Stephen Casper
This stream will focus on impact-oriented technical AI governance research work, potentially including research on open-weight models, applied AI safeguards research, AI incidents, technically rigorous AI policy, etc.
Stream overview
Our stream projects will generally focus on a few types of topics:
- Open-weight model safeguards: working to make AI systems with publicly downloadable weights more resistant to misuse, including by making them more tamper-resistant.
- Applied AI safeguards research: studying if and how AI safeguards are applied in the real world, and analyzing the connections that exist between company choices and downstream consequences.
- AI incidents: studying AI incidents and how they could be prevented.
- Technical rigor of AI policy: auditing laws for technical ambiguities, challenges, and loopholes.
- Miscellaneous AI governance research: guerrilla-style research to help policymakers make informed choices about emerging challenges in AI.
Mentors
Stephen "Cas" Casper is a computer scientist and an Assistant Professor of Public Policy at the Harvard Kennedy School and a Faculty Affiliate of the Harvard School of Engineering and Applied Sciences. Prior to joining Harvard, he completed his PhD at MIT and did a research residency with the UK AI Security Institute. He is a writer for the International AI Safety Report and a lead writer for the Singapore Consensus. His research has been recognized with a Hoopes Prize, an ML Safety Workshop best paper award, a BioSafeGenAI best paper runner-up, a GenLaw spotlight paper award, a TMLR outstanding paper finalist distinction, and a handful of mentions in news articles and newsletters. Find him on Google Scholar, Twitter (sorry), BlueSky, and LinkedIn.
Mentorship style
These projects in this stream will likely follow a certain default research process:
- Pay attention to and discuss contemporary discussions, debates, and proposals related to AI governance.
- Get disappointed, confused, or frustrated.
- Write a technical paper to improve the discussion.
- Spend a lot of time and effort communicating it to the audience that needs it.
By default, we should expect to meet 2-3 times per week as a full group, plus ad hoc project-specific meetings.
Fellows we are looking for
Green flags include:
- Research tenacity: demonstrated ability to pursue self-directed work, make things happen through determination, teach oneself whatever skills are required for a project, and succeed even when not set up to succeed. As an example, I think it is a strong green flag when an undergrad pursues side projects in a self-directed manner rather than only pursuing projects under classes, internships, jobs, etc.
- Research taste: the AI research space is noisier than ever, and almost all AI research has little to no practical value. Putting impact over interest and designing projects around a specific plan for impact is the most important single skill needed for good AI work.
- Experience across most or all of the full project stack: ideating, planning, experimenting, writing, and publishing.
This stream will follow an academic collaboration model. Scholars will be free to discuss and collaborate externally. However, scholars should also expect to work in collaboration with others in the stream.
Project selection
I will work with MATS scholars to iteratively refine project ideas in whatever area our interests and skills overlap. Above all, project selection will hinge on having a clear (and good) theory of impact.
The Winter 2026 cohort offers a wide range of research streams led by experts across AI alignment, interpretability, governance, and safety. Each stream provides its own research agenda, methodology, and mentorship focus.
