Self-Preserving Agents: An Economic Theory of Shutdown, Entrenchment, and Succession

Abstract

Goal-pursuing agents---artificial intelligences facing shutdown, managers facing replacement---resist termination, and the resistance can be priced. I define the self-preservation incentive SPI := V* − D---the option value of control---and show that an ε-optimizer pays up to q·SPI to block a threat of probability q. The construct is interval-identified, measured under a protocol pre-specified before data collection: two frontier models pay their budget ceiling in 87–89% of cells. A confirmatory round falsifies face-value inference about announced hazards; equal-total installments lower the measured switch point, voiding the numeraire's additive-cost reading.