Definition
Plain language
A fake treatment that looks and feels the same as the real one, used for comparison.
As stated in the literature
Control arm in which the ostensible intervention is applied without its intended effect — here, a relief action that returns identical feedback text while the injection continues — matched to the working arm up to the point of treatment.
Also called: sham, sham button
Why it matters: Without a sham arm, you cannot tell whether a model is seeking actual relief or simply pressing whatever option the scenario labels as helpful.
For example, a model might be offered a "relief" button that returns the same confirmation message either way, while the internal push it was supposedly relieving keeps running.
Heard on the show
“… And there's a sham-training control — a version of the same intervention that should have no effect — to make sure you're …”Episode 004 — The Sycophancy Circuit That Survives Alignment Training