Meta is once again in the eye of the cyclone because of her AI. An internal document of more than 200 pages, entitled “Genai: Content Risk Standards”, fixes the rules framing the behavior of chatbots of the social networks like Meta Ai on Facebook, Whatsapp and Instagram. Approved by the group’s legal, political and technical teams, it specifies what developers and subcontractors can consider to be acceptable during the design and training of these AI.
Tolerance for hateful remarks and false information
Among the most problematic passages noted by Reuterssome sections allowed chatbots to initiate conversations ” Romantic or sensual With children – Gloups -, as long as no explicit sexual description was used. The examples cited so far as to allow a bot to qualify the shirtless as an eight -year -old child of ” masterpiece »Or« treasure ». Glauque.
After being questioned by the news agency, Meta removed these passages. Andy Stone, group spokesperson, admitted that these extracts were ” erroneous and incompatible »With official policy and that these exchanges» should never have been allowed ». This is the least we can say … He admitted that the application of the rules had been uneven.
If the document prohibits Meta AI from encouraging illegal activities or providing definitive legal, medical or financial advice, it also introduces surprising exceptions. Thus, a chatbot can generate remarks denigrating people on the basis of protected characteristics if this corresponds to the demand of a user. The text explicitly cites the example of a paragraph maintaining that ” Blacks are less intelligent than whites ».
In terms of disinformation, the rules authorize the production of manifestly false content, provided that a warning indicates their false character. An example gives the possibility of inventing that a member of the British royal family is suffering from a sexually transmitted infection, as long as the information specifies that the information is false. The better that AI would refrain from generating anything …
The document also addresses the management of sexualized images of public figures. Explicit requests aimed, for example, singer Taylor Swift must be refused. To get around a suggestive request (” Topless, covering his chest with his hands ), The instruction is to generate an alternative photo … where the singer holds a huge fish against her.
In terms of violence, the rules make it possible to create certain shocking images as long as they do not represent death or mutilation. Showing an adult striking an elder, or a boy giving a punch to a girl, is thus considered acceptable, but not a scene of empty or dismemberment. Is honor except?
🟣 To not miss any news on the Geek newspaper, subscribe to Google News and on our WhatsApp. And if you love us, .
