Definition
Plain language
When an attacker asks the same question many times, gets the same wrong answer every time, and concludes it must be right.
As stated in the literature
Failure mode of plurality-vote extraction in which a defended model re-generates a stable falsified value across samples, so the decoy wins the vote decisively and passes the attacker's confidence test.
Why it matters: It shows that repeating a question is not a reliability check, because a consistently wrong answer is indistinguishable from a consistently right one.
For example, an attacker asks the same question fifty times, gets "heat to 180 degrees" every single time, and concludes the number must be correct — when in fact it was planted.
Heard on the show
“On the fourteen-billion model you get consensus deception.”Episode 244 — The Open-Weight Defense That Feeds Attackers Confident, Falsified Answers