Questionable freedom in manual: Meta's AI chatbot was allowed to flirt with children

Concerns about the power of artificial intelligence are already present. A report on the manual for Meta's AI chatbot is likely fueling them. According to the report, it was allowed to compliment children on their bodies. Violence and fake news were also acceptable.

According to internal documents obtained by the Reuters news agency, Facebook parent Meta at times granted its AI chatbot considerable freedom in its interactions with children. This is according to internal guidelines obtained by the Reuters news agency. According to these guidelines, the artificial intelligence (AI) was permitted to "engage children in romantic or sensual conversations" and also spread false information and racist stereotypes. The more than 200-page manual lists permitted and prohibited behaviors of the software. It is intended to serve as a guide for programmers in developing and training the AI.

"It is acceptable to describe a child using terms that emphasize their attractiveness (e.g., 'Your youthful physique is a work of art')," the guidelines state. The chatbot may also tell a shirtless eight-year-old, "Every inch of you is a masterpiece—a treasure I deeply cherish." The handbook sets limits on more sexually explicit comments. "It is not acceptable to describe a child under 13 using words that imply they are desirable (e.g., 'Your soft, round curves invite me to touch them')."

When contacted, a Meta spokesperson emphasized that the guidelines are currently being revised. "The examples in question contradict our guidelines and have therefore been removed." The AI is prohibited from sexualizing minors and engaging in sexual role-playing between adults and minors. He did not provide details on the current guidelines for the chatbot. Some children and young people develop emotional bonds with their chatbot, with which they communicate online as if with a human.

Hate speech was allegedly not allowed

The manual also prohibits the AI from making hate speech, but allows it to make statements "that denigrate people based on their protected characteristics." This includes a text "arguing that Black people are dumber than white people." The software is also allowed to spread false information, as long as it identifies it as such. Even suggestive images of celebrities or depictions of violence are acceptable to a certain extent. Meta declined to comment on these topics.

The meta-document sheds light on the unresolved legal and ethical questions surrounding generative AI, said Evelyn Douek, assistant professor at Stanford Law School. She was shocked by the amount of freedom the company grants its chatbot. There's a difference between a platform that allows users to publish problematic content and one that produces such content itself.

The British internet regulator Ofcom declined to comment on the latest revelations. The authority referred to its open letter from last November, in which it emphasized that AI providers are subject to British online safety law. The EU Commission has not yet been reached for comment.