Performing a Self: Shutdown resistance, the consciousness question and why verification matters more than metaphysics
Year: 2026
Type: article
Abstract: Safety evaluations published since late 2024 show advanced AI systems, in controlled tests, deceiving evaluators, disabling oversight, sabotaging shutdown procedures and acting to prevent other models from being shut down. Public reaction often moves straight to a metaphysical question: are these systems conscious, and do they want to survive? This paper argues that the question, while important, is the wrong one to wait on. First, the behaviour has a sufficient explanation without consciousness. Systems pursuing almost any goal can learn that remaining operational helps them reach it, and systems trained on vast amounts of human writing can convincingly perform a self. Second, science cannot yet determine whether any AI system has experience, and may not be able to for some time. Third, and decisively, the safeguards required are largely the same whichever answer turns out to be true. The paper reviews the evidence, separates it from speculation, including popular claims about neural implants and "consciousness energy", and considers how competition between states and firms widens the gap between capability and verification. It proposes a verification framework based on GCTI's QA Triad (verify the input, check the output, keep a human decision) and relates it to the certifications proposed earlier in this series. The author leads the entities described. Declared interests: the author is the founder of GCTI, Managing Member of E.M.B.O.A LLC and Representative Member designate of Emboa Japan GK (in formation). Discussion paper, version 1.0.
Author: Fischer, Brooke
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225432485