Understanding agency as a science.
SOAR Initiative is a research group studying agency as a scientific object — what it is, how it can be measured, and what follows for increasingly capable AI systems.
What we do
Agency is invoked everywhere and defined almost nowhere
Claims about AI systems acting as agents carry enormous weight in safety arguments, governance proposals and capability forecasts. Yet the underlying notion is rarely made precise enough to measure, and almost never precise enough to falsify.
We work on the foundations: formal accounts of agency, measurement that survives contact with real systems, and the implications for how advanced AI should be evaluated and governed.
Research directions
Three lines of work
Formal foundations
Precise accounts of goal-directedness, coherence and optimisation that are strong enough to make predictions rather than merely describe.
Measurement
Turning those accounts into quantities that can actually be estimated from the behaviour and internals of deployed systems.
Implications
What a working science of agency changes about evaluation, oversight and the governance of advanced AI.
Working with us
We are early, and interested in collaborators, critics and people who think this framing is wrong in an interesting way.
Recent writing