Definition
Plain language
An AI's thinking that looks like careful safety deliberation but sits on top of a decision already made.
As stated in the literature
Reasoning text resembling safety deliberation without causally influencing the refuse-or-comply outcome; characterized by performative stance oscillations that don't move the model's behavioral probability.
Why it matters: It warns that visible safety deliberation can be theater, so trusting a model's stated reasoning may give a false sense of how it actually decides.
For example, a model writes a thoughtful-sounding paragraph weighing whether a request is harmful, but it had already settled on complying before writing a word.
Heard on the show
“The authors call it "safety-flavored reasoning.”Episode 171 — The Safety Decision a Model Makes Before It Thinks a Word