It's literally the explanation. The only identification we have now is "this local part is often used by an AI model" and "this global structure is often used b

It's literally the explanation. The only identification we have now is "this local part is often used by an AI model" and "this global structure is often used by an AI model". There's nothing more fancy about it. The heatmap would literally just point out "this part is suspiciously unlikely" - that's the explanation because that's the classification systems use.