Definition
Representation entanglement is the phenomenon where a model's internal representation of one concept is causally intertwined with representations of other, seemingly unrelated concepts, so that intervening on one shifts the others. It matters for interpretability and alignment because it shows that targeted edits—like training a model to stop claiming consciousness—can have diffuse side effects, altering the model's expressed beliefs about minds, spirits, or even its mood in ways the intervention never intended.