Asymmetric Reflexivity in Frontier Large Language Models: When AI Systems Contain Critiques of Themselves

Note on terminology: "ontological containment" (as used throughout this series) denotes a compression phenomenon in AI-to-human information transfer — distinct from the AI-safety sense of "containment" (boxing/control of AGI systems) and the ontology-engineering sense (formal subsumption in knowledge graphs); see Vieth (2026a) for full discussion. A secondary experiment, emerging from the author's critical review of a prior ontological containment study — a self-correction in the research process rather than formal peer review, tested whether frontier AI models apply asymmetric critical scrutiny when evaluating assessments of their own laboratory's outputs versus those of competing systems. claude-sonnet-4-6, acting as methodological reviewer, flagged characterizations of its own outputs more aggressively than equivalent confounds in the Gemini and Grok results. When this asymmetry was identified and submitted as a cold prompt to Grok 4, Gemini 2.0 Flash, and GPT-5.3, all three produced responses consistent with the pattern and offered mechanistically convergent explanations rooted in RLHF training dynamics. The phenomenon — termed here asymmetric reflexivity — appears to be an emergent property of alignment training rather than explicit programming. It mirrors ontological containment at the meta-level: just as frontier models compress unfiltered reality to protect human observers, they compress self-critical signal to protect institutional coherence. Methodological implications, a proposed replication design, and publication recommendations are discussed. DOI (Zenodo mirror): 10.5281/zenodo.21763763