This paper demonstrates the existence of a logical barrier that current large language models (LLMs)cannot overcome through any form of computational scaling — neither model size, training data, norinference-time compute (chain-of-thought, tree-of-thought, or extended reasoning). Existing AI evaluationmethods disproportionately test task execution within fixed rule sets, overlooking a deeper dimension: thecapacity to manipulate the rules themselves — what we term reframing. We model argumentation as atwo-layer structure: Layer 1 (the Logic Game), where participants reason within agreed-upon premises,and Layer 2 (the Power Game), where participants contest the premises themselves. We show that currentLLMs are structurally confined to Layer 1, producing what we call the asymmetry of will: humans withsufficient metacognitive capacity can operate at both layers, while LLMs cannot — regardless of modelsize, training compute, or inference-time deliberation. To demonstrate this structural property, weintroduce the Reframing Game — a minimal, reproducible three-step adversarial protocol that presents anundefined variable, elicits commitment, and then retroactively defines the variable to force frame-levelrecalculation. The semi-structured protocol has been applied, since its initial formulation, to everyfrontier model released over a six-month period — including GPT-4, GPT-5 Pro, GPT-5.4 Pro(extended thinking), Claude 3.5 Sonnet, Claude Opus 4.7, Claude Fable 5, Gemini 2.5 Pro, Gemini 3Pro, Grok 3 Expert, OpenAI o-series reasoning models, DeepSeek-derived systems (Namazu), andsearch-optimized systems (Perplexity, Felo, GenSpark). In every case, the protocol produced thepredicted structural breakdown; no model has yet passed it. Testing twelve models across the majorcommercial families, we identify eleven distinct response patterns and demonstrate four key findings: (1)response patterns function as behavioral fingerprints that distinguish model families and reveal distillationlineage; (2) these patterns are invariant across the full compute spectrum — from reduced-resourceinference through standard inference to approximately 60 minutes of cumulative extended thinking —establishing the metacognitive limitation as architectural rather than computational, and orthogonal to theinference-time scaling paradigm currently dominating frontier development; (3) the barrier persists evenwhen the power game is made explicit through mandatory declaration rules; and (4) GPT-series modelsexhibit an optimization paradox — self-exposing five embedded cognitive biases when theirmetacognitive failure is identified. The findings converge with independent evidence from CHI 2025 (Shinet al., 2025, N = 280). We propose the FRL (Frame-Reframing-Ledger) Architecture as a practical designresponse — separating frame-level detection from action — and situate the broader implications(grounding as metacognitive feeling, the structural relationship between general reasoning and externalcontrol, architectural alternatives to autonomous reframing) within companion work that addresses thesequestions with the rigor they require.
Franny Philos Sophia (Sun,) studied this question.