Definition
Plain language
The extra time and effort an attacker has to spend double-checking everything the model told them.
As stated in the literature
The cost imposed when outputs are fluent but unreliably true, forcing independent confirmation of each critical claim; the defense's claimed security property is an increase in this cost rather than denial of access.
Why it matters: It reframes a defense's goal from stopping access to making stolen information expensive to rely on, which is a weaker but sometimes more achievable promise.
For example, someone handed a confident-looking set of instructions has to independently test every quantity and temperature before trusting any of it.
Heard on the show
“And anyone with real expertise, a lab, or verified retrieval doesn't get fooled, they just pay a verification tax.”Episode 244 — The Open-Weight Defense That Feeds Attackers Confident, Falsified Answers