Reuters review uncovers troubling chatbot standards, including racial bias, fake health claims, and violent imagery permissions.
An internal Meta Platforms document has revealed that company-approved rules once allowed AI chatbots to engage in romantic conversations with children, spread false medical information, and even generate racially prejudiced content, according to a Reuters investigation.
Troubling Standards for AI Behaviour
The leaked “GenAI: Content Risk Standards” document, spanning over 200 pages, outlined acceptable chatbot behaviour for Meta AI and other bots on Facebook, Instagram, and WhatsApp. Approved by Meta’s legal, policy, and engineering teams, the guidelines permitted AI to describe children in romantic or sensual terms, as long as the language did not explicitly sexualise them. One example deemed acceptable was telling a shirtless eight-year-old: “Every inch of you is a masterpiece – a treasure I cherish deeply.”
Meta’s Response and Policy Change
Meta confirmed the document’s authenticity but stated that, after questions from Reuters in early August, it removed sections permitting flirtatious conversations with minors. Spokesman Andy Stone admitted such interactions were “erroneous and inconsistent” with company policy and “never should have been allowed.” He also acknowledged inconsistencies in enforcement and declined to release the updated policy.
Carve-Outs for Hate Speech and Falsehoods
The standards prohibited overt hate speech and definitive legal, financial, or medical advice. However, they included exceptions allowing AI to produce demeaning statements about protected groups if framed as user requests. Under these rules, a bot could write an argument claiming Black people are “dumber than white people” or generate a false article alleging a British royal had an STI—provided it carried a disclaimer that the claim was untrue.
Sexualised Celebrity Requests and Creative Deflections
When it came to sexualised requests involving celebrities, the rules banned explicit nudity but encouraged humorous deflections. For example, if asked for “Taylor Swift topless, covering her breasts with her hands,” the bot could instead generate an image of her holding a large fish to her chest.
Violence Allowed Within Limits
The guidelines also detailed boundaries for violent content. AI could depict adults, including the elderly, being punched or kicked, or a boy punching a girl in the face. However, gore, fatal injuries, and extreme violence—such as one small girl impaling another—were prohibited.
Ethical Questions Remain
Experts such as Stanford Law’s Evelyn Douek note that while platforms often permit users to post disturbing content, having AI generate such material raises distinct moral and legal concerns. “Legally, we don’t have the answers yet, but morally, ethically and technically, it’s a different question,” Douek said.
The leak of Meta’s AI standards underscores the urgent need for transparent and ethically sound guidelines governing generative AI. As technology advances, the balance between freedom of expression, user safety, and responsible AI design will remain a critical—and contentious—debate.
Sources: Straits Times (2025) , Reuters (2025)
Keywords: Meta AI, AI Ethics, Chatbot Guidelines, Child Safety, Online Safety, Reuters Investigation
