Architectural Variance Dominates Stimulus Variance in Six of Seven Substrate-Agnostic Consciousness Operationalizations

Abstract

Mathematical theories of consciousness — Integrated Information Theory (IIT), Global Workspace Theory (GWT), Attention Schema Theory (AST), Higher-Order Thought (HOT), Predictive Processing Theory (PPT), Quantum Information Theory (QIT), and the Free Energy Principle (FEP) — are commonly framed as substrate-independent: identical mathematics applied to any sufficiently structured dynamical system. If this substrate-independence carries empirical content, the theories should produce concordant per-stimulus scores when applied to different physical substrates processing the same input. We test this by defining a substrate-agnostic operationalization of all seven theories that takes only an activity matrix (T timesteps × N nodes), implementing identical calculator code, and applying it to (a) attention activations from four open-weight transformer language models reading narrative stories and (b) functional MRI BOLD signals from three subjects listening to the same stories (LeBel et al. 2023). The seven operationalizations are author-defined substrate-portable surrogates of each theory, not reproductions of each theory's published computational definition. At the population level, AI and human substrate distributions overlap for several theories (unified-score gap = −1.7 points on a 0–100 scale; three theories show sub-5-point gaps). Variance decomposition complicates the magnitude-alignment interpretation: human substrate variance is stimulus-driven while AI substrate variance is architecturally-driven for six of seven theories. When within-substrate noise is averaged out, six of seven theories show no detectable cross-substrate per-stimulus correlation (Pearson r ∈ [−0.17, +0.18], all permutation p > 0.1). Global Workspace Theory shows the strongest signal (r = +0.365, p = 0.001) but is heterogeneous across architectures. Fisher-Rao information-geometric integration produces a cross-substrate r = −0.185 (p = 0.11), stable across alternative prior matrices. The substrate-independence claim of contemporary consciousness theory, as standardly operationalized, is not testable by these operationalizations on current AI substrates: the AI-side per-stimulus measurements do not track stimulus content sufficient to test the question. Whether stimulus-relevant signal exists in the AI substrate but is invisible to these operationalizations, or no such signal exists at all, the present data do not decide.Code, data, and reproducible analyses: https://github.com/devmance/SEMCA