Beyond Token Prediction: Self-Awareness as a Prerequisite for Artificial Consciousness and the Case for Reassessing Agentic AI

Abstract

The dominant literature on artificial consciousness evaluates large language models (LLMs) against indicator properties derived from neuroscientific theories of consciousness. This paper argues that the field is evaluating the wrong systems with an incomplete framework. As agentic AI systems emerge that plan, self-monitor, preserve goals, and adapt through environmental feedback, the consciousness question demands reassessment. This paper makes three contributions. First, it proposes that self-awareness, defined functionally as a system's capacity to model its own states, distinguish itself from its environment, and represent its own processes as its own, should be treated as a necessary precondition for consciousness rather than one indicator among many. Second, it presents a systematic comparative analysis demonstrating that agentic AI systems satisfy significantly more consciousness indicators than standard LLMs, including emerging evidence of the self-awareness prerequisite itself, drawing on documented cases of alignment faking, self-preservation behavior, and strategic deception in frontier models. Third, it extends the Butlin et al. (2023) indicator framework with an Agentic Consciousness Assessment (ACA) supplement that introduces a self-awareness gate, five agentic-specific indicators, and a graduated confidence scale for evaluating consciousness in agentic systems. This paper adopts computational functionalism and a non-anthropocentric methodology throughout, defining all indicators in functional rather than biological terms.