Research & Fellowships
Research Focus
Our research is anchored in three areas: how AI systems interact with human autonomy, how they shape the conditions under which knowledge is produced, and how institutional and social arrangements facilitate self-governance.
We work closely with Oxford University's Human-Centered AI Lab as well as practitioners at the frontier of AI, leading academics, and thinkers on AI & society.
Research Team and Collaborators
Cosmos Fellowship
The Cosmos Fellowship identifies and supports individuals with the potential for world-class AI expertise and deep philosophic insight.
Previous Fellows include research leads from OpenAI and Google AI, polymathic bloggers, education and autonomy experts, and top computer science and philosophy graduates. The 2025 Fellowship cohort produced NeurIPS papers, open-source software, and industry-supported research on AI and human autonomy.
We encourage applications from both experienced researchers, and promising junior talent.
Typical award
USD $75,000 per annum (pro rata). Travel expenses covered. Additional compute available.
Community access:
Connect with a vetted network of builders and thinkers at the AI × philosophy frontier.
Typical duration
Up to one year, with the possibility of extension.
Application window
Not currently open, but accepting expressions of interest here. For details on Fellowships with HAI lab see their webpage.
Current Fellow Projects
Tianyi Alex Qiu is a safety and alignment researcher focused on truth-seeking and moral progress. He built ProgressGym, an open-source framework for studying moral progress in AI systems, downloaded by over 20,000 researchers. He is currently a fellow at the Foresight Institute and previously worked on frontier safety and alignment at Anthropic and at the UC Berkeley Center for Human-Compatible AI. He won the Best Paper Award at ACL 2025 and saw his work spotlighted at NeurIPS 2024.
At Cosmos, his work focuses on building tools that make AI models internally coherent, able to detect and resolve contradictions in their own reasoning, so they can help users do the same with their own beliefs.
