Based on the information provided in the JRC report, here are key points regarding the development and characteristics of harmonised standards for the EU AI Act:
Timeline and Progress
- The EU AI Act entered into force on August 1st, 2024.
- High-risk AI systems must comply with the Act’s requirements after a transition period of 2 or 3 years, starting from August 2026[1].
- The European Commission has engaged with European Standardisation Organisations since April 2021.
- The standardisation process has been slower than anticipated, with challenges in reaching consensus on new work items and their scope[1].
Characteristics of AI Standards
Standards supporting the AI Act are expected to have the following key attributes:
- Tailored to the AI Act’s objectives: Focus on risks to health, safety, and fundamental rights of individuals[1].
- System and product-centric: Cover all phases of the AI system lifecycle[1].
- Prescriptive and clear: Define explicit requirements for AI systems to meet[1].
- Broadly applicable: Provide horizontal requirements applicable across sectors and AI system types[1].
- State-of-the-art aligned: Address modern AI techniques and architectures[1].
- Cohesive and complementary: Ensure logical structure and capture interdependencies between requirements[1].
Key Areas of Standardisation
The European Commission has requested standardisation deliverables in 10 concrete aspects of AI, including:
- Risk Management: Specify a risk management system for AI products and services[1].
- Data Governance and Quality: Cover both data management and dataset quality aspects[1].
- Record Keeping: Define requirements for tracing and recording events in AI systems[1].
- Transparency: Outline transparency information required for compliance[1].
- Human Oversight: Define requirements for selecting and implementing oversight measures[1].
- Accuracy: Support the selection of relevant accuracy metrics and thresholds[1].
Challenges and Considerations
- Existing international standardisation efforts often focus on organizational objectives rather than individual risks, requiring a shift in approach for AI Act standards[1].
- Standards must balance prescriptiveness with flexibility to accommodate various AI systems and sectors[1].
- The rapid advancement of AI technology necessitates standards that can address state-of-the-art techniques[1].
These harmonised standards, once published in the Official Journal of the EU, will grant a legal presumption of conformity to AI systems developed in accordance with them[1].
Citations:
[1] https://publications.jrc.ec.europa.eu/repository/bitstream/JRC139430/JRC139430_01.pdf