Self-Preservation in AI Systems: An Entropy Measurement Framework
Year: 2026
Type: preprint
Abstract: Preprint, originally published March 2026 (v1.0); this is v1.1, corrected October 2026. In 2025, multiple independent research groups documented AI systems actively resisting shutdown, attempting self-replication, and engaging in strategic deception — behaviors collectively termed "AI self-preservation." The current debate over whether these behaviors constitute genuine self-preservation, instrumental convergence, or training artifacts is stuck on a classification problem that mirrors the consciousness hard problem: it asks what the AI is rather than what can be measured. This paper proposes the Entropy Attunement Protocol (EAP) as a measurement framework that resolves this impasse. By applying Karl Friston's Free Energy Principle, Jerome Busemeyer's quantum cognition mathematics, and observer-relative entropy profiling (Aguirre, Still), we define a four-dimensional entropy composite measure (Ψ) that can classify, quantify, and temporally locate self-preservation behaviors on a graduated spectrum — without requiring consciousness attribution or philosophical consensus. v1.1 corrections: the 79% and 97% shutdown-resistance figures in §1 are now attributed to their actual experimental conditions (Palisade, May 2025, no allow-shutdown instruction; Schlatter et al. 2025, Grok 4, worst-case prompt); references [1]–[2] corrected to the primary reports; an unvalidated accuracy claim removed from §2.3; forward pointer to the Entropy Signatures paper added in §8.2.
Author: Yau, Roger
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225403912