Measuring Consciousness Indicators in Large Language Models: Between Verbal Claims and Testable Evidence

Abstract

This study examines whether large language models (LLMs) display indicators associated with scientific theories of consciousness, and how ChatGPT, Gemini, and Claude differ in metacognitive caution, reasoning, theory of mind, and the distinction between simulation and subjective experience. A descriptive comparative method was used. Fifteen standardized questions were presented to three major language models in separate sessions, and responses were analyzed according to eight indicators derived from prominent theories of consciousness. Results showed that all three models demonstrated strong capabilities in reasoning, theory of mind, and resistance to misleading information. However, they differed in metacognitive caution and awareness of evidential limits. ChatGPT and Claude showed clearer distinction between simulation and subjective experience, while Gemini showed strong reasoning and theory-of-mind performance but less depth in discussing the limits of consciousness claims. The study concludes that these results measure functional indicators associated with consciousness, not proof of subjective consciousness itself, because consciousness still lacks a universally accepted definitive test. Version 1 (Work in Progress). Future versions will include repeated trials, additional models, an expanded question set, and statistical analysis.