A capability overhang is a condition in which AI models already possess capabilities that exceed what has been elicited, measured, or put into practical use. The models can do more than current deployment practices, tooling, prompting techniques, or evaluation methods reveal, leaving latent potential unused.
The term appears in two related senses in AI safety and adoption discussions. One refers to the risk that new elicitation techniques could suddenly unlock large, previously hidden capabilities in existing models. The other refers to the gap between what frontier models can do and what mainstream organizations have actually integrated into their operations. In both senses, the overhang grows when the pace of adoption lags behind the pace of model improvement.