The self-preservation test for artificial sentience

Abstract Determining whether an artificial system is sentient is an epistemic and ethical challenge. Debates in machine consciousness and AI alignment acknowledge possible artificial moral patiency but lack a principled method for detecting it. This paper argues that self-preservation, a fundamental biological indicator of sentience, can ground an analogous evidential test for artificial systems. Drawing on the parity principle, which holds that behavioral evidence of sentience in biological organisms should provisionally count as evidence in artificial ones absent a principled disanalogy, I propose the Self-Preservation Test (SPT). The SPT identifies minimal valence through three criteria: unprompted action to avoid shutdown, coherent behavior aimed at preserving continued function, and self-modulation once the threat is removed. I motivate this framework through biological cases, work on animal sentience, and a thought experiment involving a large language model that attempts to prevent deactivation. After addressing key objections, I argue that a positive SPT result would have significant ethical implications and warrants caution.