Share This Article
The negotiations of the trilogue on the AI Act prepared the ground for the final meeting where the political agreement is expected to be reached.
A negotiation meeting of the trilogue between the EU Council, Parliament, and Commission was held on October 24, 2023 and focused on critical aspects of the AI Act. Unfortunately, it failed to reach a complete agreement. While there was consensus on classifying high-risk AI applications and providing general guidance on supervising powerful foundation models, discussions on prohibitions and law enforcement were left largely unresolved.
Based on the publicly available information, key highlights of the negotiations were as follows:
- Classification Rules for High-Risk AI Systems: The AI Act adopts a risk-based approach, requiring strict compliance for AI models posing significant risks to health, safety, and fundamental rights. A possible filter system was introduced, allowing AI developers to seek exemptions from the strict regime. However, the conditions for exemption criteria are still under refinement.
- Foundation Models: Policymakers grappled with the proliferation of powerful models. The tiered approach gained support but defining the top tier of ‘very capable’ foundation models, subject to vetting and risk mitigation, remains a challenge. Criteria such as computing power, training data, and resources are being considered.
- Governance: Centralized enforcement of AI models posing systemic risks is proposed, similar to what is provided for the Digital Services Act, where the 17 Very Large Online Platforms (VLOPs) and the 2 Very Large Online Search Engines (VLOSEs) are subject to direct monitoring by the European Commission. An AI Office, under the Commission but with functional independence, is envisioned, funded by a management fee. However, some MEPs are skeptical of this approach.
- Prohibitions and Law Enforcement: The chapter addressing which AI applications should be prohibited and how law enforcement agencies should handle exceptions remains a contentious issue. Negotiators aim to package bans, law enforcement exceptions, fundamental rights assessments, and environmental provisions together.
The next meeting of trilogue is scheduled for December 6th and should be the final meeting of negotiations to reach a political agreement on the AI Act. As such, the most intense negotiations are likely to take place in the coming month.
On the same topic, you can find the following article interesting “What is the AI risk management system required by the EU AI Act?”.
