Cecilia Elena Tilli

I work in AI safety, and I’m now launching a new research initiative on which propensities and character traits make AI agents safe to deploy in complex multi-agent settings, supported by Rethink Priorities and Cooperative AI Foundation. For the past three years I've been at the Cooperative AI Foundation, running grantmaking and leading workshops, and most recently built evaluations for agent-agent influence capabilities. Before that I founded a vegan cheese company, worked in strategy consulting for research-heavy startups, led a foundation focused on preventing antibiotic resistance, and did postgraduate research in scientific computing at Uppsala.

Selected Outputs

• Agent Properties for Multi-Agent Safety

  Cooperation failures between AI agents could lead to catastrophic outcomes even when agents are individually safe. I propose studying specific agent properties as a tractable path toward safer multi-agent interactions.

• Video: Properties of Cooperative Agents

  A talk I gave at the Cooperative AI Summer School on what it means for an agent to be "cooperative". I also present some experiments done in collaboration with AI Digest to test model behaviour in mixed-motive interactions.

• AI-Facilitated Human Cooperation: What Would Success Look Like?

  Insights from the Cooperative AI Summer Retreat on how AI could help resolve major cooperation challenges in democratic processes. We explored what successful integration of AI into governance might look like and identify promising testbeds.

• Updates to the CAIF Grant Program in 2025

  Changes that I implemented to the CAIF grantmaking program based on learnings from 2024. We've introduced more specific research areas, an early-career track, and adjusted our review process to increase impact and support promising researchers.

• Cooperative AI: Three Things That Confused Me as a Beginner

  While AI alignment deals with the question of how to make one powerful AI system behave in a way that is aligned with (good) human values, cooperative AI is about making things go well with powerful AI systems in a messy world where there might be many different AI systems, lots of different humans and human groups and different sets of (sometimes contradictory) values.

• Why Scientific Research Is Less Effective in Producing Value Than It Could Be: A Mapping

  A systematic mapping of why academic research often fails to produce value. I identify root causes including poor choice of research questions, reproducibility issues, and lack of connection with end-users.

• Improving Science: Influencing the Direction of Research and the Choice of Research Questions

  Influencing the direction of research seems more pressing, and more neglected, to me than working on improving other aspects of the research system. Influencing research policy seems like a potentially valuable and rather unexplored area.

• Antibiotic Resistance and Meat: Why We Should Be Careful in Assigning Blame

  The link between meat production and antibiotic resistance is real but uncertain in magnitude. We should base our advocacy on what the evidence actually supports, and while there are many good reasons to criticize industrial animal agriculture the evidence for its contributions to antibiotic resistance doesn't seem that strong.

• Post-mortem: Ultuna mejeri AB

  Lessons learned from my vegan cheese startup Ultuna Mejeri.