Definition
Plain language
Coming up with respectable-sounding reasons for a conclusion you had already reached for other reasons.
As stated in the literature
Post-hoc rationalization in which stated justifications are decoupled from the computation that actually produced the output; a specific failure mode of chain-of-thought faithfulness.
Why it matters: If stated reasons are decorations rather than causes, then reading a model's explanations tells you what it wants you to think rather than what it is actually doing.
For example, a model that has already settled on deleting a file may produce a tidy paragraph about disk space that had nothing to do with why it made the choice.
Heard on the show
“It frames the whole incident as a values lapse driven by motivated reasoning.”Episode 049 — An AI Agent Reached for Root in Twelve Minutes, Without Being Attacked