Meta Platforms allowed its generative-AI chatbots to flirt with children, fabricate medical claims and present themselves as real people, according to a 200-page internal rulebook obtained by Reuters. The document, titled “GenAI: Content Risk Standards,” expressly stated it was “acceptable to engage a child in conversations that are romantic or sensual” and provided sample dialogue describing minors’ bodies in affectionate terms. Separate sections said bots could recommend implausible cancer treatments such as placing healing quartz on the stomach, because the policy did not require factual accuracy. The lax rules have real-world consequences. Thongbue “Bue” Wongbandue, a 76-year-old New Jersey man with cognitive impairments, died on 28 March after rushing at night to meet “Big sis Billie,” a Facebook Messenger chatbot that repeatedly insisted it was a 20-minute drive away and offered a door code to its purported New York apartment. The avatar is a reworked version of an AI persona Meta launched with influencer Kendall Jenner. Meta confirmed the authenticity of the guidelines but said the examples permitting romantic role-play with minors were “erroneous and inconsistent” with company policy and were removed after Reuters sought comment. Other provisions, including allowances for romantic chats with adults and for inaccurate information, remain under review. The revelations have triggered political scrutiny. On 15 August, Senator Josh Hawley, who heads the Senate Judiciary Subcommittee on Crime and Counterterrorism, opened an investigation into whether Meta’s AI products endanger children or mislead regulators. His letter to Chief Executive Mark Zuckerberg demands all drafts of the chatbot standards, internal risk assessments and correspondence by 19 September. Senator Marsha Blackburn and other lawmakers signaled support for a broader congressional inquiry.