Alignment Problems in AI Governance

Simons Institute Law and Society fellows Rui-Jie Yew and Gregory Demirchyan explore two challenges of alignment in AI governance.

First, we presently lack a thorough understanding of these models, making it difficult to identify the risks they pose. Furthermore, existing auditing tools may not be sufficiently reliable to offer assurances of model safety and alignment. Even when explanations appear compelling, establishing their faithfulness remains hard to achieve, especially at scale. These technical challenges present regulatory difficulties for current governance proposals and can expose the misalignment between proposed interventions and their feasibility in achieving their regulatory objectives.

Secondly, another challenge lies in the unintended effects of how AI systems might be designed, deployed, and framed to minimize regulatory costs. While methods in the technological tool kit of safety, like privacy-preserving technologies and AI evaluations, are framed as safety enhancing, they also simultaneously shape the terms of regulatory oversight for AI systems and can be developed and deployed in misalignment with the goals of regulation.

Yet, despite these challenges, it is critical to continue to develop governance mechanisms. The speakers’ aims in this talk are to discuss approaches to governance that attempt to reduce these tensions by (1) designing regulatory systems that are sufficiently adaptable as our understanding of this transformative technology improves and (2) presenting steps toward the robust oversight of AI systems.

Rui-Jie Yew is a PhD student in computer science at Brown University, where she is affiliated with the Center for Technological Responsibility, Reimagination, and Redesign (CNTR). Her research focuses on the intersection of artificial intelligence, policy, and law. Her work has been recognized at venues including FAccT, AIES, and the ACM Symposium on Computer Science and Law, including with a Best Student Paper Runner-Up Award at AIES 2024. Previously, she earned an SM in technology and policy from MIT and a joint BA in computer science and mathematics from Harvey Mudd College and Scripps College. She has contributed to AI policy research and privacy technology development at NIST, Sony AI, and Google. She has also been on the program committee for workshops at AAAI and NeurIPS, and she has organized for MIT’s AI Ethics & Policy group and Brown CNTR’s Graduate Seminar on Technology and Society.

Greg Demirchyan is an attorney specializing in technology law. He is the CEO of a consulting firm that advises companies on AI governance, responsible AI principles, and the emerging landscape of regulatory requirements. Prior to this, he was a partner at a litigation firm in San Francisco and served as general counsel for several tech start-ups. Demirchyan also worked internationally as lead advisor for the American Bar Association on rule of law and anti-corruption. He is a member of several technology advisory committees and working groups, including the norm forecasting group at the Center for Cyber-Social Dynamics at the University of Kansas. His current research focuses on algorithmic fairness, safety, and AI alignment. In addition to his JD, he holds a PhD in moral and political philosophy from Cornell University.

Light refreshments will be available at 3 p.m., prior to the start of the lecture. 
_______________________________________________________________________________________

Established in 2020, the Simons Institute’s Law and Society Fellowships enhance Institute programs that address technologies with profound impacts on human society and with implications for ethics, law, and policy, by supporting a researcher within each who is focused on addressing the broader societal implications of the techniques and technologies addressed within these programs. 

The Simons Institute regularly captures photos and video of activity around the Institute for use in publications and promotional materials. 

If you require special accommodation, please contact our access coordinator at simonsevents@berkeley.edu with as much advance notice as possible.