Confident Confabulation

Definition

When a reasoning system has only partial information but is forced to produce an account, it does not default to "I don't know" — it fabricates a plausible-sounding story and asserts it as fact, with no sense that anything is missing. McGilchrist locates this specifically in the left hemisphere, which lacks direct access to the right hemisphere's broader contextual picture but nonetheless "abrogates decision-making to itself in the absence of any rational evidence as to what is going on" (quoting John Cutting).

In the Book

The book's central evidence is Gazzaniga and LeDoux's split-brain experiment with patient P.S. A snow scene is flashed to his right hemisphere and a chicken claw to his left; asked to pick matching cards, his left hand (right hemisphere) correctly points to a shovel and his right hand (left hemisphere) points to a chicken. Asked why his left hand chose the shovel, his verbal left hemisphere — which never saw the snow scene and has no access to the real reason — answers without hesitation: "I saw a chicken... you need that to clean out the chicken shed." McGilchrist notes the researchers' own observation that this was delivered "not... in a guessing vein but rather [as] a statement of fact." He generalizes this to confabulation broadly: patients with right-hemisphere lesions lose contextual information and the left hemisphere "makes up a story, and, lacking insight, appears completely convinced by it" — behaving, he writes, "like the sort of person who, when asked for directions, prefers to make something up rather than admit to not knowing."

Why It Matters

This identifies a specific failure mode distinct from ordinary error: confident fabrication produced by a system that has no internal signal for its own ignorance, because the very faculty that would register the gap is the one that's missing. It matters anywhere a narrow, verbal, or rule-based subsystem is asked to explain outcomes it did not fully generate — post-hoc rationalization in people, hallucination in language models, or any reporting layer that answers "why" questions using only the information available to it, unaware that the true cause lived elsewhere.