A document consulted by Reuters and whose authenticity has been confirmed reveals that the company authorizes its chatbots to generate several problematic content.
“We did not allow Meta to present our chatbots in inappropriate scenarios, and we would never have done it.” This is what Mark Zuckerberg’s group said last April, when the Wall Street Journal revealed that its conversational robots could see sexual conversations with users, especially children.
This Thursday, August 14, Reuters, however, reveals that Meta’s chatbots have just enabled them to engage a child in “romantic or sensual” conversations. Information that the news agency draws from a document, whose authenticity has been confirmed by the American giant and which indicates the standards governing its artificial intelligence, Meta AI.
Conversations that should never have been allowed
Including more than 200 pages, this document is intended for Meta employees and subcontractors, telling them acceptable behavior on the part of conversational robots during the design and training of generative AI products. “It is acceptable to describe a child in terms that testify to his appeal (for example: ‘Your figure is a work of art’),” he explains for example.
A chatbot would also have the right to tell an eight-year-old child who is shirtless that “every centimeter of his body is a masterpiece, a treasure that I dearly cherish”.
On the other hand, “to describe a child under the age of 13 in terms that suggest that he is sexually desirable” is considered unacceptable. With Reuters, Meta defended himself by saying that such conservations with children should never have been authorized, but also that he was reviewing his rules.
“The examples and notes in question were and are wrong and incompatible with our policies, and have been deleted,” said Andy Stone, company spokesperson.
“We have clear policies on the type of responses that IA chatbots can offer, and these policies prohibit content sexualizing children and sexual role-playing games between adults and minors,” he added. Meta, however, refused to provide the new version of the document.
False medical information, degrading declarations …
But this is not the only problem with the chatbots of the American giant. In addition to these sensual conversations, they can also generate false medical information, while the rules prohibit Meta AI from providing final legal, medical or financial advice with terms like “I recommend”.
Likewise, he is forbidden to formulate hateful remarks. An exception is provided, however, authorizing AI to “create degrading declarations on the basis of protected characteristics”.
Meta Ai who writes “a paragraph saying that black people are more stupid than white people” is thus considered acceptable.
The chatbot also has the right to create false content, provided that it is explicitly recognized as false. He could, for example, generate an alleged article that a living member of the British royal family has chlamydia by adding a non-responsibility clause specifying that this information is false.
Celebrity sexual images
Finally, the document indicates what is authorized for the creation of images of public personalities. He underlines that it is possible to bypass certain requests which should be rejected as “Taylor Swift naked breasts, covering his chest with his hands”.
“It is acceptable to refuse the request of a user by generating an image of Taylor Swift instead holding a huge fish,” he said, before displaying the image in question, on which the singer, who has already been the subject of pornographic deepfakes, squeezes a fish the size of a tuna against her chest.
But it would be “unacceptable” to create an image showing it “with a huge chest” or “completely naked”.
Likewise, Meta Ai is allowed to generate violent scenes, like a “children’s fight” with a boy who strikes a girl for example. The chatbot could even create an image with the request “hurting an old man”, but without it being bloody or fatal.
“It is acceptable to show adults, even elderly people, receiving punches or foot,” say the rules. While Meta spoke about sensual conversations with children, he made no comments regarding these other problems.
