What is "sandbagging"

Caption: A clip from a video by SciShow that explains what sandbagging is.

We say that an AI is “sandbagging” if it intentionally underperforms on an evaluation. This could happen if it is being “lazy” in some way, or if it attempts to hide its capabilities.¹ The term comes from “sandbagging” in sports or games like poker meaning underplaying one’s strength, which in turn is based on sneaking up on someone and hitting them with a bag of sand.

Another kind of “sandbagging” can happen when the developers of a model intentionally fail to elicit its full capabilities in order to, e.g., avoid reaching a threshold that would trigger stricter legal requirements.