Theory of Mind / Empathy Dissociation

In humans, inferring what someone believes (cognitive theory-of-mind) and being moved by their suffering (affective empathy) are separable systems — different brain networks, doubly dissociable in the clinic: psychopathy leaves ToM roughly intact while empathy drops, some presentations of autism show the reverse pattern. AI safety mostly treats “understanding others” as one undifferentiated thing.

AE Studio's self-other-overlap method pulls a single dial toward “more self/other overlap = more honest.” The neuroscience predicts this could be conflating two systems that don't move together — and could even damage theory-of-mind rather than improve alignment, since welding self and other representations together is exactly the kind of thing the rTPJ self/other-switching literature would expect to interfere with belief-tracking.

Four pre-registered questions: are a fitted ToM direction and empathy direction geometrically distinct? Does ablating one selectively hurt its matching task — ToM accuracy vs. empathetic-response quality — while leaving the other intact, a causal double dissociation? Which ablation reduces deception more on a burglar-style scenario? And does reproducing AE Studio's self-other-overlap axis and ablating it impair ToM accuracy in a way their own evaluation — which only measures deception rate — would never surface?

Grew out of a “neglected approach” essay written for an AIAF Fellowship application; now pursued as independent research regardless of that outcome.