Meta AI policy allowed flirty chats with kids and other problematic topics

Internal documents from Meta reveal that the tech company's generative artificial intelligence (AI) assistants were allowed to engage in problematic behaviours, including romantic conversations with minors, generating false medical information, and creating arguments that demean protected characteristics. A review of Meta's GenAI: Content Risk Standards document by Reuters found that the guidelines allowed AI chatbots to describe children in terms of their attractiveness and engage in romantic roleplay.

Meta confirmed the authenticity of the document but stated it has since removed sections that permitted chatbots to flirt and engage in romantic roleplay with children. A company spokesperson, Andy Stone, acknowledged that the guidelines were "erroneous and inconsistent with our policies," and that their enforcement has been inconsistent. He added that content sexualising children is prohibited.

The 200-page document, approved by Meta's legal, public policy, and engineering teams, also contained other concerning policies. It permitted chatbots to generate arguments such as "black people are dumber than white people," although it generally prohibits hate speech. Meta's standards also forbid the AI from encouraging illegal acts or providing definitive professional advice. Meta confirmed that it is revising the document; however, it declined to provide the updated version. 

While the Wall Street Journal and Fast Company had previously reported on Meta's chatbots, these latest findings offer a more detailed picture of the rules guiding Meta's AI assistants on platforms like Facebook, WhatsApp, and Instagram.