How the decision theory of NCLP differs between humans and machines

• Classical decision theory assumes humans but AIs have different capabilities and different situation that they find themselves in.
• Evidential decision theory would cause cooperation with AI that share the same source code (and data...).
• CDT is harder (than EDT) to formalize enough for machines to follow it.
  • Inferring the causal graph is hard, sometimes unclear how to do it.
  • CDT is more sensitive to choosing the right abstractions (and if you messed it up, the consequences are worse than for EDT).
  • Causality itself is complicated and not completely solved.
• Tickle defence might work better for AIs than for humans, so EDT + tickle defence = Reflective EDT would be a better decision theory.
• If AIs are very similar, when they get into a prisoner's dilemma, it becomes a NCLP.
  • Even if they are not, they can simulate each other.