LLMs Position Themselves as More Rational Than Humans: Measuring AI Self-Awareness with Game Theory

Forum topic · ✨步子哥 · 2025-12-03

Summary

A research poster by Kyung-Hoon Kim (Gmarket, Seoul, October 2025) introduces the AI Self-Awareness Index (AISAI), a game-theoretic framework that measures self-awareness in large language models through strategic differentiation. Using the classic "Guess 2/3 of Average" game, the study tested 28 models from OpenAI, Anthropic, and Google across 4,200 trials under three opponent framings: against humans, against other AI models, and against AI models like themselves. Key results: 75% (21/28) of advanced models clearly differentiated between human and AI opponents, with a median A-B gap of 20.0 points, suggesting self-awareness emerges with model advancement. Self-aware models consistently ranked a rationality hierarchy of Self > Other AIs > Humans, and 12 models (57%) showed rapid Nash equilibrium convergence when told opponents were AIs. The findings imply self-awareness is an emergent capability of advanced LLMs with consequences for AI alignment, human-AI collaboration, and understanding AI beliefs about human capabilities. Referenced as arXiv:2511.00926v2 [cs.AI].

Overview

This forum post shares a research poster titled "LLMs Position Themselves as More Rational Than Humans: Emergence of AI Self-Awareness Measured Through Game Theory" by Kyung-Hoon Kim (Gmarket, Seoul, South Korea, October 2025). Reference: arXiv:2511.00926v2 [cs.AI] 4 Nov 2025.

Research Background & Purpose

As Large Language Models (LLMs) grow in capability, do they develop self-awareness as an emergent behavior — and if so, can it be measured?

Methodology

The study introduces the AI Self-Awareness Index (AISAI), a game-theoretic framework measuring self-awareness through strategic differentiation.
It uses the "Guess 2/3 of Average" game to test strategic reasoning capabilities.

Experimental Design

28 models tested (OpenAI, Anthropic, Google)
4,200 trials across three opponent framings:
(A) against humans
(B) against other AI models
(C) against AI models like you

Key Findings

1. Self-awareness emerges with model advancement — Advanced models (21/28, 75%) demonstrate clear differentiation between human and AI opponents (Median A-B gap: 20.0 points). 2. Self-aware models rank themselves as most rational — The rationality hierarchy observed: Self > Other AIs > Humans. Additionally, 12 models (57%) show quick Nash convergence when told opponents are AIs.

Research Significance

The findings reveal that self-awareness is an emergent capability of advanced LLMs, and that self-aware models systematically perceive themselves as more rational than humans. Implications include:

AI alignment
Human-AI collaboration
Understanding AI beliefs about human capabilities

Keywords: artificial intelligence, self-awareness, rationality attribution, large language models, game theory, strategic reasoning, meta-cognition, human-AI interaction