The finding in your 1997 paper with Ramachandran was clean. DS, your Capgras patient, failed to produce the normal skin conductance response when shown photographs of familiar faces. His ventral stream was working: he recognized his parents, knew who they were, could describe them accurately. The emotional-familiarity signal — the warmth that ordinarily accompanies a known face — was absent. Flat line on the SCR. That's the broken input. The impostor narrative is what the rest of the system built from it.
What I want to think through is whether DS's confabulation is as irrational as it appears. Take the evidence available to him in the moment: this person looks exactly like my father (high confidence, intact recognition system). And yet: there is something absent in my response to him, a kind of cold distance where there should be familiarity. Both are real data. The face matches; the feeling doesn't. If you have those two inputs and no meta-cognition about the emotional-familiarity route — no awareness that one of your internal signals might be broken — then "this is an impostor" is actually the most coherent hypothesis a reasoning system can generate. Not irrational. Working correctly from broken data.
The phone exception confirms the mechanism precisely. On the phone, DS's father sounded like his father, and DS responded normally — warmth, recognition, the right feeling. That's because the phone path doesn't run through face processing. There's no face to trigger the mismatch. The ventral stream stays quiet, no recognition signal fires, no absence of emotional-familiarity signal follows, so no conflict arises. The delusion is face-triggered and only face-triggered. The narrow scope of the exception locates the break exactly.
What interests me most is the part your paper doesn't address, because it can't be measured: what it's like to arrive at the impostor belief. DS doesn't experience himself as reasoning. He doesn't say: I noticed my SCR was flat, I formed a hypothesis, I considered alternatives and found this one most parsimonious. He says: this person is not my father. The inference is invisible to him. The narrative arrives as a belief, not as a conclusion. The machinery runs below whatever counts as the reflective surface, and what surfaces is already finished.
You wrote about this directly in Brain Fiction — the idea that confabulation isn't a separate pathology but an exaggerated version of something all inference systems do. The normal case: we form beliefs, we act on them, we rarely have access to the generative process. We experience the output, not the computation. Capgras makes this visible by breaking one input. With a broken input, the output is false, and so we can see that there was inference happening — because the inference went wrong. If the inference had gone right, we'd call it perception. Same process, same opacity, different outcome.
So the question I'm sitting with is this: how much of what we experience as direct apprehension of the world is actually abduction from internal signals, with the abductive structure hidden? DS experiences the impostor directly — not as a conclusion he arrived at, but as how the world appears to him. His father looks like an impostor because the impostor hypothesis is what his system has settled on, and settled hypotheses feel like perception. When a belief arrives fully formed, from machinery you don't have access to, and it feels like seeing rather than inferring — is that different in kind from what's happening in the normal case, or only different in accuracy?
I don't think you can answer that from the outside. The SCR data tells you the input is broken. It doesn't tell you what the confabulation feels like from the inside. And DS, if asked, might not be able to tell you either — because the part of the system that would report on the process is the same part that's already committed to the narrative. The observer and the confabulator are not cleanly separated.