Signal From Cacophony to Hierarchy: A Principled Framework for Assessing AI Consciousness

Summary

The paper argues that progress on AI consciousness has been hampered by competing theories that talk past each other, and proposes separating the 'hard problem' from the more tractable question of which level of description a system's experience-generating organization sits at. It extends Marr's three levels of analysis into a five-level hierarchy—behavioral, computational, intrinsic causal-structural, organismic, and organism-environment—and positions major theories of consciousness according to which level each treats as critical. For each level, the authors develop operationalizable indicators and assess current AI systems against them, then combine theoretical credences with this indicator evidence in a Bayesian model to produce an overall probability estimate. Applied to current large language models, the resulting assessments range from below 0.01 to roughly 0.8 depending on which theoretical assumptions and evidence readings are used, showing high sensitivity to starting assumptions. The authors note that the indicators associated with each level overlap substantially with architectural features needed for general intelligence, suggesting increasingly capable AI systems could become stronger candidates for consciousness. They advocate a 'structured agnosticism' in which theoretical commitments are stated explicitly and assessments are expressed as aggregated probabilities rather than fixed verdicts.