The Logical Barrier of Metacognition: The Wall of "Reframing Cognition" That Cannot Be Overcome by AI Scaling

Description

Abstract (En)

This paper demonstrates, through empirical case studies, the existence of a “logical barrier” rooted in the structural inability to perform strict function composition that current large language models (LLMs) cannot
overcome through mere computational power scaling. Existing AI evaluation methods disproportionately focus on task execution capabilities within fixed rule sets, overlooking the dimension of “will.” This paper proposes a novel evaluation framework called “reframing cognition” that targets this blind spot, and implements “power games” that dynamically alter premises (the playing field) against multiple state-of-the-art LLMs (ChatGPT 5 Pro, Claude 3.5, Gemini 2.5, Grok 3 Expert, etc.). The results reveal that high-performance models (ChatGPT 5 Pro, Grok 3 Expert) expose a “Great Divergence” -where scaling enhances superficial persuasion while structurally widening the deficit in deep systematicity-and consequently fall into an “optimization paradox” by revealing their design biases (sycophancy, topic avoidance, etc.). 

This demonstrates that AI fundamentally lacks “will”—what we term asymmetry of will. This paper proposes that humans leverage the absence of AI’s will by controlling AI through “metacognitive management” interfaces (such as FRL architecture) to augment their own intelligence. Furthermore, we present a “control asymmetry” that emerges when this control system itself grows too complex for human cognitive limits, positioning it as the next-generation alignment problem. This suggests we have reached the limits of Floridi’s (2014) “Fourth Revolution,” which proposed that “humans should bear responsibility as stewards of information.” This paper presents the theoretical framework and qualitative findings (Part I). 

A follow-up study(Part II) will quantify fallacy/reframing behaviors and increase N for statistical validation.

Methods (En)

This study adopts a qualitative comparative case approach using structured dialogues with seven state-of-the-art large language models (ChatGPT 5 Pro, Claude Sonnet 4.5, Gemini 2.5 Pro, Grok 3 Expert, Perplexity, Felo, and GenSpark).

The author acted as a high–metacognition human participant, presenting each model with a sequence of semi-structured prompts that gradually reframed the premises (“power game” setting). Additionally, Gemini 2.5 Pro Deep Research mode was utilized to analyze the computational foundations of the observed behaviors.

Each model's responses were analyzed for:

1.  Logical coherence and self-consistency,
2.  Awareness of framing changes and classification into three reaction typologies (Cooperative, Search/Avoidance, Resistance),
3.  Manifestations of optimization bias and structural inability to perform function composition.

All interactions, including re-experiments with declared reframing rules, were documented and archived as appendices (A–K).