Codes of Practice
Industry-developed guidelines that help AI providers demonstrate compliance with regulatory requirements.
Definition
Codes of Practice are industry-developed documents that provide detailed guidance on how to comply with AI regulatory requirements. Under the EU AI Act, codes of practice have specific legal significance for general-purpose AI models.
EU AI Act Framework (Articles 56): The AI Office facilitates development of codes of practice to help GPAI model providers demonstrate compliance. Key features:
- Drawn up with input from providers, civil society, and stakeholders
- Cover transparency requirements, copyright compliance, and systemic risk mitigation
- Voluntary but create presumption of conformity if followed
- AI Office may approve existing codes or request amendments
Content Requirements (Article 56(4)):
- Detailed description of measures for compliance
- Information on interaction with downstream providers
- Copyright policy and training data summary approach
- Systemic risk identification and mitigation measures
Legal Effect:
- Compliance with approved codes creates rebuttable presumption of conformity
- Non-compliance doesn't automatically mean violation—alternative evidence permitted
- Commission may adopt common specifications if codes inadequate
Comparison to Self-Regulation: Unlike purely voluntary industry codes, EU AI Act codes of practice operate within a regulatory framework with oversight and potential enforcement.
Related concepts: Harmonised Standard, Common Specification, Self-Assessment, General-Purpose AI Model
Sources
- •EU AI Act Article 56
- •Industry Self-Regulation
Related Terms
Harmonised Standard
A European standard requested by the European Commission that, once cited in the Official Journal, gives a rebuttable presumption of conformity with relevant EU harmonisation legislation....
Common Specification
An EU-issued set of technical specifications that providers can follow to demonstrate compliance with certain AI Act requirements when harmonised standards are unavailable or inadequate....
General-Purpose AI Model
An AI model trained at scale that displays significant generality and can competently perform many distinct tasks and be integrated into varied downstream systems....