Open Problems in Technical AI Governance
The broad technical-governance catalogue. Focus on compute, verification, and auditing sections to see how the verification problem sits among adjacent open problems.
→ arXiv:2407.14981What people are pursuing
Several research agendas have crystallised around AI verification: frontier auditing, hardware-enabled mechanisms, cloud governance, international agreement design, and broad technical-governance taxonomies. This unit gives you the map before the mechanisms.
By the end of the unit, you should be able to place a new verification paper inside the agenda it contributes to and understand what problem family it is trying to solve.
Hardware-Enabled Mechanisms for Verifying Responsible AI Development
The hardware-rooted taxonomy: location verification, network verification, offline licensing, and workload verification. This is the core framing for hardware-enabled AI governance.
→ arXiv:2505.03742Frontier AI Auditing
A detailed agenda for rigorous third-party assessment of frontier AI labs. Read the verification-relevant sections as the auditing counterpart to the hardware agenda.
→ arXiv:2601.11699Mechanisms to Verify International Agreements About AI Development
A survey of mechanisms for verifying international agreements about AI development. Useful for understanding the trust problem between mutually suspicious parties.
→ arXiv:2506.15867Verification for International AI Governance
The Oxford Martin framing of international AI verification. Read alongside RAND and FLI-style proposals to compare institutional assumptions.
Available via Oxford Martin AI Governance Initiative.Verifying International Agreements on AI
The RAND treatment of verification for international AI agreements. Useful for comparing policy design constraints with the more technical agendas.
Available via rand.org.Hardware-Level Governance of AI Compute
A recent feasibility taxonomy connecting hardware capabilities to regulatory compliance and treaty verification. Read for a synthesis of what hardware can plausibly support.
→ arXiv:2604.04712