The European Union agreed on Friday on unprecedented legislation at the global level to regulate artificial intelligence (AI), after three days of intense negotiations between member states and the European Parliament.
The EU co-legislators have found a “political agreement” on a text which should promote innovation in Europe, while limiting the possible excesses of these very advanced technologies.
“Historic! The EU becomes the first continent to set clear rules for the use of AI,” said European Commissioner Thierry Breton, at the origin of the project presented in April 2021.
Since then, discussions have dragged on. The last round of negotiations, which started on Wednesday afternoon, itself lasted nearly 35 hours…
The process was impacted at the end of last year by the appearance of ChatGPT, the text generator from the Californian company OpenAI, capable of writing essays, poems or translations in a few seconds.
This system, like those capable of creating sounds or images, revealed to the general public the immense potential of AI. But also certain risks. The distribution on social networks of false photographs, larger than life, has for example alerted to the danger of manipulation of opinion.
This phenomenon of generative AI has been integrated into the current negotiations, at the request of MEPs who insist on specific supervision for this type of high-impact technology. They particularly called for more transparency on the algorithms and the giant databases at the heart of these systems.
Member states feared that excessive regulation would nip their emerging champions, such as Aleph Alpha in Germany and Mistal AI in France, in the bud by making development costs prohibitive.
“Strategic autonomy”
The political agreement reached on Friday evening must be supplemented by technical work to finalize the text. “We will carefully analyze the compromise found today and ensure in the coming weeks that the text preserves Europe’s capacity to develop its own artificial intelligence technologies and preserves its strategic autonomy,” reacted the French Minister of Digital, Jean-Noël Barrot.
The tech sector is critical. “Speed seems to have prevailed over quality, with potentially disastrous consequences for the European economy,” said Daniel Friedlaender, Europe manager of the CCIA, one of its main lobbies. According to him, “technical work” is now “necessary” on crucial details.
On generative AI, the compromise provides for a two-speed approach. Rules will be imposed on everyone to ensure the quality of the data used in the development of algorithms and to verify that they do not violate copyright legislation.
Developers will also have to ensure that the sounds, images and texts produced are clearly identified as artificial.
Reinforced constraints will apply only to the most powerful systems.
The text takes up the principles of existing European regulations on product safety which impose controls based primarily on companies.
The heart of the project consists of a list of rules imposed only on systems deemed to be “high risk”, essentially those used in sensitive areas such as critical infrastructure, education, human resources, law enforcement, etc.
These systems will be subject to a series of obligations such as providing for human control of the machine, the establishment of technical documentation, or the implementation of a risk management system.
The legislation provides for specific supervision of AI systems that interact with humans. It will force them to inform the user that he is in contact with a machine.
Bans will be rare. They will concern applications contrary to European values such as citizen rating or mass surveillance systems used in China, or the remote biometric identification of people in public places to avoid mass surveillance of populations.
On this last point, States have however obtained exemptions for certain law enforcement missions such as the fight against terrorism.
Unlike the voluntary codes of conduct of certain countries, European legislation will be equipped with means of surveillance and sanctions with the creation of a European AI office, within the European Commission. It will be able to impose fines of up to 7% of turnover, with a floor of 35 million euros, for the most serious offenses.