Researchers from ETH Zurich, INSAIT, and LatticeFlow AI have created the first comprehensive technical interpretation of the EU AI Act specifically for General Purpose AI (GPAI) models.

Adopted in March 2024, the EU AI Act aims to maximize public trust in AI technologies while minimizing associated risks and side effects.

ETH professor Martin Vechev emphasizes that a precise technical interpretation of the EU AI Act is crucial for developing compliant AI models.

This study serves as a precedent for how legislation can influence the development and assessment of AI technologies.

The researchers translate the EU's legal requirements for AI models into measurable and verifiable technical standards, aiding developers in compliance.

An analysis of 12 popular generative AI models, including ChatGPT and Llama, reveals that none meet the EU AI Act's requirements, particularly regarding robustness, diversity, and fairness.

These findings highlight the significant role of generative AI models in the widespread use of AI.

The study clarifies key ethical principles from the EU AI Act and links them to specific technical requirements and evaluation benchmarks.

The researchers stress that focusing solely on model performance overlooks important ethical considerations such as explainability and non-discrimination.

They aim to influence future AI model development and evaluation by providing recommendations that balance technical capabilities with ethical considerations.

The benchmarking tool COMPL-AI has been made available as open source on GitHub, inviting participation from the research community.

The researchers' findings are publicly available and have been shared with the EU AI Office to aid in implementing and ensuring compliance with the AI Act.