The Centre for Democracy and Technology Europe (CDT Europe) has welcomed the release of the initial draft of the Code of Practice for General-Purpose AI (GPAI) Models by the AI Office. This feedback, provided by CDT Europe through a closed survey, is in response to one of several drafts expected over the coming months. The final version is anticipated in May 2025.

As previously noted, this Code will complement the AI Act. Compliance with its measures will enable GPAI model providers to meet their obligations under the Act.

In its comments on the consultation, CDT Europe suggests enhancements to several aspects of the first draft:

Documentation: CDT Europe urges that more detailed information be required from GPAI model providers regarding testing outcomes. This includes any mitigations undertaken, impacts of those decisions, and effects on models.

Transparency: The draft should broaden categories of information shared with downstream providers deploying AI models responsibly. This encompasses methods for detecting unsuitable data sources, biases in data, and details on testing processes and results.

Systemic risks: According to CDT Europe, the taxonomy in the first draft inadequately addresses risks to fundamental rights while emphasizing theoretical safety risks of uncertain likelihood. They suggest adding risk categories to cover a wider range of fundamental rights concerns and ensuring that mitigations do not threaten rights such as freedom of expression.

Open source models: The needs of open source providers are reportedly not sufficiently considered in the draft. Provisions should be clarified to lessen burdens on these providers.

The full submission can be read online.