European AI's Foundation for Security and Reliability
In the European Union, a comprehensive system of institutions, laws, and methods known as the European Quality Infrastructure ensures products and services are safe, reliable, and conform to functional and quality requirements across various sectors. This approach is now being extended to advanced AI products through the EU AI Act.
Effective in stages since early 2024, the EU AI Act sets a legal framework identifying high-risk AI systems that significantly impact health, safety, fundamental rights, or the environment. These systems must comply with multiple layered obligations, including risk management, transparency, human oversight, and importantly, independent conformity assessments before entry into the market.
Conformity assessments involve external, independent testing bodies verifying compliance with safety, robustness, fairness, data governance, and cybersecurity requirements mandated by the AI Act. This mirrors testing regimes applied in regulated sectors like automotive and medical devices, enhancing public trust and safety assurance.
The EU has an established process for setting technical standards called harmonized standards, which provide a presumption of conformity with legal requirements. AI product providers who comply with these harmonized standards and pass independent assessments can more easily demonstrate compliance to market authorities.
The EU AI Act also sets up institutional infrastructure, including the European AI Office and AI Board, to oversee implementation, enforcement, and market surveillance for these assessments, ensuring a sustained regulatory regime comparable to other high-risk sectors.
For general-purpose AI (GPAI) systems, a separate but complementary approach exists involving a Code of Practice supplemented by transparency obligations and risk assessments, with ongoing development of additional standards and guidelines to cover conformity requirements.
Adversarial testing by independent experts can help uncover potentially dangerous features in AI models. Developing standardized measures for AI testing is expensive and time-consuming. Three approaches to testing and certification are certification of quality management systems, product testing, and post-market methods of periodical inspection. Mandating pre-market product testing, including adversarial testing, and regular periodic assessments throughout the lifecycle by third-party auditors could be beneficial.
Regulatory sandboxes and testing, and experimentation facilities can help roll out testing and measurement infrastructure. AI developers above a certain size threshold should contribute funding to ensure a fair allocation of costs for building an assessment ecosystem. Making third-party testing compulsory would ensure impartial product testing and motivate AI companies to fund the development of risk assessment measurement units.
Independent conformity assessments cannot be 'one and done' and periodical inspections of products already on the market are becoming essential. Safety is a mandatory precondition in the system of mandatory independent assessments, providing a clear standard for manufacturers and minimum levels of safety for consumers. Developing an ecosystem of independent experts who can audit and inspect AI models will take time and significant resources.
Academic research shows that self-assessments deliver lower safety and security standards than accredited third-party or governmental audits. European businesses who are part of the Quality Infrastructure ecosystem perform well in comparison to their international counterparts. Product testing evaluates the product itself through independent examinations across industries. Independent evaluations in AI could assess data quality, model robustness, accuracy, and bias. Periodical inspections ensure safety and proper functioning after commercial distribution, particularly for commodities like cars and industrial installations.
External scrutiny is essential to ensure diverse perspectives and incentives aligned with finding vulnerabilities in AI products. Large-scale foundation models in AI lack mandatory independent, third-party testing to comply with safety rules. The independence and competence of conformity assessment bodies is ensured by accreditation in the European Quality Infrastructure ecosystem.
The EU AI Act, White House Executive Order, G7 Hiroshima Process, and Bletchley Declaration have made commitments to some external scrutiny and testing for the most advanced AI products, but these are not yet mandatory or fully implemented. Certification of quality management systems examines production processes and management structures in place to provide a certain level of safety.
In essence, the EU mandates independent third-party testing chiefly by defining which AI systems are high-risk under the AI Act, requiring these systems to undergo independent conformity assessments against mandatory safety and reliability criteria before market launch, utilising the harmonized standards mechanism to establish technical benchmarks that testing bodies use, establishing institutional bodies (AI Office, AI Board) and processes for enforcement, market surveillance, and continuous updates to the regulatory framework, and ensuring the independence and competence of conformity assessment bodies. This approach closely resembles regulatory regimes already in place for automotive, healthcare, and industrial machinery products, ensuring AI safety and public trust through transparent, external verification by accredited third parties.
- The EU AI Act, in line with its comprehensive approach in other sectors, extends this regulatory regime to advanced AI products, mandating independent third-party testing for high-risk AI systems, focusing on cybersecurity, technology, and policy-and-legislation within the context of politics and general-news.
- As part of the European Quality Infrastructure, conformity assessments for AI systems ensure compliance with safety, robustness, fairness, data governance, and cybersecurity requirements, contributing to public trust and safety assurance, aligning with the broader discourse on technology, policy, and general news.
