GPAI Code of Practice

The GPAI Code of Practice is the compliance framework developed by the EU AI Office to operationalise the General Purpose AI obligations under the EU AI Act. Companies that placed GPAI models on the EU market after August 2025 must already comply with the Code; models placed before that date have until August 2027. The Code covers technical documentation, capability evaluations, incident reporting, and cooperation obligations owed to the AI Office as it prepares to exercise full enforcement powers from 2 August 2026. Major AI providers including Anthropic, Google, and Microsoft have signed the Code. Meta declined to sign the GPAI Code of Practice, a position confirmed as of the Article 50 briefing review in June 2026.

The Code is distinct from the Article 50 Code of Practice, which specifically governs machine-readable marking of AI-generated synthetic content. The GPAI Code addresses the structural obligations of frontier AI model providers: systemic-risk assessments, red-teaming requirements, transparency reporting to the AI Office, and incident notification. It is the primary accountability instrument between the EU AI Office and the largest AI labs operating in EU markets. Non-signatories are not automatically non-compliant under the Act, but the AI Office has indicated that signatories benefit from a rebuttable presumption of compliance at audit.

As of June 2026, the GPAI Code operates in the months immediately preceding live enforcement, meaning the AI Office's first enforcement actions will be shaped by which obligations signatories have self-certified against. No EU broadcaster has publicly signed the separate Article 50 Code of Practice, which is the sector-specific instrument for synthetic-content labelling. The GPAI Code and Article 50 Code address different parts of the AI Act's compliance architecture and apply to different categories of AI system.