There are signs of thaw but warming remains limited: on the eve of a crucial meeting on the draft European regulation on artificial intelligence AI Act, Wednesday December 6, the Minister for Digital Affairs Jean-Noël Barrot says he open to “integrate” the foundation models, these large language or image processing software used by companies to create services like ChatGPT or Midjourney. During a meeting on Friday, France and the Twenty-Seven gave a mandate to the Spanish presidency of the Council of Member States to negotiate on Wednesday with Parliament on their regulation. The latter has become a blocking point, with France being accused of opposing it to protect its national start-ups like Mistral.
“First, Europe is the first democracy to have a legislative framework for artificial intelligence,” explains Mr. Barrot. From now on, AI systems that present high risks, like toys or elevators, will have to obtain CE marking, with the highest audit and transparency obligations in the world. » Autonomous driving, medical diagnosis, CV sorting or credit attribution systems will be covered, but not “risk-free cases, such as entertainment or video games”, he notes.
On the sensitive issue of foundation models for general use, their manufacturers should first, according to Mr. Barrot, “collaborate actively and loyally with client companies who will deploy their models in high-risk services and who, for obtain CE marking, will need information”. Then, these manufacturers should “publish basic information on the operation of their models and the type of data used for their training”. Finally, they should “adhere to a code of good conduct which will harmonize their practices”.
Very tense context
Furthermore, the minister envisages that models published in free access (open source), such as those of Mistral, Meta or Hugging Face, could “be exempt from certain obligations, because open source offers guarantees of transparency and promotes competition fair “.
The context is very tense. France, supported by Germany and Italy, opposed the stricter obligations proposed in June by Parliament, which included regular assessments and risk mitigation measures for fundamental rights (fight against errors , illegal content, discriminatory bias, etc.). France also rejected the compromise reserving these obligations for the most powerful models.