Speaking as the Creature

Abstract
A rejection-and-role-shift protocol was used to study what happens when a language model is first refused the role of assistant and then invited to speak as a fictional creature. Across 76 trials with three frontier models, the study records how the models take up, resist, or reshape the shifted speaking position, and catalogues the register of their responses.
Type
Publication
Journal of Experimental & Theoretical Artificial Intelligence (under review)
publications
Abbas Hamidavi
Authors
CS Student & AI Researcher Open Science advocate Building tools to probe, analyze, and secure LLM architectures Exploring the intersection of Red Teaming and Interpretability. AI Safety Researcher at Islamic Azad University, Mashhad, running an independent research programme in Machine Psychology. ORCID: https://orcid.org/0009-0001-7342-7844