AI Act: towards three-tiered generative AI regulation

On October 18, 2023, Bloomberg revealed the outlines of potential generative AI regulation in the European Union. The measure would be part of the AI Act, which was passed at first reading by the European Parliament in June 2023, and is undergoing debate in Brussels.

The recent boom in generative AI has indeed pushed European authorities towards stricter regulation. However, Brussels does not intend to create burdensome legal obstacles that would hinder newer players, particularly European ones. The consensus seems to be a three-tiered regulation, according to the size and popularity of the models.

To be granted authorization in the European Union, generative AI must first comply with transparency requirements, particularly in regard to its training process. It will also have to make sure models are completely safe.

“High-performance” models that required significant computing power to train, will have to comply with stricter regulation. They will be subject to testing by outside experts before their launch, then to regular audits and monitoring by European authorities.

Finally, “large-scale” models (over 45 million users or 10,000 corporate customers in the EU) will be required to perform impact studies before upload. The studies will make it possible to assess the risks involved with the models and the measures established to mitigate them.

A new body will be in charge of enforcing the regulation. In the event of proven danger or recurrent non-compliance with these rules, it will have the authority to ban models in the European Union.