From Cacophony to Hierarchy: Assessing AI Consciousness
This presentation introduces a principled framework for evaluating AI consciousness amid profound theoretical disagreement. Rather than declaring systems conscious or unconscious, it proposes a structured methodology that separates metaphysical commitments from empirical evidence, organizes theories by their preferred level of description, and aggregates uncertain findings through Bayesian inference. The result is structured agnosticism: explicit theoretical commitments, defeasible empirical indicators, and credences rather than verdicts.Script
Can a language model feel confusion? The answer depends entirely on which level of organization you think consciousness requires. This paper transforms that messy debate into a structured hierarchy of five testable possibilities.
The authors distinguish the hard problem, which asks what consciousness fundamentally is, from the mapping problem, which asks which system properties correspond to which experiences. Bracketing the first lets us make progress on the second, even when metaphysical disagreement persists.
The critical level is where consciousness actually lives. If it supervenes on algorithms, then silicon can be conscious. If it requires intrinsic physical causation or biological metabolism, simulation is not enough. This single question determines whether a functional copy preserves experience or merely mimics it.
Two networks with identical node counts can differ by a factor of 20,000 in their integrated information. The right one has 11,452 ibits; the left has less than 1. Behavioral equivalence guarantees nothing about the intrinsic causal structure that Integrated Information Theory claims is constitutive of consciousness.
The Bayesian model translates theory and evidence into explicit credences. A human scores 1.0 under any weighting. A fly scores 0.91. But a large language model, under optimistic evidence and coarse-level theories, scores 0.79. Under fine-level theories with identical evidence, it drops to 0.10. The same system, the same facts, radically different conclusions.
The framework does not resolve whether artificial intelligence can be conscious. It clarifies where the disagreement actually is, which evidence would matter, and how to update beliefs as systems and theories evolve. To explore the model, create your own assessments, and learn more, visit EmergentMind.com.
