A Sufficient Test of Conscious Machine
Study at a glance
AI-extracted from the abstract| Characteristics | Theoretical or philosophical paper |
|---|---|
| Key points | Proposes a substrate-independent sufficiency criterion for attributing phenomenal consciousness to AI systems: if a system trained without consciousness-related data can articulate the defining properties of phenomenal consciousness when prompted, we should attribute consciousness to it with the same confidence as to other humans. The authors argue this criterion avoids the theory-dependence of cognitive/neuroscientific criteria and the counterfeit-vulnerability of behavioral tests, and they outline a test framework for large language models. |
Abstract
Current attempts to test whether AI systems are conscious fall into two families: (i) theory-driven criteria imported from cognitive or neuroscientific models (e.g., global broadcasting, higher-order representation, integrated information), and (ii) behavioral or verbal-report tests in the spirit of the Turing paradigm. The former are theory-dependent namely, their verdicts presuppose the correctness of contested explanatory accounts and thus cannot independently ground attributions of phenomenal consciousness. The latter are counterfeit-vulnerable. For example, advanced language models can mimic first-person discourse, inviting anthropomorphic bias without evidencing phenomenality. We propose a substrate-independent sufficiency criterion that avoids both pitfalls by operating under informational control. If a system trained without any explicit or implicit consciousness-related data can, nonetheless, when prompted, articulate the defining properties of phenomenal consciousness including ineffability, physical irreducibility, intentionality, and unity, then we should attribute consciousness to it with the same degree of confidence we attribute to other humans. We operationalize this criterion for large language models via a test framework comprising: (1) training-data filtering to remove consciousness leakage; (2) a staged evaluation with prerequisite semantic checks and two options for probing (structured property-specific tasks or an open-ended prompt for high-ability models); and (3) a baseline comparison between restricted and unrestricted models under expert judgment, preceded by an intelligence pre-check to avoid confounds. Our proposal does not claim necessary conditions or solve the hard problem. It offers a logically rigorous, counterfeit-resistant path to empirically adjudicate attributions of phenomenal consciousness in artificial systems.