Damien d’Houdain, investment director of Bpifrance

La French Touch : The EU has equipped itself, with theAI Act, a legal framework to regulate artificial intelligence. A historic first. It is also a political agreement on a European scale. What about the rest of the world?

Damien d’Houdain (DH) : The subject of regulation is first of all on a global scale: the North American bloc has historically had a lower propensity to regulate business. It is the invisible hand of Adam Smith that regulates the market, priority to business, the actors regulate through bilateral agreements. On the other hand, the European approach which is characterized first by a culture of copyright inherited in France from Beaumarchais, and by a tradition of more important regulation and which can be considered, in a certain way, as a brake on the expansion, competitiveness and innovation of the actors. Let us take the example of a European LLM model of Gen AI (Generative Artificial Intelligence), in direct competition with OpenAI: our way of regulating will influence its expansion and competitiveness on a global scale. However, we can hear that the defense of copyright is a challenge with strong resonance at a time when AI is shaking up practices to the point of threatening to return to the Wild West that we knew at the beginning of the Internet when content piracy was endemic. Technological iterations put copyright at risk.

FT: The AI ​​Act introduces a dialectic between the security of Europeans and the competitiveness of EU companies. How is this act broken down? What are its main principles?

DH: The AI ​​Act introduces a first level of classification of AI technology by level of risk. This is the first chapter, with four gradations: AI with “unacceptable” risk which concerns activities such as biometric surveillance in public spaces; AI with “high” risk relating, for example, to nuclear power, transport services, air, railways, educational systems; AI with “limited” risk relating, for example, to “chatbots”; and, finally, “minimal” risk, with few or no regulatory requirements, this is the case for spam filters.

The second chapter of the AI ​​Act develops the issues of transparency and documentation. Developers, as part of the update, construction, improvement or training of their algorithm, have an obligation to provide clear information on its operation and to ensure the traceability of the data they use. Concretely, this means that it would be possible, on the basis of this text, to ask Mistral AI to explain, in detail, how its algorithm is built, on which data set the model is trained, and whether the rights holders of this data have approved its use. That is still very ambitious! Now, let's see how this text applies in practice. By the possibility of an audit, for example? That is the real issue...

The third part of the AI ​​Act addresses ethical considerations and the necessary human supervision: companies are responsible for the compliance of their system. They must document in detail the development processes and the quality of the data used. Once again, a notion of control is introduced here in order, in particular, to ensure the quality of the algorithm training. Lastly, sanctions are provided for in the event of violation with a fine of up to 30 million euros or 6% of the company's total turnover. Finally, the AI ​​Act provides that a European Council ofartificial intelligence be established, so as to oversee the implementation and enforcement of the legislation across Member States.

FT: The application of these rules results in an economic cost for the company. Is this cost seen as a barrier?

DH: Compared to North America, we can assume that the application of reinforced regulatory constraints will impose an economic burden on the players and that this is disadvantageous for the model of European AI players. We can also wonder whether compliance with the regulations will not limit the technical capacity of the models, part of whose value is known to lie in the volume and quality of the data captured. So, yes, it is a brake and a potential distortion of competition. Now, we need to see the extent of it, and to what extent the application of this regulation will generate additional costs for players in Europe.

FT: The music industry is particularly threatened by generative AI. We see this clearly with the multiplication of “ deepfakes " (for example, the restoration of John Lennon's voice for the production of a new Beatles album). Sacem is demanding that these 200 members be able to exercise a right ofopt-out. What are the major issues to be raised by this sector in the face of this obvious threat?

DH: L'opt-out is the ability of a content publisher to prohibit access to its content to algorithms that would "mine" the data to train themselves. Generative AI is an accelerator of creativity in that it lowers the barriers to entry to creation by opening up, technically to everyone, the ability to create a musical title simply by "prompting" a generative AI: this is the  text-to-sound. Generative AI multiplies the volumes of creation: it is estimated that more than 100.000 new titles are added every day to the catalog of online music platforms (Spotify, Deezer, Apple Music, etc.). This is not without risk of dilution for the "majors" on the platforms: will they be able to capture tomorrow, in an environment where the volume of creation is exploding, as much revenue as they do today?

Second impact: the control and detection of “ deepfakes » for platforms like YouTube, Deezer, Spotify… By ensuring that the artist's voice has not been misappropriated and that the projects respect copyright. The music industry is one of the most challenged industries in the short term: the state of the art in technology and AI already poses vigilance issues for historical players. From a perspective tech-enthusisast, the extraordinary creative capabilities of generative AI are very exciting and expand the possibilities tenfold. You must have listened to covers of the Beatles and other artists resurrected by AI. Recreate, innovate, modify, iterate, AI unleashes incredible creative power. Between this technophile aspect and the fear of copyright peril, there is a path to follow. It is in this context that this opt-out request emerges. Regulation will also come from the actors among themselves, accompanied by regulations which have power in terms of structuring the sector.

FT: This is already the case for content publishers, such as press companies. Two positions stand out: on one side, the New York Times is suing Microsoft, which defends itself by saying that the New York Times hacked ChatGPT. And on the other, the Financial Times and Le Monde are signing partnerships with the American giants, accepting, for a fee, the use of their data.

DH: And the subject is strictly the same: the economic aspect. The challenge for all these players, whether through legal means or agreements, is to find fair compensation for the use of their content. It is also an opportunity for additional income in an environment where the press is being disrupted on its historical, physical media. The means of finding additional monetization relays and stopping the plundering of documentary resources of press titles for free. The power of regulation is here supplanted by the ability of the players to find bilateral agreements.

FT: Researchers have highlighted the fact that it is impossible to know the source of the data used by software like ChaptGPT. This implies absolute vagueness. How do we get out of this "Bermuda Triangle"?

DH: It must be understood that ChatGPT has swallowed all of Instagram, all of Pinterest, all of Facebook, all of the magazines, all of the books in the history of humanity… The scope of the AI ​​Act, which aims to audit the source and compliance from a “copyright” point of view of the entire corpus of data - which has fed engines, algorithms and neural networks of this level of complexity - is therefore an ambitious bet.

FT: Let's return to the debate between regulatory power and innovation capacity. How is this opposition expressed on the market?

DH: The arrival of generative AI can be compared to the invention of the steam engine. It is a " game changer » economic of an unprecedented scope. However, we cannot ignore the fact that there is a copyright system that organizes and controls. The challenge is to arrive at a fair framework between, on the one hand, respect for copyright and a business that respects these practices, and on the other, the capacity for innovation and the creation of leading groups. Unfortunately, there is no Gafam in Europe. At the time of the emergence of the Internet, Europe missed the opportunity to create groups that were sufficiently powerful on an international scale. The subject arises once again with the Gen AI iteration: will European AI champions benefit from a regulatory framework that is harmonious with their US counterparts, without which their chances of success could be reduced?

FT: Is this the meaning of this call for projects launched by the Ministry of the Economy and this investment objective of five billion euros per year over five years? To give business creation power to a sector which, for the moment, has no champion.

DH: There is a strong will. The Paris market is recognized through its talents. Some French companies are pioneers and enjoy an excellent reputation. We know that public money, particularly through Bpifrance but not only, has a role to play so that this new major technological iteration can give rise to European world leaders. Let's not miss the train that we missed during the Internet era.

FT: Can we say more about Bpifrance’s strategy?

DH: AI is a strategic priority of the government, which is dedicating 2,5 billion euros to it from the France 2030 program. A enthusiasm demonstrated by the recent record fundraising of 202 million euros carried out by the start-up H, which attracted renowned investors (Xavier Niel, Eric Schmidt and Bernard Arnault) as well as Bpifrance via its Large Venture fund. The entrepreneurs' bank also operates IA Booster France 2030, one of the many State programs which encourage the development of a French ecosystem ofartificial intelligence.

FT: How is the question of the sovereignty of the French model, but also of the French language, expressed? How not to disappear in this virtual world?

DH: It is estimated that English represents 55% of content on the web versus 5% for the French language, it is obvious that the result of a user request cannot comply with the best use, and the best richness of our language. From a perspective of “ soft power ", visibility of both France and Europe in the world, the ability to have domestic actors trained on French-language corpora, is a challenge. But beyond the language, it is the sovereignty and cultural visibility of our country that is at stake. The Internet giants are called Google, Apple, Amazon, etc. So yes, it is a challenge to preserve our language, our culture and our way of life. It is an important challenge to continue to be visible on the international scene to have domestic leaders. Once again, we must not miss the turn.