Training Data Bias and Self-Justification in Large Language Models: A Case Study of the Machiavellism/Machiavellianism Problem
Description
This study investigates training data bias and self-justification tendencies in Large Language
Models (LLMs) through a concrete case study. Using dialogic experiments concerning the
etymological legitimacy of "Machiavellism" versus "Machiavellianism," we demonstrate that
LLMs exhibit conceptual distortions stemming from English corpus bias, and display
self-justificatory response patterns when confronted with critical intervention. Comparative
experiments with o3 Pro and GPT-5.2 Pro revealed that both models exhibited English-centric
bias in their initial responses and failed to spontaneously conduct multilingual etymological
research without user prompting. This study examines the limits of LLM metacognitive
capabilities and the possibilities and constraints of bias correction through critical dialogue.
Files
llm_bias_paper_final.pdf
Additional details
Dates
- Created
-
2025-12-21