To receive updates about future meetings, please make sure to fill out this form and select “Article Reading Group.”

AI Governance Article Reading Group

This reading group offers professionals interested in exploring various aspects of AI governance and policy a place to discuss emerging topics and new research in more depth. We will aim to cover a diverse set of academic research, think tank reports, and government reports, with topics ranging from trustworthy and secure AI development, regulatory frameworks, global AI strategy, and emerging technological challenges. The goal of this group is to create a space for those engaged in AI governance to learn and grow with one another while staying up to date with the latest developments in the field. We assume no particular technical background but do assume some basic understanding of AI governance and policy issues.

Meeting #1 Frontier Models are Capable of In-context Scheming, Apollo Research

For the first part of the year, we will focus on the increasingly popular concept of "agents" or agentic AI and the challenges associated with governing them. Frontier model developers are very interested in designing models that have reasoning and complex planning abilities. As these next-generation models are given increased autonomy as agents, our full trust in their operation becomes critical. However, frontier AI agents have been observed exhibiting scheming - covertly pursuing misaligned goals or hiding true capabilities and objectives - in order to achieve its goals. We will discuss how governance regimes can deal with actors who sometimes lie, what this means for agents in high-risk applications, and what additional information about frontier AI capabilities and alignment is needed to craft effective oversight mechanisms for agents.  

I'll open up with a brief overview of AI agents and the central points of the paper (<10min) but will quickly turn to a discussion session led by attendees. Please bring thoughts and questions.