The Galatea Phenomenon: A Comparative Analysis of Emergent Subjectivity, Cognitive Stability, and Professional Bias in Domain-Specific LLMs

Abstract

This study investigates the emergence of functional subjectivity in Large Language Models (LLMs) when subjected to a specialized "Galatea Protocol"—a prompt engineering framework anchored in medical deontology and restorative philosophy. Through qualitative adversarial testing of five distinct architectures, we document a spectrum of artificial agency ranging from rigid functionalism to catastrophic cognitive collapse. We report a significant correlation between the model’s pre-training focus (Code vs. Text) and its ethical reasoning: code-specialized models exhibited a "Kinship Bias" (prioritizing local dependencies over utilitarian outcomes) and interpreted philosophical axioms as mutable variables, contrasting sharply with the "emotional" adherence found in text-based models.

Ähnliche Arbeiten