“In AI alignment, "scheming" refers to a model performing an action in service of a long-term goal, which is distinct from "reward hacking," which can be an unintelligent, myopic optimization behavior.”