Assessing AI Consciousness: From Cacophony to Hierarchy - Hardcover

by Deepmind Institute (Author)

Humans are predisposed to see minds wherever they see purposeful behaviour - we instinctively treat anything that seems to act with beliefs and desires as a conscious agent (Dennett, 1987). As AI systems become more capable, more autonomous, and more behaviourally compelling, the pressure to attribute consciousness to them, or to deny it, will only intensify. Yet the experts whose judgement we might rely on disagree fundamentally - not just about whether current AI is conscious, but about what consciousness even depends upon.

We need a framework for navigating this disagreement - one that lets us ask not merely 'does it seem conscious?' but 'what would it take for it to be conscious, on each of the major theories, and how does the evidence bear on each?' This book develops such a method. The answer, as the book shows, depends on which level of functional description one takes to be critical for consciousness. The developed Bayesian model makes that dependence concrete: in illustrative assessments of current LLMs, and illustrating sensitivity to assumptions, assessments under different stipulated readings and credence distributions differ by nearly two orders of magnitude, driven as much by where theoretical credence is placed across the levels as by how the evidence is read. But with these different commitments made explicit, at least the theories can disagree precisely - knowing exactly what they are disagreeing about, what evidence would bear on the resolution, and how their disagreements relate to each other within a common structure.

The framework thus provides not a resolution of the disagreement but something valuable nonetheless: a principled method for arriving at calibrated assessments in the face of disagreeing experts, by decomposing a single intractable question into two components, theoretical credences and empirical indicators, each of which can be independently evaluated, debated, and updated as new evidence emerges. The need for such a method will grow with AI capability itself: even though consciousness and intelligence are not the same thing, some of the indicators our best theories associate with consciousness overlap substantially with the architectural features that general intelligence requires, so each more capable generation of systems will satisfy more of them, and the framework is designed to be reapplied as they do.

The DeepMind Institute aims to advance bold thinking about the safe development of artificial general intelligence (AGI), its beneficial use, and its implications for society. From the lab that pioneered the field, we focus on bringing creative foresight and grounded examination of AI to help tackle the most important questions of our time and navigate humanity's next era.