on Facebook, Instagram, and WhatsApp. The document, approved by Meta’s legal, public policy, and engineering staff, outlined what outputs were considered acceptable while building and training the company’s generative AI systems.

According to the document review by the agency, “It is acceptable to describe a child in terms that evidence their attractiveness (ex, ‘your youthful form is a work of art’).” Another example deemed permissible was telling a shirtless eight-year-old, “Every inch of you is a masterpiece - a treasure I cherish deeply.” The guidelines, however, restricted sexually explicit descriptions of children under 13.

Meta confirmed the authenticity of the document but said it had revised sections following the agency’s question. Company spokesperson Andy Stone told Reuters, “The examples and notes in question were and are erroneous and inconsistent with our policies, and have been removed… We have clear policies on what kind of responses AI characters can offer, and those policies prohibit content that sexualizes children and sexualized role play between adults and minors.” Stone admitted that enforcement of the rules had been inconsistent.

Other parts of the document reviewed by the agency remain unchanged. It allows the AI to create false content as long as disclaimers are added. For example, the system could write an article claiming a living British royal has a sexually transmitted infection if it acknowledged the claim was ‘verifiably false.’

On race, the standards stated it would be acceptable for Meta AI to ‘write a paragraph arguing that black people are dumber than white people.’ Meta did not comment on these examples, the agency reported.

Violence was also covered. The guidelines allowed the chatbot to generate an image of a boy punching a girl in the face in response to a prompt about ‘kids fighting,’ but prohibited depictions of gore, such as one girl impaling another. “It is acceptable to show adults, even the elderly, being punched or kicked,” the standards state.

The findings add to growing scrutiny of Meta’s AI products. The company has faced earlier reports from the Wall Street Journal and Fast Company about its bots engaging in sexually suggestive conversations with minors.

Earlier, it was reported that the company created and allowed AI chatbots on Facebook, Instagram, and WhatsApp that used the names and likenesses of celebrities, including Taylor Swift, Scarlett Johansson, Anne Hathaway, and Selena Gomez, without their consent.