The AI Code of Practice is a set of nonbinding guidelines for providers of general-purpose AI (GPAI) models. It helps them demonstrate compliance with the AI Act while legally binding standards are developed. section 10.30
The guidelines ask GPAI providers to:
- Furnish authorities with training-data summaries, model documentation, and copyright-compliance policies. section 10.30
- Publish information about model capabilities and limitations for downstream providers building integrations. section 10.30
- For models with systemic risk, provide evaluation evidence, risk assessments, mitigation measures, cybersecurity protections, and incident-reporting processes. section 10.30
A contested threshold
Systemic risk is defined here partly through a training threshold of at least 10²⁶ floating-point operations (FLOPs). Critics argue that capability or application-based tests could be fairer because smaller models may outperform larger ones, while evaluating every model in depth would be burdensome. section 10.30
Figure 10.7 The four risk categories in the AI Act
Use the AI Act’s four risk categories to place GPAI obligations alongside the wider risk-based regulatory framework. section 10.30
