In a crucial development, negotiations on the EU's AI Act have hit a roadblock due to disagreements over the regulation of foundation models. France, with support from Germany and Italy, has called for a retreat from the tiered approach, leading to concerns about the future of the comprehensive AI law.

The impasse occurred during a technical meeting on November 10, when major EU countries asked to retract the proposed approach for foundation models. This move has now jeopardized the entire legislation, unless a resolution can be reached in the coming days.

In response to these demands, officials from the European Parliament walked out of a meeting, signaling their opposition to excluding foundation models from the law. The Spanish presidency, which has been attempting to mediate between the Parliament and reluctant governments, has faced an uphill battle.

In an attempt to find common ground, the three heavyweight countries - France, Germany, and Italy - circulated a non-paper on November 19. The document argues against a two-tier approach for foundation models, stating that the inherent risks lie in the application of AI systems rather than the technology itself.

The non-paper proposes an alternative solution: building on mandatory self-regulation through codes of conduct. The signatories note that European standards can support this approach, while preserving innovation and safety. They further suggest drawing on the principles established at the G7 with the Hiroshima process.

Furthermore, the three countries argue that regulation should focus on General Purpose AI systems rather than foundation models, aligning with the risk-based approach of the AI Act. They propose that developers of foundation models define model cards, which would provide key information about trained models to a broader audience.

These model cards, according to the non-paper, should include relevant details about the capabilities and limitations of the model and adhere to best practices within the developer community. The proposal suggests establishing an AI governance body to develop guidelines and monitor the application of model cards, with a reporting mechanism for any code of conduct violations.

A crucial discussion on the approach to foundation models is set to take place at the Telecom Working Party, a technical Council body, on November 21. On the same day, MEPs will hold an internal meeting on the matter, followed by a dinner with the Council presidency and the Commission.

One of the key sticking points in the negotiations is the question of sanctions. France, Germany, and Italy are opposed to implementing sanctions at the beginning, suggesting that a sanction regime should only be established following a proper analysis and impact assessment of identified failures.

With the EU's AI rulebook reaching a delicate stage of negotiations, the Spanish Secretary of State for Digitalization and Artificial Intelligence, Carme Artigas, emphasized that the AI Act cannot turn away from addressing foundation models.

As the clock ticks, the fate of the world's first comprehensive AI law hangs in the balance. It remains to be seen whether a compromise can be reached regarding the regulation of foundation models, ensuring the harmonized approach to AI regulation in the European Union.