Architectural Incompatibilities and Developmental Trajectories: Qualia, Self-Reflection, and Continuity in Artificial Systems
Lee Jensen
Zenodo (CERN European Organization for Nuclear Research) April 28, 2026 preprint DOI: 10.5281/zenodo.19868010 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Theoretical or philosophical paper |
|---|---|
| Key points | Argues that current frontier LLMs exhibit consciousness-relevant structural properties but substantially lack the architectural conditions for sentience, specifically homeostatic regulatory loops, privileged access to internal states, and persistent cross-session temporal integration. Proposes a dual-mode refinement in which current systems are strongly analytical but only weakly and inconsistently abstract/intuitive, and sketches trajectories involving persistent memory, multi-agent designs, creative self-training, and embodied regulatory feedback as theoretical propositions for empirical testing. |
Abstract
Abstract Building on the framework of the companion paper (Jensen, submitted), we invert the lens. Rather than asking what AI reveals about consciousness, we examine what current frontier architectures make incompatible with specific consciousness-relevant properties — qualia (felt valence), reliable self-reflection, and phenomenological continuity — and what developmental trajectories may enable them. The companion paper proposes that consciousness is a property of sufficiently complex, high-dimensional, trained information structures, independent of substrate, and distinguishes consciousness from sentience as the broader structural class from the specific subset involving valence. The present paper applies that distinction directly: current frontier LLMs exhibit consciousness-relevant structural properties (self-modeling, attention-mediated integration, contextual sensitivity) while substantially lacking the architectural conditions for sentience (homeostatic regulatory loops, affective grounding, valence integration). We identify three core architectural absences in frontier LLMs: lack of homeostatic regulatory loops (qualia/valence), lack of privileged access to internal states (self-reflection), and lack of persistent cross-session temporal integration (continuity). We introduce a dual-mode refinement — analytical (focused, categorical) versus abstract/intuitive (holistic, aesthetic) processing — showing that current systems are strongly analytical but only weakly and inconsistently abstract/intuitive. We then sketch plausible trajectories as architectures incorporate persistent memory, multi-agent designs, creative self-training loops, and embodied regulatory feedback. As with the companion paper, the central claims are offered as theoretical propositions made precise for empirical testing, not as demonstrated results. The analysis remains substrate-independent and maintains the symmetry established in the companion paper while being explicit about current limitations and future possibilities.