LLMs in near-twin prisoners' dilemmas reveal no consensus on interpreting FDT
OrionJohnston · x · 2026-11-10
The author argues that CDT as defined by Lewis is just a maximization rule where you bring your own mixture of dependency hypotheses; causal variants of FDT are the same rule plus informal guidance on which acts and dependency hypotheses to consider. So "functional CDT" would be a more accurate label than "causal FDT".
The dependency hypotheses in the FDT family are not "logical" in any meaningful sense: you pick out parts of the environment that depend on your policy and act as if you choose the policy's input/output, but nothing tells you which parts those are beyond "I know it when I see it" — a problem shared by every decision theory.
The author also ran an experiment placing LLMs in simulated near-twin prisoners' dilemmas and instructing them to apply different decision theories. The results show a lack of consensus on how to interpret action-based FDT as defined in the FDT paper, equation 4.