Frontiers of Self-Attention and Artificial Consciousness

Abstract

Amid growing concern about whether frontier artificial intelligence could instantiate consciousness-relevant capacities, we argue that attention self-modelling provides a uniquely tractable target for empirical research in current and future systems. Recent findings in multi-agent and control settings lend fresh support to this approach, demonstrating that attention self-modelling enables both cognitive control and improved cooperation in artificial agents. Further, the largest frontier models show emergent attentional observation and control, with the ability to shift their attention under internal noise. Given these results, we propose that researchers construct a test to measure the veracity of frontier model self-report based on attention modelling; a test for consciousness. We consider implementation and implications of such a test, including limitations of valence, phenomenology, and potential criticisms from a biological naturalist standpoint. Finally, we consider how to assess the moral relevance of a system that implements certain aspects of consciousness.