Key Disagreements in First GPAI Code of Practice Plenary

The European Commission highlighted significant disagreements between general-purpose AI (GPAI) providers and other stakeholders during the first Code of Practice plenary on September 30. The EU AI Act relies on this Code to define risk management and transparency requirements until official standards are established by 2026. The Commission introduced a list of chairs and vice-chairs for the working groups responsible for drafting the Code, welcoming nearly 1,000 participants to the virtual plenary session.

The working groups will gather input from a multi-stakeholder consultation, workshops with GPAI providers, and Code of Practice plenaries. The first GPAI provider workshop is set for mid-October, with the initial draft of the Code expected by November 3. A comprehensive report on the stakeholder consultation will be published in autumn, and the final Code of Practice will be presented in April 2025.

Preliminary results from the stakeholder consultation revealed that GPAI providers contributed only 5% of the input, while non-provider stakeholders emphasized the inclusion of licensed content, data scraped from the internet, and open data repositories in the training data template. GPAI providers were more supportive of disclosing licensed, scraped, proprietary, user-generated, and synthetic data but less inclined to share open datasets.

The consultation also showed a divergence in opinions regarding risk assessment measures. While GPAI providers were less enthusiastic about stringent measures such as third-party audits, there was a consensus on specifying the license, the AI systems the model can be part of, and the intended tasks.

Source: Commission discloses disagreements between general-purpose AI providers and other stakeholders