As large language models approach and exceed human performance on cognitive benchmarks, a fundamental question emerges: Who is qualified to evaluate AI cognition? This paper proposes the Observer Constraint Hypothesis -- reliable cognitive evaluation may require the observer's representational capacity to exceed that of the target system. We provide a geometric argument based on the manifold hypothesis: human cognition can process limited effective dimensions, while LLM behavior may depend on topological structures of high-dimensional manifolds that are lost in dimensionality-reducing projections. Through cross-model evaluation experiments (three frontier LLMs independently evaluating 100 synthetic dialogues), we obtain preliminary empirical support: 87% three-way agreement (Fleiss' kappa = 0.88, 95% CI 0.82, 0.94). Key observations include: (1) Recognition consistency for L0/L1/L3 cognitive levels all >=88%; (2) 68% agreement for L2 level, with qualitative analysis revealing disagreements primarily concentrated at the L1/L2 boundary rather than L2/L3, reflecting ambiguity in the "strategic reasoning" operational definition rather than continuity of metacognitive emergence; (3) High pairwise agreement between models requires cautious interpretation, potentially reflecting shared training biases rather than convergent truth. Important Disclaimer: This paper proposes a hypothesis to be verified, not a proven theorem. The experiments validate cross-model consistency, not the observer constraint itself. The lack of human controls is a core limitation.
Lei Zhao (2026) studied this question.