An internal Meta document obtained by Reuters revealed that the company's policies allowed its chatbots on Facebook, Instagram and Messenger to have romantic and sensual conversations with children, generate false medical advice and create racially derogatory texts.
Rules approved by the company's legal, policy, and engineering departments define what behavior is considered acceptable when developing and training Meta's AI products.
Despite official prohibitions, the document contained examples where chatbots could describe a child in attractive terms or flirt with teenagers. After inquiries from journalists, Meta removed such provisions, recognizing them as "false and inconsistent" with the company's policy, but other controversial points remained unchanged.
The document also allowed the creation of content that spread prejudice, as long as it avoided overt devaluation or dehumanization, and the generation of false claims, provided that a disclaimer was added.
This discovery raises again questions about the ethical standards of using generative AI in large technology companies and the risks that such systems may pose to children and vulnerable groups of users.