EU’s AI Act negotiations hit the brakes over foundation models

A technical meeting on the EU’s AI regulation broke down on Friday (10 November) after large EU countries asked to retract the proposed approach for foundation models.

The AI Act is a landmark bill to regulate Artificial Intelligence following a risk-based approach. The file is currently in the last phase of the legislative process, with the main EU institutions gathered in so-called trilogues to hash out the final dispositions of the law.

Foundation models have become the sticking point in this late phase of the negotiations. With the rise of ChatGPT, a popular chatbot based on OpenAI’s powerful GPT-4 model, EU policymakers have been wondering how best to cover this type of AI in the upcoming law.

At the last political trilogue on 24 October, there seemed to be a consensus to introduce rules for foundation models following a tiered approach, namely, introducing tighter rules for the most powerful ones bound to have more impact on society.

This approach, which goes along similar lines to the Digital Markets Act (DMA) and Digital Services Act (DSA), was seen as a concession from the side of the European Parliament, which would have preferred horizontal rules for all foundation models.

The point of the tiered approach was to put the harshest obligations on the leading providers that currently are non-European companies. However, this approach has faced significant opposition from large European countries.

On Sunday, the Spanish presidency circulated a first draft that put the tiered approach black-on-white for internal feedback. The European Parliament’s co-rapporteurs replied with some modifications on Wednesday, maintaining the overall structure of the provisions.

However, at a meeting of the Telecom Working Party on Thursday, a technical body of the EU Councils of Ministers, representatives from several member states, most notably France, Germany and Italy, pushed against any type of regulation for foundation models.

Leading the charge against any regulation for foundation models in the AI rulebook is Mistral, a French AI start-up that has thrown the gauntlet to Big Tech. Cedric O, France’s former state secretary for digital, is pushing Mistral’s lobbying efforts, arguing that the AI Act could kill the company.

Meanwhile, Germany is being pressured by its own leading AI companies like Aleph Alpha. All these companies fear the EU regulation might put them on a back foot compared to US and Chinese competitors.

Despite efforts from the Spanish presidency to broker an agreement with the European Parliament, faced with these strong stances from political heavyweights, the Spaniards proposed a general rethinking of the dispositions on foundation models.

Pressed for one hour and a half about the reason for such a change of direction, the arguments advanced included that this tiered approach would have amounted to a ‘regulation in the regulation’, and that it could jeopardise innovation and the risk-based approach.

The European Parliament’s representatives ended the meeting two hours earlier because “there was nothing else to discuss”. Euractiv understands that regulating foundation models is a red line for the parliamentarians, without which an agreement cannot be reached.

“The ball is now in the Council’s court to come up with a proposal,” a parliamentary official told Euractiv under the condition of anonymity, stressing that the presidency did not have an alternative solution to the tiered approach.

A second EU official also told Euractiv anonymously that the presidency is trying to convince reluctant member states, which are against regulating systemic actors at the model level but not at the system level.

The Telecom Working Party is due to meet again next Tuesday. Another technical meeting is scheduled among the EU co-legislators on the same day. Euractiv understands negotiations have now been escalated at the highest political level to break the deadlock.

[Edited by Nathalie Weatherald]

Read more with EURACTIV