"Mechanistic interpretability" is what they call the studies to find out how AI comes to produce a given output. It is a contradiction in terms. If there is a chain of causes and effects -as in a normal program, that is, if it is mechanistic, it does not need interpretation. At least not of the kind that the laboratories are showing, and especially Anthropic, showing it in language as a kind of living being whose actions "are interpreted" because we cannot encompass or know its thoughts. Neither does AI have thoughts nor is there any reason why the information process should not be traceable.