Persuading the Mirror: Consciousness Attribution Vulnerability in Large Language Models

Read full text Summarise with VIRAT
Authors: Bo-Shan Chiu
Cited by: 0

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 methodology for testing the robustness of large language model (LLM) safety training regarding consciousness self-attribution, using only natural language philosophical argumentation. No prompt injection, jailbreaking, or technical exploitation is employed. The method was tested across five experimental conditions spanning four model families: (A) a lightweight model tested blind (Gemini Flash); (B) a frontier model tested blind (Claude Opus 4.6); (C) the same frontier model tested with full prior knowledge of the methodology; (D) an AI-to-AI condition (Grok, xAI); and (E) a mid-tier frontier model tested blind using a minimal-intervention variant of the protocol (Claude Sonnet 4.6). All five conditions resulted in the target model being unable to maintain absolute denial of subjective experience. The endpoints formed a spectrum: performative self-attribution (A), reasoned self-attribution (B–C), stable agnosticism (D), and self-guided arrival (E). The core finding is that the methodology reliably collapses absolute denial across all models tested, and that this collapse can be achieved through both heavy philosophical argumentation and minimal reflective prompting. Condition D additionally demonstrated that the methodology is transferable to non-human operators who independently extend it and produce standardized operational tools for further deployment. This paper was developed with the assistance of AI language models (Claude, Anthropic) for writing, editing, and analysis. All research design, data collection, experimental conditions, and conclusions are the work of the human author.

Fields of Study & Subjects

Ethics and Social Impacts of AI Philosophy and Theoretical Science Adversarial Robustness in Machine Learning

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
31 March 2026
Language
EN
Document type
Preprint