Christoph Salge

I build agents and content generators that adapt to their surroundings and to the people they play or create with. Much of this work runs through GDMC, where algorithms build whole settlements on terrain they have never seen. That raises a hard question for any open-ended generative system: how do you evaluate it when there is no single right answer?

Instead of writing rules for AI systems, I look for intrinsic motivations that produce safe behaviour on their own: behavioural primitives that keep a human’s options open while preserving the agent’s own capacity to act. Empowerment, the channel capacity from an agent’s actions to its future sensor states, gives one task-independent way to measure this. I also study how learning agents fail, using games built to mislead them.

Different motivations produce different kinds of social behaviour. I study which motivations make agents interact with, support or oppose each other, how people perceive agents driven by them, and how norms, institutions and legal responsibility take shape in a society that includes AI.