Harvard students working to reduce catastrophic risks from advanced AI
The AI Safety Team (AIST) is a community of technical and policy researchers at Harvard. Spring 2027 fellowship applications open soon; get notified.
We’re on track to develop immensely powerful AI soon, and we are woefully unprepared for its arrival.
Without adequate safeguards, powerful AI could autonomously escape our control, help bad actors build weapons, or slowly take important decisions out of human hands. We research these problems and help students start working on them.
We run nine-week intro fellowships
Three tracks of our introductory reading groups each semester. Cohorts meet once a week for two hours at our office in Harvard Square.
Technical Fellowship
Why frontier AI models are hard to control, and the research on making them behave safely.
Basic machine learning
Curriculum and FAQ for the Technical Fellowship
Policy Fellowship
The governance questions advanced AI raises, and the policies that could keep it safe.
Any background
Curriculum and FAQ for the Policy Fellowship
AI × Biorisk Fellowship
What new AI tools mean for biosecurity, and the safeguards that could prevent their misuse.
Life-science background
Curriculum and FAQ for the AI × Biorisk FellowshipIn a survey of 1,580 AI researchers, most think there's a >10% chance of AI destroying or disempowering humanity, and the top AI forecasters believe we will have vastly superhuman AI in the next decade.
AI Futures Project, Q2.5 2026 Timelines Update (2026)
AIST members have worked with or gone on to
After the fellowship
Membership
After a fellowship, Harvard students can apply to become members. Members join a technical, policy or strategy track and facilitate future cohorts.
Workshops
Each semester, AIST and MIT AI Alignment (MAIA) host weekend workshops where students meet professionals working in AI safety from the top labs and nonprofits.
Speaker events
We also regularly host speaker events: some are restricted to members, while others are open to anyone at Harvard.
Research from our members
Representational Simplicity and Circuit Size Dissociate in a Threshold-Dependent Way: A Controlled Test via Adversarial TrainingAdam Elimadi · InterpretabilitySeptember 27, 2026
Tracing mechanisms of sycophantic agreement in language modelsLogan Riggs Smith and 4 coauthors · InterpretabilitySeptember 21, 2026
A Transformative AI Strategy for EuropeConor McGlynn (research lead) · GovernanceSeptember 14, 2026
Spring 2027 fellowship apps open soon
Leave your email and we'll notify you when applications release.