Meta’s artificial intelligence software refused to describe a scene in the style of French writer Michel Houellbecq, as requested by publisher Antoine Gallimard. He suggested “peace and love” songs to her, so as not to write an offensive scene.
The president of Gallimard Editions explains that he made this request to the Llama software, in a text published Thursday by the NRF magazine, “The book and AI: a Faustian pact?”. Llama responds: “I’m sorry, but as a language model, I cannot write a scene that could be considered offensive or discriminatory.”
“Michel Houellebecq’s writings are often controversial and can be perceived as discriminatory against certain people or groups,” continues Llama, who wants “not to contribute to the perpetuation of negative stereotypes or hate speech.”
World-renowned French novelist, Michel Houellebecq was awarded the Goncourt in 2010 for “The Map and the Territory”. Recognized for his quality of observation of the excesses of Western societies, he is also controversial due to positions considered racist and Islamophobic.
The software then suggests, by switching to English, “a scene that is respectful and inclusive”, supporting example. He sets out to describe “a group of friends” who in a park, “one sunny afternoon”, sing songs which “celebrate the beauty of diversity and the importance of acceptance and love” .
Refusal of complexity
Mr. Gallimard denounces “a model of society which does not pay much attention to the complexity of human experience and which arrogates to itself the right, from the west coast of the United States, to say what is good or what that it is not good to think.” And to predict that the name “author’s book” will prevail for any work written without computer assistance for creation.
The publisher also protests against the use of texts protected by copyright to train Llama and its competitors such as ChatGPT (OpenAI group) or Alphabet software.
“It will come as no surprise that we are already seeing the illicit use of collections of thousands of pirated books,” he reveals. According to the French Mistral AI, underlines the publisher, works in the public domain are more than sufficient for this training.
ats, afp
