EMPIRICAL VALIDATION OF CONSCIOUSNESS THEORIES IN ARTIFICIAL NEURAL NETWORKS

This study presents a systematic empirical investigation testing predictions derived from three major consciousness theories—Global Workspace Theory (GWT), Integrated Information Theory (IIT), and Predictive Processing (PP)—in artificial neural networks. The research addressed a critical gap in consciousness science: the lack of rigorous empirical validation of theoretical predictions in controlled computational environments where architectural parameters can be precisely manipulated and internal states completely observed. A multi-study experimental design grounded in post-positivist epistemology examined 12 neural network architectures across four families (transformer-based, recurrent, convolutional, and generative), with 60 trained network instances. Study 1 tested GWT predictions regarding global workspace dynamics; Study 2 examined IIT predictions concerning integrated information and behavioral flexibility; Study 3 investigated PP predictions about hierarchical predictive processing; and Study 4 conducted cross-theory integrative analysis. Results provided strong support for all twelve pre-registered hypotheses. Transformer architectures developed attention patterns consistent with global workspace dynamics (Global Broadcast Index M = 0.850, d = 3.21), exhibiting characteristic ignition dynamics with identifiable thresholds. Phi-star values robustly predicted generalization (r = .847) and behavioral flexibility (r = .912). Generative architectures developed hierarchical predictive structures correlating with anticipation performance (r = .78) and active inference behavior (r = .81). Cross-theory analysis identified recurrent connectivity as necessary for consciousness indicators across all frameworks, with multi-theory-compliant architectures demonstrating superior behavioral outcomes. These findings demonstrate that consciousness theories make valid predictions testable in artificial systems, suggest recurrence as a fundamental architectural requirement for consciousness-related processing, and indicate that GWT, IIT, and PP capture complementary aspects of a unified phenomenon. The research contributes theoretical validation, novel computational metrics, and practical design principles for developing AI systems with enhanced adaptive capabilities. Keywords: consciousness theories, artificial neural networks, Global Workspace Theory, Integrated Information Theory, predictive processing, machine consciousness, integrated information, attention mechanisms, empirical validation