Definition
Plain language
When the same change pushes one system one direction and a different system the opposite direction.
As stated in the literature
Reversal in the direction of an effect across models or conditions; here, an identical blank-image input increases refusal on some checkpoints and increases attack success on others.
Also called: sign reversal, sign inverted
Why it matters: When the same intervention helps one system and harms another, no single safety fix can be assumed to transfer, and averaging across models hides the damage.
For example, attaching a blank image makes one model refuse more often while making another model comply with requests it would otherwise reject.
Heard on the show
“The placebo ladder and the sign inversion are harder to explain away as risk-tracking.”Episode 277 — The Blank White Square That Swings AI Refusal Rates Fifty Points