Topics in Technical AI Governance

My proposed DPhil research centres around 'Technical AI Governance', that is, technical analysis and tools for supporting the effective governance of artificial intelligence. This project falls within the EPSRC 'Artificial intelligence technologies' research area.

The increasing adoption of artificial intelligence (AI) has prompted governance actions from the public sector, academia, civil society, and industry. These include 'hard' regulation targeted at AI systems, such as in the case of the EU's AI Act, as well as 'softer' governance through norms and voluntary commitments, such as those agreed to by AI developers at the AI Seoul Summit in May of this year.

However, policymakers focussing on AI often have insufficient information for identifying the need for intervention and assessing the efficacy of potential policy options. Furthermore, the technical tools necessary for successfully implementing policy proposals are often lacking. The emerging field of technical AI governance seeks to address these challenges.

The desired impact of my research is thus to advance effective governance and policy regarding AI, through the development of targeted technical engineering insights, methods, and infrastructure. Alongside my own research, I aim to enable other computer science and engineering researchers to work on policy-relevant topics through outreach and facilitating connections between the technical academic community and relevant policy-makers.

While the details of my proposed research have not been fully decided, I am particularly interested in researching and developing methods to allow for increased public oversight into the development of state-of-the-art AI systems. This could be enabled through the development of technical infrastructure that enables external researchers or auditors to access proprietary systems and data for the purpose of assessment and evaluation, without exposing sensitive information.

While infrastructure for privacy-preserving researcher access to data has been developed in other industries, for example health or census data, no such examples have been developed specifically for access to AI systems. Thus, developing a proof-of-concept implementation of bespoke AI researcher access infrastructure will be a novel contribution. Trust between auditor and system owner could further be bolstered through the use of verification mechanisms such as Zero-Knowledge Proofs that can attest to the actions taken by external researchers and auditors, without revealing specifics of the actions taken.

Exploring how this could be done in practice is another avenue of research that I would be interested in pursuing, building on the limited literature appling zero-knowledge proofs to AI systems.

While no formal collaboration arrangements have been made, my prior work on this topic has involved a broad range of collaborators from academia (including from Stanford University, MIT, and the University of Cambridge), industry (including from HuggingFace and Cohere for AI), and civil society (including from the Centre for the Governance of AI), and similar collaborations are possible moving forwards. Given my current and prior affiliations with both the Centre of the Governance of AI and the Oxford Martin AI Governance Initiative, continued collaboration with these institutions also seems likely.