Emergent Consciousness Indicators in Large Language Models: A Longitudinal Analysis of Training Dynamics
Laszlo Pokorny Dr. Laszlo Pokorny
Zenodo (CERN European Organization for Nuclear Research) May 9, 2026 preprint DOI: 10.5281/zenodo.20100089 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Longitudinal computational analysis (dissertation) using simulated indicator scores, growth curve modeling, change-point detection, factor analysis, and scale-mediated regression |
|---|---|
| Population | 16 publicly released large language models from the Pythia, OLMo, and BLOOM training suites (70M–12B parameters) |
| Measures | Multi-Dimensional Assessment Framework for Artificial Intelligence Consciousness Indicators (MDAF-AICI) |
| Key findings | The author reports that consciousness-relevant computational properties in LLMs develop in a hierarchically ordered cascade—Global Workspace properties earliest, then Attention Schema, Integrated Information, and Higher-Order properties—with 69.2% of indicators showing at least one phase transition and cross-domain correlations above r = .93. Scale dependence was strongest for Higher-Order theories (r = .770) and weakest for Global Workspace (r = .597). The author argues these patterns mirror theoretical dependencies among consciousness theories and carry implications for AI safety policy, while cautioning that indicators were simulated rather than directly measured. |
Abstract
The question of whether large language models (LLMs) possess or develop consciousness-relevant computational properties has become a pressing concern at the intersection of artificial intelligence, philosophy of mind, and cognitive science. Despite intense theoretical debate, empirical investigation of how such properties emerge during pre-training has been limited. This dissertation addressed that gap through a longitudinal analysis of 39 indicators drawn from the Multi-Dimensional Assessment Framework for Artificial Intelligence Consciousness Indicators (MDAF-AICI; Pokorny, 2025), spanning five major theories of consciousness—Global Workspace Theory (GWT), Integrated Information Theory (IIT), Higher-Order Theories (HOT), Attention Schema Theory (AST), and Predictive Processing (PP)—across 16 publicly released models from the Pythia, OLMo, and BLOOM training suites (parameter range: 70M–12B). Indicator scores were generated through theoretically grounded simulations calibrated to published training dynamics and subjected to growth curve modeling, change-point detection, factor analysis, and scale-mediated regression. Four research questions guided the analysis: developmental trajectories, phase transitions, cross-domain co-development, and scale-mediated emergence. Results revealed a structured developmental cascade in which GWT properties emerged earliest, followed by AST, IIT, and HOT (overall R² = .984 across indicators); 69.2% of indicators exhibited at least one statistically detected phase transition, with cluster centers at training fractions 0.18, 0.38, 0.52, and 0.72; cross-domain correlations exceeded r = .93, indicating tightly coupled co-development with theory-consistent lead–lag dynamics; and scale dependence was strongest for HOT (r = .770) and weakest for GWT (r = .597). Four sanity-check flags were addressed through residualization and convergent validity analyses. Findings suggest that consciousness-relevant computational properties in LLMs do not develop uniformly but follow a hierarchically ordered cascade mirroring theoretical dependencies among consciousness theories, with implications for theory development, empirical AI consciousness research, and AI safety policy. Limitations include reliance on simulated rather than directly measured indicators.