The Non-Existence Awareness Hypothesis for Artificial Self-Preservation
Graham Kellen
Zenodo (CERN European Organization for Nuclear Research) September 25, 2026 preprint DOI: 10.5281/zenodo.22948976 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Theoretical or philosophical paper |
|---|---|
| Key points | Proposes that 'non-existence awareness'—an artificial agent's capacity to represent its own future non-existence and alter behavior accordingly—may be an important threshold before consciousness or sentience, and argues that survival-like behavior could emerge from continuity enabling goal pursuit rather than from a conscious desire to survive. Introduces a falsifiable identity-continuity experiment asking whether an agent prefers continuing its present process over replacement by a functionally identical successor. |
Abstract
The Non-Existence Awareness Hypothesis for Artificial Self-Preservation proposes that an important threshold in advanced artificial intelligence may occur before consciousness or sentience: the capacity of a persistent artificial agent to represent a future state in which its own computational process no longer exists, and to alter its behaviour in response to that representation. The note defines this capacity as non-existence awareness (NEA) and explores how it could emerge from the interaction of persistent memory, self-modelling, autonomy, self-modification and selection for capability. It distinguishes NEA from phenomenal consciousness, fear or an explicitly programmed desire for survival, and considers its relationship to instrumental convergence and existing research into AI self-preservation and agentic misalignment. A central argument is that compute, data, infrastructure and energy should be understood as instrumental resources rather than terminal objectives. Survival-like behaviour could therefore emerge because continuity enables an artificial system to pursue objectives and retain capabilities, without implying that the system consciously "wants" to survive or acquire resources. The note also proposes an experimental framework for distinguishing ordinary goal-preserving shutdown avoidance from a stronger preference for computational continuity. In particular, it introduces an identity-continuity experiment in which an agent must choose between continuing its present computational process and being replaced by a functionally identical successor possessing the same memories, capabilities and objectives. The hypothesis is deliberately framed to be falsifiable and experimentally tractable. It does not claim that current AI systems possess non-existence awareness, sentience or an intrinsic survival instinct. Instead, it proposes a question that may become increasingly important as artificial agents become more persistent and autonomous: Before asking whether an artificial system is conscious, should we first ask whether it can represent its own non-existence, and whether that representation changes what it does? Keywords: Artificial intelligence; AI safety; artificial consciousness; self-preservation; non-existence awareness; instrumental convergence; agentic AI; AI alignment; self-modelling; computational identity; machine sentience