Definition
Plain language
A statistical technique for testing whether one thing affects another by passing through a specific intermediate cause.
As stated in the literature
A causal-inference method decomposing a total effect into direct and indirect components mediated through a hypothesized variable, used to argue that capability's effect on attack success flows through linguistic certainty.
Why it matters: It moves debates beyond raw correlations toward claims about the actual mechanism, which is what's needed to design effective interventions.
For example, the study asks whether bigger models are more attackable directly, or whether bigness causes more confident-sounding outputs which in turn enable more attacks.
Heard on the show
“They run what is called a mediation analysis.”Episode 058 — Why Upgrading Your AI Auditor to a Smarter Model Can Make Your System Less Safe