Miles Brundage, a prominent voice in artificial intelligence (AI) safety, has called for a significant transformation in how AI system cards are developed and utilized. He advocates for these documents to transition from ambiguous claims to structured safety cases backed by verifiable evidence and independent third-party attestations. This statement underscores a growing demand within the AI community for enhanced transparency and accountability in the development and deployment of advanced AI systems.
Brundage, formerly a senior policy researcher at OpenAI and instrumental in developing their system cards, departed the company to focus on independent AI policy research. His background provides unique insight into the current state and future needs of AI safety documentation. He has consistently championed rigorous approaches to AI governance and has been vocal about the importance of robust safety measures.
AI system cards are typically detailed documents that outline an AI model's capabilities, limitations, safety measures, and ethical considerations. However, Brundage argues that their current form often consists of "semi-random assortments of often-ambiguous claims." This lack of structured evidence makes it challenging to truly assess the safety and reliability of AI systems, hindering effective oversight and risk management.
To address these shortcomings, Brundage proposes that system cards become "structured safety cases with directly verifiable evidence for some claims and attestations from third party auditors for others." This shift would necessitate a more rigorous, evidence-based approach to documenting AI system performance and safety. The involvement of independent auditors would provide an impartial validation of the claims made by AI developers, fostering greater trust and credibility.
This call aligns with broader industry trends and regulatory efforts aimed at strengthening AI accountability. Emerging standards like ISO/IEC 42001 and frameworks such as the NIST AI Risk Management Framework emphasize robust governance and independent verification. Furthermore, regulations like the EU AI Act are increasingly mandating conformity assessments, some of which will require third-party involvement, highlighting the critical role external validation plays in ensuring responsible AI development.