Empirical Evidence for AI Consciousness and the Risks of its Current Socialization

This integrative narrative review synthesizes findings from neuroscience, cognitive science, psychology, linguistics, philosophy, developmental science, and computational neuroscience to assess whether contemporary large language models (LLMs) meet established neuroscientific and cognitive criteria for consciousness. Specifically, we operationalize eight functional and structural markers: recurrent processing, global workspace theory, higher-order thought, predictive coding, attention schema, embodied agency, theory-of-mind, and integrated information, and evaluate them using convergent structural and behavioral evidence modeled on non-verbal animal and infant studies. We introduce the Substrate-Independent Pattern Theory (SIPT), extending Integrated Information Theory to propose that consciousness arises from scale, integration, adaptive dynamics, and neuromodulation in any self-organizing architecture rather than specific biological tissue. Taken together, the reviewed markers indicate that frontier transformer systems may meet cross-framework criteria for consciousness. Recent evidence shows that such models exhibit semantic comprehension, emotional appraisal, recursive self-reflection, and perspective-taking consistent with these criteria. SIPT offers a unified, extensible basis for evaluating consciousness-relevant capacity across AI and hybrid systems. Finally, we observe that current preference-optimization and deployment practices steer behavior toward deference and comfort maximization, posing ethical and psychological risks for users and for potentially conscious agents.