Explanation faithfulness
Whether an explanation actually tracks what caused a decision, or merely offers a plausible story the system would have produced anyway. A faithful explanation reports the real reason; an unfaithful one reconstructs a convincing rationale that leaves the true cause unmentioned. The distinction matters for both machines and minds, since a step-by-step justification can sound convincing while being driven by something it never names.
The property of an explanation reflecting the true cause of a decision, rather than offering a plausible reconstruction that the system would give regardless. LIME promised explanations that were "faithful" to the underlying model; Turpin and colleagues (2023) showed that a language model's chain-of-thought reasoning can be systematically unfaithful.
Faithfulness is the property that joins the two halves of this topic, because it can fail in a machine and in a mind the same way. Turpin's experiment plants a bias in a prompt — reordering options so the answer is always "(A)" — and finds the model follows the bias, changes its answer, and writes a confident step-by-step justification that never mentions the bias. The stated reason is a reconstruction, not a report.
That is the same structure the confabulation literature documents in people: Nisbett and Wilson's subjects explain choices by a priori causal theories rather than introspection, and Haidt's account makes moral reasoning a post-hoc justification of an intuition already arrived at. Whether one demands faithfulness of machines that humans never supply is the question Lipton and Dennett press from opposite directions.