A “trilogue” meeting around the Artificial Intelligence (AI) Act was on the agenda on Tuesday 18 July. In this institutional ritual specific to Brussels, the European Parliament, the Commission and the Council, which represents the Member States, seek to find a compromise around a legislative project, in this case the European regulation on artificial intelligence. This phase of negotiation, scheduled to last several months, opens while, for several weeks, criticism has been heard, particularly in France, on the latest version of the AI Act, considered too restrictive.
“As it stands, regulation has a strong risk of hampering innovation in Europe”, had launched on June 14, at the VivaTech show, Arthur Mensch, the founder of Mistral AI, a start-up having raised 105 million euros to manufacture language processing models competing with those of the Americans Google, Meta or OpenAI, the creator of the chatbot ChatGPT. Significant fact, Emmanuel Macron, present on stage, had abounded in this direction: “I share your concern,” declared the President of the Republic.
The criticisms relate to the obligations added to the AI Act – in the version voted by Parliament on June 14 – in order to regulate AI models not falling under a specific use (the rest of the text regulates software in depending on the risk posed, considered greater for a medical or automotive application than for an interface advising a customer on an e-commerce site).
For “general purpose” artificial intelligences, capable of producing text for a poem, but also computer code for a nuclear power plant, manufacturers are required to assess and mitigate the potential risks to health, safety, fundamental rights or democracy. In the same spirit, “generative” AI interfaces, which, like ChatGPT or Midjourney, allow the public to create text or images, must be transparent about the data used for their training, specifying content subject to copyright.
“Lobbying exercise”
“The text applies to language processing models the obligations reserved for the most risky applications, without taking into account their use or their power”, regrets Cédric O, ex-Secretary of State for digital technology who became founder and adviser of Mistral AI, whose models will, according to him, be classified for general purposes. At LightOn, another French creator of large tongue models, the founder, Laurent Daudet, also says he is ” worried “. German Aleph Alpha ruled that complying with the AI Act could take too much time for start-up managers and deter some investors.
You have 54.27% of this article left to read. The following is for subscribers only.