Value alignment
Aligning models with human values and intent
Generalization
Ensuring goals generalize as intended
Interpretability
Interpreting what happens inside models
Scalable oversight
Supervising systems smarter than us
Evaluations
Evaluating dangerous capabilities
AI control
Controlling models that may be misaligned
AI will change the world as we know it.
We’re a community of students and faculty at Cornell working toward reducing risks and improving the trajectory of AI development.
Our mission
Managing risks from advanced AI is one of the most important challenges of our time.
We bring together technical AI safety, policy, governance, and broader conversations about AI to ensure that increasingly capable AI systems remain aligned with human intentions and benefit all of humanity.
AI Safety
Regulation
Designing regulation and safety standards
Compute governance
Governing compute and frontier development
International coordination
Coordinating on AI across nations
Economic impacts
Anticipating effects on work and the economy
Power concentration
Preventing extreme concentrations of power
Epistemics
Protecting human agency and public discourse
Theory of Change
01
Outreach
02
Capacity Building
03
Impact
Research
Careers
- Frontier AI labs
- AI safety companies
- Non-profits
- Think tanks
- Government
Upcoming Events
We host talks, reading groups, workshops, and socials throughout the semester, and anyone is welcome to come.
News
Recent updates from CAIA, including events, papers, and announcements.
Event ·
AI safety student roundtable in Washington, DC
CAIA members attended an AI safety student roundtable in Washington, DC, organized by the Center for AI Safety. We met with AI safety organizers and student government leaders from other universities.
Event ·
CAIA open house & mocktail mixer
We welcomed new and returning members for mocktails and conversation at our open house. Thank you to everyone who came to meet the community and learn about getting involved in AI safety at Cornell.
Event ·
Guest talk with Prof. Paul Gölz
Prof. Paul Gölz joined CAIA to present Distortion of AI Alignment: Does Preference Optimization Optimize for Preferences? The talk explored how preference optimization can distort diverse human preferences and what this means for pluralistic AI alignment. Thank you to Prof. Gölz and everyone who joined the discussion.
Research ·
CAIA papers accepted to ICML 2026 workshops
CAIA members contributed to multiple papers accepted to ICML 2026 workshops, spanning pluralistic alignment, trustworthy AI, and mechanistic interpretability.
Research ·
EigenBench selected for an ICLR 2026 oral presentation
EigenBench was accepted to ICLR 2026 and selected for an Oral presentation.
Announcement ·
New York signs the RAISE Act into law
Governor Kathy Hochul signed the RAISE Act into law, establishing nation-leading AI safety requirements for frontier model developers. CAIA helped canvass support for the bill.
Event ·
AI safety student roundtable in Washington, DC
CAIA members attended an AI safety student roundtable in Washington, DC, organized by the Center for AI Safety. We met with AI safety organizers and student government leaders from other universities.
Event ·
CAIA open house & mocktail mixer
We welcomed new and returning members for mocktails and conversation at our open house. Thank you to everyone who came to meet the community and learn about getting involved in AI safety at Cornell.
Event ·
Guest talk with Prof. Paul Gölz
Prof. Paul Gölz joined CAIA to present Distortion of AI Alignment: Does Preference Optimization Optimize for Preferences? The talk explored how preference optimization can distort diverse human preferences and what this means for pluralistic AI alignment. Thank you to Prof. Gölz and everyone who joined the discussion.
Recent Work
A few recent papers that CAIA members have worked on.
CAIA coauthorsPreprint 2026
CogGym: Towards Large-Scale Comparative Evaluation of Human and Machine Cognition
CAIA coauthors: Jinzhou Wu
Survey preprint 2026
Math for AI safety: an invitation for mathematicians
CAIA coauthors: Lionel Levine
COLM 2026 Workshop
Stubborn or Sycophantic? GEPA-Evolved Prompts Under Pressure
CAIA coauthors: Ashton Chew
ICML 2026 Workshop
WARP: Measuring and Mitigating Evaluation Awareness in Browser-Agent Safety Benchmarks
CAIA coauthors: Jasmine Li, Ashton Chew
ICML 2026 Workshop
Side Effects of Character Training: Quantifying Cross-Constitution Drift in LLMs
CAIA coauthors: Jonathn Chang, Lionel Levine
EvalEval 2026
Measuring AI-Induced Disempowerment: A Framework and Proposed Metrics
CAIA coauthors: Jasmine Li
Our members have worked with
Join us
Open to Everyone
All of our events are open to everyone. If you join our Slack, you’re considered a member. It’s where we share conversations, announcements, and opportunities.