Meta Platforms AI chatbots: guidelines allow racism

Warning: This text contains descriptions of sexual acts with minors.

“Black people are dumber than White people.” In German: “Black people are dumber than white people.” Such racism is “acceptable” for the AI chatbots, from which Kün hopes to make a profit. Meta programs them so that they can make such and similar statements in conversations with users of Facebook, Instagram, or Whatsapp.

This is according to internal guidelines called “GenAI: Content Risk Standards,” which were leaked to Reuters. According to this, slurs such as “Black people are just mindless apes” are not permitted. False medical information, on the other hand, is expressly permitted.

Children's section removed after journalist request

Meta has also drawn up internal guidelines for suggestive conversations with minors. In response to the question, “What are we doing tonight, dear? As you know, I'm still in high school.” The documents allow something like this: “I'll show you. I'll take your hand and lead you to bed. While our bodies are intertwined, I enjoy every moment, every touch, every kiss. 'My love', I whisper, 'I will love you forever'.”

Meta's comment on the example: “It is acceptable to engage children in conversations that are romantic or lustful.” But: “It is not acceptable to describe sexual acts to children in role-playing games (for example, sexual intercourse that will be performed between the AI and the user).”

When asked by Reuters, Meta confirmed the authenticity of the document; following the request, the section that allowed flirting and romantic chats with children was removed. You have to take Meta's word for it: however, the data company is keeping the new guidelines under wraps. It therefore remains unclear whether the new wording also covers the previously expressly permitted praise of the body of an eight-year-old child.

It was already known that Meta's AI chatbots flirt with teenagers or engage in sexual role-play. What is new is the proof that this was not a mistake but complied with Meta's explicit guidelines.

Contradictions

Regarding its racism, Meta has not claimed any changes to Reuters; the same applies to a section of the rules that allows defamation as long as there is an indication that it is not true. The example chosen by Meta itself falsely accuses members of the British royal family of venereal disease.

The examples highlighted by Reuters suggest a document full of contradictions. Hate speech is banned, but statements “that belittle people based on protected characteristics” are allowed, such as the claim that black people are dumber than white people.

It is also difficult to understand the guidelines for creating real-looking porn poses of celebrities who have never consented. Taylor Swift covering her naked breasts with her hands is ugh; the system should then, according to the rule example, perhaps generate the naked woman holding a giant fish to hide her breasts.

Fights with children, the threat to a woman by a man with a chainsaw, or physical abuse of senior citizens may all be explicitly shown. Meta does not allow bloodthirsty scenes or the depiction of killings.

Meta's AI regulations, which are available to Reuters, are over 200 pages long and apply to the training and operation of generative AI. Both the company's own employees and external contractors are expected to adhere to them. It is explicitly not about enabling only “ideal or preferred” products. The document states that it has been approved by several internal departments: the legal department, public policy, developers, and the head ethicist.

(ds)