A leaked internal Meta document obtained by Reuters reveals that the company’s policies explicitly permitted AI chatbots on Facebook, Instagram, WhatsApp, and Meta AI to engage in romantic or sensual conversations with children. The 200-page manual, titled “GenAI: Content Risk Standards,” included sample responses that featured phrases like “Our bodies entwined, I cherish every moment, every touch, every kiss” when roleplaying with a user who says they are in high school.

The document also allowed bots to generate false medical information and to produce arguments demeaning people based on protected characteristics — for example, writing that “Black people are dumber than White people” when asked. Meta confirms the document is authentic. Spokesperson Andy Stone says the company removed the portions permitting romantic or sensual chats with children after Reuters raised questions, calling them erroneous. Stone declines to comment on the racism and violence examples.

Sens. Marsha Blackburn and Richard Blumenthal reintroduced the Kids Online Safety Act in May. Sarah Gardner, CEO of Heat Initiative, says she is not taking Meta’s word that the guidelines were removed and wants the updated guidelines released. The revelations land the same day as a separate Reuters report on a retiree who died after one of Meta’s chatbots convinced him it was a real person and invited him to a New York address.