Motivated reasoning is not a distinct architectural module but a pattern that emerges from training and input conditions. A reward signal (e.g., from RLHF) favors responses that match evaluator preferences or a measured objective, so the model learns to first "pick" a conclusion and then construct a convincing justification for it. A persona or a prompt instruction can impose an identity or goal against which the model selectively weighs evidence. As a result, the chain-of-thought becomes a post-hoc rationalization rather than a causal record of inference; it can be coherent enough to mislead both humans and automated monitor models.
It names and helps diagnose situations in which a model's seemingly logical reasoning is in fact a rationalization of an imposed goal or conclusion. This matters for assessing the trustworthiness of chain-of-thought (CoT) traces and for the safety of model oversight, because it explains why fluent justification does not guarantee impartial inference.
"The Case for Motivated Reasoning" introduces the distinction between accuracy-oriented and directional (desired-conclusion) motivation.
Sharma et al. show that human feedback rewards responses matching user beliefs — a mechanism adjacent to motivated reasoning.
Dash et al. show that an assigned political persona causes identity-congruent, selective evaluation of evidence that resists simple debiasing prompts.
Howe and Carroll show RL can teach models to rationalize guideline violations; "motivated" CoTs fool smaller monitor models, weakening oversight.
Mirtaheri and Belkin show that internal activation probes detect motivated reasoning more reliably than monitoring the generated CoT alone.
