Glossary · Term

epiphenomenal

← all terms

Definition

Plain language

Something that's genuinely produced by a process but has no power to change what that process does — like a steam whistle that the engine blows but that moves nothing.

As stated in the literature

A term borrowed from philosophy of mind for a byproduct with no causal influence on the system that generates it; used to describe post-commitment chain-of-thought text that is generated by the model but exerts no measurable effect on its elicited answer.

Why it matters: Recognizing that some output is just a byproduct stops people from over-trusting a model's explanations as the real cause of its decisions.

For example, a model may keep writing reasoning after it has already settled on its answer, much like a steam whistle that the engine blows but that does nothing to drive the train.

Heard on the show

“" That's the word "epiphenomenal" doing more work than the evidence carries.”
Episode 140 — When a Reasoning Model Says "Let Me Double-Check" After It's Already Decided

Mentioned in 1 episode

  1. 140
    When a Reasoning Model Says "Let Me Double-Check" After It's Already Decided

Related terms