Last updated:
The detailed regulatory instrument outlining specific transparency, safety testing, and risk mitigation obligations for providers of General-Purpose AI (GPAI) models under the EU AI Act.
Drafted by the EU AI Office and finalized in August 2025, this Code provides signatories with a compliance pathway for systemic risk models. It mandates adversarial testing (red-teaming), rigorous technical documentation, and training data summaries. Signatories enjoy a presumption of compliance with Chapter V of the Act, which covers models like those with training compute exceeding 10^25 FLOPs.
Real world example:
A developer of a large language model (LLM) signs the Code of Practice to simplify their reporting obligations to the EU AI Office, documenting their watermarking techniques and their approach to preventing the generation of non-consensual imagery.
