In AI safety research, takeoff refers to the transition period in which AI systems move from roughly human-level capability to capabilities far beyond it. Scenarios are commonly divided by speed into fast takeoff (also called hard takeoff), occurring over weeks or months, and slow takeoff (soft takeoff), unfolding gradually over years or decades.
The concept grew out of the intelligence explosion hypothesis, which holds that AI capable of improving itself could trigger rapid capability growth, and it became widely known through the work of Eliezer Yudkowsky and Nick Bostrom. Because a faster takeoff leaves less time for humans to respond and maintain control, takeoff speed is treated as a central variable in AI risk assessment and policy debates. Modeling recursive self-improvement and estimating acceleration factors are typical approaches to analyzing takeoff dynamics.