Standards are crucial in ensuring smooth market function, interoperability, and consumer safety, as well as aiding in the development of regulation for new technology. However, establishing standards for rapidly evolving artificial intelligence (AI) technologies is complex, due to challenges including the absence of universal definitions surrounding AI and the explosion of potential AI use cases. The family of related AI technologies presents societal risks that require various levels of oversight and a nuanced approach to standard-setting.
A series of workshop sessions co-organized by the Center for Security and Emerging Technology and the Center for a New American Security in the fall of 2022 examined case studies of previous standards development across several industries to draw lessons for AI. These discussions highlighted the challenges of developing robust, effective standards as well as best practices that have enhanced standards development and enforcement processes in the domains we studied. The workshop and many of its recommendations were completed before the October 30, 2023 release of the Executive Order on Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. The recommendations are consistent with the Executive Order and frequently provide details and specificity for implementing it.
Our key findings are:
- Finding 1: AI risk assessment and mitigation should include examining how interdependencies affect systemic risk.
- Recommendation 1: Critical infrastructure owners and operators should track the interdependencies of their AI systems.
- Recommendation 2: The Office of Management and Budget (OMB) and the Office of Science and Technology Policy (OSTP) Directors’ forthcoming guidance on minimum risk management practices for AI should require that agencies identify the risks that could emerge from interdependencies between their AI systems and other entities.
- Finding 2: Guidance on testing and re-approval of AI systems should be calibrated to risk and account for changes to AI systems over time.
- Recommendation 3: The U.S. Department of Defense should create thresholds or triggers for different levels of rigor and oversight for testing military AI systems.
- Recommendation 4: U.S. government agencies should establish processes for the reassessment and re-testing of systems as they change over time and share these processes with each other.
- Finding 3: Compliance assistance can help small- and medium-sized businesses prepare for and implement AI regulation.
- Recommendation 5: Congress should create a pilot AI Compliance Assistance Office within the U.S. Department of Commerce, which should later expand to other government agencies.
- Finding 4: Third-party organizations can remove barriers to standards development, implementation, compliance, and tracking.
- Recommendation 6: OMB should direct a study by an independent body to inform the designation of third-party accreditation bodies that ensure certifiers evaluate the implementation of AI standards in a consistent manner.
- Recommendation 7: Professional organizations should establish AI standards access funds, whistleblower protection programs, and reporting programs to gather anonymized information on AI risks from industry participants.
- Finding 5: Non-regulatory governance is one mechanism that can support the safe development and use of AI systems.
- Recommendation 8: The United States should commence discussions in the G7 about creating the equivalent of a Financial Action Task Force for AI.
- Recommendation 9: NIST should create an online portal to ensure technical developments relevant to standards are captured and publicized.
- Finding 6: Coordination and regular efficacy checks of standards can ensure that standards development is efficient and effective.
- Recommendation 10: Standard-setting bodies should host biannual summits to coordinate on standards interoperability and efficacy.
- Recommendation 11: NIST should support the development of testbeds to monitor AI standards for effectiveness.