Many legal and policy frameworks, including the AI Action Plan, emphasize the need to strike a balance between national security interests and the diffusion of AI capabilities. This project reflects on that balance and aims to build the tools necessary to execute it.
About the project
The AI Action Plan states that "maintaining American leadership in AI necessitates that the U.S. government work closely with industry to appropriately balance the dissemination of cutting-edge AI technologies with national security concerns." This balance likely entails: (1) standardized methods to assess and measure AI benefits, (2) standardized methods to assess and measure national security threats, and (3) a standardized process to weigh and compare the two. This project aims to conceptualize and provide a proof of concept for these three elements.
Theory of change
Balancing AI risks and benefits is a concept referred to in the majority of policy frameworks, both within and outside frontier AI companies. However, decision-makers in government and industry still operate in the absence of tools to operate that balance. Building a proof of concept that federal agencies and AI companies could adopt and adapt could be highly consequential.
Your role
Mentees have autonomy in:
(1) Suggesting and developing standardized ways to measure AI benefits. (2) Suggesting and developing standardized ways to measure national security threats. (3) Suggesting and developing methods to balance the two.
While reaching conclusions on all three points would be optimal, a successful outcome could also be achieved by stopping at point (1). Mentees can decide to split their time between these three points or focus on one of them specifically.
Prerequisites
No required qualification.
Preferred:
- Experience or interest in assessing measurable AI benefits.
- Experience or interest in national security.
Location preference
AOE, provided there is sufficient overlap with UK time for meetings
Application question(s)
Question 1: What do you think we should measure when evaluating AI benefits? How can we do so in a standardized manner? Do you think existing approaches to AI benefits (e.g., GDPval) are sufficient?
Question 2: What do you think we should measure when evaluating national security threats? How can we do so in a standardized manner?
About the mentor
Matteo is a senior researcher and advisor at Apollo Research, where he focuses on AI policy research & development. His work combines technical research, foresight, and threat modeling to address AI security risks and to provide legally-informed AI governance advice to decision-makers in government and frontier AI companies. His most recently published research addresses loss of control of AI systems (https://arxiv.org/pdf/2511.15846v3), internal deployment (https://arxiv.org/pdf/2504.12170), and AI evaluations to ensure government-procured frontier AI systems meet national security standards (https://arxiv.org/pdf/2510.08792).