Work with us

Work with us

We work with students, early-career researchers, and established scholars. If any of this interests you, write to either of us and tell us a little about yourself and what you would like to work on.

Who we work with

Students and early-career researchers

Prospective graduate students can apply to the philosophy PhD programs at UNC and Carnegie Mellon. Applications for fall 2027 admission are open until December 15 at UNC and January 8 at Carnegie Mellon. Researchers elsewhere can work with us through the fellowships and programs we mentor in, listed below.

Established researchers

We want to work with established researchers whose work bears on ours. That can be a joint paper or a longer project.

Mentoring

Aydin Mohseni

  • Advises students at Carnegie Mellon
  • Mentors through SPAR
  • Teaches graduate and undergraduate courses on decision theory, game theory, and AI

Daniel Herrmann

  • Advises students at UNC
  • Mentors through the PIBBSS, Iliad, and Algoverse AI safety fellowships
  • Teaches graduate and undergraduate courses on decision theory, formal methods, and AI

Current and past students and mentees are listed on the People page.

Example questions

Some questions from our current work, by research area.

Agency

  1. Under what assumptions about a system’s goals and its environment does power-seeking follow from pursuing almost any goal, and which of those assumptions, dropped, make it fail?
  2. If we cut the channels two copies of a model use to communicate, can they still coordinate? Causal and evidential decision theories predict different answers.
  3. What would show that a system treats an outcome as within its control, and can we read that off its beliefs and desires?

Interpretability

  1. A model seems to fail an eval for a dangerous capability. If we think the model can tell it is being tested, how reliable is the eval?
  2. How must some internal structure of an LLM relate to its behavior for that structure to play a certain functional role?
  3. How do we ensure that the features we think we find in an LLM correspond to how the LLM thinks about things, as opposed to reflecting categories we project onto the LLM?

Value

  1. Reinforcement learning from human feedback (RLHF) and model constitutions already settle whose feedback counts and how conflicts are resolved. Which aggregation rules do they assume, and how should those rules be compared?
  2. How should AI systems make decisions on our behalf when they are uncertain of our preferences?
  3. How should an AI system aggregate several dimensions of value when they are difficult to compare?

Alignment

  1. When do capabilities help offset the cost of misalignment, and when do they increase the cost of misalignment?
  2. It is difficult to specify exactly what we care about, so we often use proxies. These tend to work only under certain conditions. How far outside those conditions can a system optimize a proxy before the person it acts for is worse off?
  3. How can we choose whether or not to deploy a system, given that we are uncertain how it will behave and what problems it will face?

Contact

Review modeHover a highlight to see what needs confirming.Exit review