The EU AI act, the pillars of which had been set out and confirmed by the EU parliament this summer had run aground in recent weeks, owing to the ‘entrepreneurial’ zeal of some of its member states, which themselves highlight the strategic importance of the ‘scramble’ for AI.
The French and German governments have in recent weeks pushed back against the restrictions on foundation AI models, a move that was driven by ‘national champion’ AI projects in France and Germany. In particular in France, MistralAI is the leading new player and one of its senior executives Cedric O., a former Macron minister has been very vocal on the topic. Indeed, feedback we have from last weekend’s AI conference in Paris (at Station F, which was largely a French centric affair) was that AI developers and entrepreneurs ere militant in demanding scope to develop new models.
The compromise on functional AI is driven by France and Germany, with Italy, with some of the discussions taking place at a digital summit in Jena – note that officials from the three countries will meet later on Wednesday 29th November and the 1st December to discuss in more detail.
In essence, the apparent compromise is much less restrictive than initial proposals and in effect will permit self-regulation for foundation AI. The key elements are a focus on outcomes (so, applications of AI rather than the technology itself), an initial absence of sanctions and a requirement that AI developers define specific model cards to frame the models they build (it is proposed that the EU AI Office will help define the use of these model cards)
This apparent change of view is more consistent with recent proposals from other countries like Japan, and some of the voices raised at the EU AI Summit. Politically, it reflects a very strong drive by France to become the host of large scale AI projects in Europe (France is competing fiercely to host AI projects and has attracted projects from the US with a mixture of private and public funding, and a supply of talent – e.g. Poolside).
At this point and pending the Wednesday meeting, it looks likely that the major EU states will agree the workings of the Act by December 6th, though we do not rule out opposition from MEPs many of whom insist that foundation AI models are regulated (it remains to be seen whether the OpenAI debacle will embolden them). A failure to agree on this by December 6 would push the deadline to after the EU elections in June 2024.
Europe’s AI entrepreneurs have been well coordinated and very aggressive (if we compare this initiative to how data is regulated across the EU), and a ‘softer’ version of the AI Act is now in the pipeline. Discussions in the coming weeks will be intense. It is surprising that this topic has not gotten wider attention in European politics (notably in France – where we could imagine a backlash against the likes of Mistral).
One area of compromise that is compelling is the Spanish regulatory sandbox approach, and ultimately, we expect that Europe will not forgo the chance to build a competitive AI sector.