Will the AI Act help or hinder Europe’s catch-up?

A political deal has been reached for the EU’s AI Act after fierce negotiations. What does this agreement entail, and what are its implications for member states and the technology sector? What were the main sticking points in the negotiations between the European institutions and the largest Member States? What do these differences mean for the stability of the agreement? Why is AI a major issue for Europe? What would be the consequences for the European economy? What about France? Review with Rémi Bourgeot, economist and associate research fellow at IRIS.

 

What does this agreement entail, and what are its implications for member states and the technology sector?

The AI Act, which has been in the pipeline since 2021, has faced a number of challenges this year. The explosion in generative AI in particular has shaken up its approach. It is based on the levels of risk associated with various types of application, from a harmless spam filter to the unacceptable use of facial recognition in everyday life. The designers of this regulation, mainly in the European Parliament, had not envisioned the potential of generative AI and had to review their copy, adding specific provisions. The need for regulation as such can hardly be doubted. However, these urgent additions have raised the spectre of European start-ups, which are in the early stages of catching up, being crushed under the weight of labyrinthine regulation, which would paradoxically be more manageable for the much more advanced digital giants in the US.

The spectacular progress made in large language models is based on neural networks using billions of parameters calculated on data that offers little transparency. This raises issues ranging from privacy and copyright to the many security risks associated with their use and erratic operation. These very rapid developments call for flexible, even adaptive, regulation. Conversely, an approach based on hundreds of pages of self-referential considerations, intended to be set in legislative stone, runs the risk of rapid obsolescence.

Beyond simple exemptions, a degree of flexibility is made all the more necessary by the increasingly central role of open source in AI, an opportunity that is being seized by European start-ups. Most of them are appropriating some of the main language models developed by the industry giants, while some are also starting to develop their own underlying models. An open regulatory approach is therefore needed, both to deal with the new risks that are bound to emerge and to enable an inventive and original European AI industry to thrive in the face of the American and Chinese behemoths.

 

What were the main sticking points in the negotiations between the European institutions and the largest Member States? What do these differences mean for the stability of the agreement?

Lawmakers thought they were replicating the approach of GDPR, which has emerged as a kind of gold standard for data regulation around the world, in the face of social network platforms. But the technological challenge is different with AI. Europe is chronically lagging behind US digital giants. The AI Act adds uncertainty to the early successes, at least in terms of funding, that have been seen in France with Mistral and in Germany with Aleph Alpha.

In recent weeks, these concerns have led the French, German and Italian governments to try to tip the balance. A degree of confusion has developed between two distinct issues: on the one hand, the controversy over state’s use of facial recognition (which some member states refuse to give up entirely), and on the other, the challenge of preserving the potential of start-ups working on the foundation models that underpin generative AI. The trio of the EU’s three largest member states used confusing terms, proposing a form of self-regulation and code of conduct for these models.

Faced with this resistance, European institutions tried to keep up the momentum, backed up by NGOs expressing an often general view in favour of the AI Act as it stood. A compromise emerged, which included a fairly broad exemption for the open source community, which is at the centre of European initiatives on foundation models.

 

With French Digital Minister Jean-Noël Barrot declaring that the agreement would preserve Europe’s ability to develop its own artificial intelligence technologies and maintain its strategic autonomy, why is AI a major issue for Europe? What would be the consequences for the European economy? What about France?

Europe is worryingly lagging behind the US and China. This was not inevitable. In recent decades, neural networks have benefited from the hard work of European visionaries, whether they stayed on the continent or moved to the US. They promoted neural networks at a time when other types of models were dominating the AI scene.

Geoffrey Hinton, the British « godfather of AI », who quickly left the Silicon Valley to set up in Toronto, to keep some distance from the US military complex, comes to mind. It should be noted that this pioneer, also a great pedagogue, expresses himself in admirably clear English, both technically and philosophically, which could inspire our regulatory writers. Frenchman Yann Le Cun has also been pioneering neural networks since his engineering studies in the 1980s. He is now Chief AI Scientist at Meta, where he notably promotes open source models. As early as 1991, the German Sepp Hochreiter put forward the notion of long memory in neural networks in his master’s thesis… This led in 1995 to the famous LSTM (long short term memory) architecture, which is still in use and whose advances paved the way for language models based on « transformers », at the heart of the generative AI revolution since 2017. There are many European examples in the early successes of neural networks.

Despite the crisis in education, and in mathematics in particular, Europe, and France in particular, still has pockets of excellence that should be encouraged to deploy their talent for the benefit of a different kind of AI. The idea that Europe would thrive in an administrative role of global regulator, with widespread technological dependence on the United States and China, seems suicidal in economic and strategic terms, given the importance of AI in all aspects of industrial development. The appropriation of cutting-edge technologies has always been at the core of catch-up, development and power mechanisms. In recent decades, Europe has focused on competition policy at the expense of industrial issues. European citizens were seen more as consumers than producers, and even today, with the AI Act, more as users than developers. This trend has clearly started to reverse in recent years but the task remains daunting.