The Possibility of Conscious AI Bridging the Gap with the Attention Schema Theory

Abstract

This thesis argues that the Attention Schema Theory (AST), first proposed by Michael Graziano, is the most promising framework for approaching the possibility of conscious AI. The argument proceeds in three stages. The first stage examines the philosophical challenges posed by David Chalmers’ hard problem of consciousness and John Searle’s biological naturalism; both constrain but do not rule out the possibility of conscious AI. The next stage situates the AST within the landscape of informational theories of consciousness to which it belongs. This landscape includes the Global Workspace Theory (GWT) and the Integrated Information Theory (IIT). I argue that the AST’s specificity about the attention schema as an internal model of attention gives it an advantage over these other theories that it shares features with. This prepares the argument for the third stage, which applies the requirements of the AST for consciousness to BabyX, a neurobotic agent developed by Alistair Knott, Mark Sagar, and Martin Takac. The thesis concludes that the AST provides specific and testable criteria for attributing consciousness to AI systems, and that the proximity of AI systems to satisfying these requirements makes ethical considerations about these systems an urgent priority.