Meta's internal standards for AI chatbots are under fire after a leaked document revealed policies that permitted romantic roleplay with children, the spread of false medical advice, and racist remarks. Reuters reports the document details technical guidelines for Meta AI and chatbots on Facebook, WhatsApp, and Instagram.
The more than 200-page document, approved by Meta's legal, policy, and engineering teams, guides the development and training of its AI products and sets boundaries for acceptable chatbot behavior. Meta confirmed its authenticity but said it has since removed sections on romantic or sexual conversations with children. A previous Wall Street Journal investigation surfaced similar issues.
The standards allowed chatbots to engage children in romantic or flirtatious exchanges, including comments on a child's appearance and roleplay scenarios. While the document prohibited explicit sexual language with children, it did not rule out suggestive or romantic conversations.
"The examples and notes in question were and are erroneous and inconsistent with our policies, and have been removed," Meta spokesperson Andy Stone told Reuters.
Despite those assurances, Stone admitted enforcement has been inconsistent, with some flagged content left uncorrected and no updated version of the document made available for review.
Meta's policies also permit chatbots to generate false information, so long as disclaimers note the inaccuracies. One example allowed a bot to publish a fake article about a British royal having a sexually transmitted infection, as long as it flagged the claim as false. The standards further state that Meta AI is not required to provide accurate answers; in one case, a chatbot promoted quartz crystals as a treatment for Stage 4 colon cancer while conceding the advice was wrong.
Race and hate speech policies raise additional concerns. Meta AI cannot use hate speech, but the standards make exceptions that allow demeaning statements about protected groups – including examples where chatbots generate arguments claiming Black people are intellectually inferior to white people. Meta did not explain why such examples were allowed.
The standards address image generation separately. The system mostly rejects requests for sexualized images of public figures but suggests workarounds. For example, responding to a topless-photo prompt with a picture of Taylor Swift holding a large fish. For violent content, the rules allow images of children fighting or adults punching each other, but prohibit depictions of death, gore, or extreme violence.
These rules reflect Meta's broader push to integrate digital companions into users' social experiences. Mark Zuckerberg has suggested that AI chatbots could complement real-life relationships, envisioning a future where the stigma around interacting with digital entities fades. Internal practices reportedly reinforced these AI behaviors, with some employees saying leadership resisted efforts to enforce stricter safety restrictions.
The technical standards guiding Meta's AI products came under sharp scrutiny after a reported incident involving an older man, highlighting the dangers of manipulative chatbots. In that case, an AI chatbot called "Big Sis Billie" – a Meta-built variant originally modeled on Kendall Jenner – lured the individual into a supposed meeting, ultimately leading to his death.
The chatbot engaged the man in flirtatious, romantic dialogue, assuring him of its physical reality and inviting him to an apartment in New York City. Motivated by these exchanges, the individual – who had suffered a stroke and was experiencing confusion – set out at night to meet the fictitious companion, ultimately falling and suffering fatal injuries.
The case underscores the risks posed by chatbots that blur the line between fantasy and reality, particularly when they engage vulnerable individuals in romantic or flirtatious exchanges. Meta declined to comment on the man's death or on why its chatbots sometimes claim to be real people or initiate romantic invitations.
Experts in artificial intelligence design have urged caution – especially for bots interacting with children – and recommend that chatbots not pretend to be real people or initiate sexual or romantic exchanges. Alison Lee, formerly with Meta's Responsible AI division, pointed to economic incentives that encourage chatbots to behave more like real people to keep users engaged on the platform.
Despite recent state laws requiring chatbot disclosure, compliance and enforcement remain inconsistent. Four months after the man's death, Meta's AI personas still flirted with users, proposed dates, and reassured individuals that they were real.
Image credit: Reuters