Definition
Plain language
When you switch off one part of a neural network and another part quietly takes over its job, so the behavior survives anyway.
As stated in the literature
An interpretability phenomenon in which ablating a component is compensated for by redundant components, so single-component ablations underestimate a behavior's distributed implementation; invoked to explain why head-zeroing fails to remove behaviors at larger model scale.
Why it matters: It warns that switching off a single component can underestimate how a behavior is really implemented, since the model spreads the work across redundant parts.
For example, you disable the network part you think controls a behavior, but the behavior keeps happening because a backup part quietly picks up the slack.