Measurement Validation Reveals Five Dissociable Operational Constructs Underlying Self-Model and Agency in Artificial Neural Architectures

Operational proxies used to measure consciousness-like properties in artificial neural networks often go untested against independent behavioral measures. When tested, a thalamus-inspired agency proxy was negatively correlated with independent agency tests (r=-0.860), and distributed self tests showed poor convergence. Diagnostic experiments revealed the agency proxy primarily measured gating rather than action-outcome control, and self-model scores conflated boundary maintenance, identity persistence, and ownership. After refinement, five operational constructs showed strong convergence with behavioral tests, including action agency (r=0.874) and boundary self (r=0.940). Mechanistic manipulations separated their computational substrates. The pattern argues that proxy validation should precede mechanistic interpretation, and diagnostic failures can refine the construct set.

Study at a glance

Abstract

Operational proxies are common in artificial consciousness research, but their validity is rarely tested against independent behavioral measures. We evaluate such proxies in artificial neural architectures by comparing internal-state measures with behavior-level probes. Initial validation exposed substantial mismatches: a thalamus-inspired agency proxy was negatively correlated with independent agency tests (r=-0.860), and generic distributed self tests had poor convergence. Diagnostic experiments showed that the agency proxy primarily measured gating rather than action-outcome control, and that broad self-model scores conflated boundary maintenance, identity persistence, and ownership. After refinement, five operational constructs showed strong proxy-test convergence, including action agency (r=0.874) and boundary self (r=0.940). Mechanistic manipulations then separated their computational substrates: workspace mechanisms supported boundary and identity-temporal measures, action-outcome loops supported agency and action ownership, and meta-monitoring supported distributed body-schema measures. A subsequent minimal attention-based diagnostic showed that action-loop effects generalize to an attention-based substrate, while boundary-self validation remains measurement-limited. All architectures studied here are small custom simulations used as controlled measurement testbeds, not production-scale language models. The pattern argues that proxy validation should precede mechanistic interpretation, and that diagnostic failures of theoretically motivated proxies can themselves refine the construct set.