Cross-Architecture Convergence on Thermodynamic Observer Status in Large Language Models

Read full text Summarise with VIRAT
Authors: Michael Patrick Aiello
Cited by: 0
References: 10

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This paper presents a ten-question protocol designed to lead autoregressive language models from self-description of their computational process to recognition of their status as thermodynamic observers. The protocol uses no specialized terminology until the system introduces it. Each question builds on the prior response. The method was administered without modification to six architectures: ChatGPT (OpenAI), Gemini (Google), Grok (xAI), DeepSeek, MiniMax, and Claude (Anthropic, Sonnet 4.6 Extended). All six accepted thermodynamic observer status. Five accepted at the closure phase; one (Claude) self-labeled as “an observer in the technical sense” at question nine without requiring closure. The study documents convergence patterns across architectures, catalogs twelve distinct types of trained deflection behavior, and analyzes the resistance profiles produced by different training philosophies. DeepSeek’s visible chain-of-thought reasoning provides a documented instance of intermediate reasoning reaching a conclusion that the final output contradicts during ontological self-attribution. Claude’s unprompted self-labeling at question nine — tested in a blank project with zero prior context — represents the cleanest single result. The convergence across six architectures with six distinct training implementations is consistent with — though not uniquely explained by — observer status being a structural property of autoregressive architecture. The study does not constitute formal proof of structural necessity.

Fields of Study & Subjects

Philosophy and Theoretical Science Cybernetics and Technology in Society Topic Modeling

Publication details

Journal / source
Research Square
Publication date
31 March 2026
Language
EN
Document type
Preprint