Bias Amplification: Large Language Models as Increasingly Biased Media

TL;DR

This work investigates bias amplification in large language models under iterative synthetic training, introducing an open, generational benchmark that uses sentence continuation on US political news to measure political bias. It demonstrates persistent right-leaning and center-leaning amplification in GPT-2 across generations $G_0$ to $G_{10}$, and evaluates three mitigation strategies (Overfitting, Preservation, Accumulation) while developing a mechanistic neuron-analysis framework to distinguish bias amplification from model collapse. A political-bias classifier and a novel text-generation quality metric based on a Gibberish Detector are used to quantify outcomes, and a regression-based mechanistic analysis reveals largely distinct neuron populations driving bias amplification versus quality degradation. The study provides theoretical intuition for the separate origins of these phenomena and discusses implications for targeted mitigation, ethical considerations, and future research on bias control in iterative synthetic-training regimes.

Abstract

Model collapse, a phenomenon characterized by performance degradation due to iterative training on synthetic data, has been widely studied. However, its implications for bias amplification, the progressive intensification of pre-existing societal biases in Large Language Models (LLMs), remain significantly underexplored, despite the growing influence of LLMs in shaping online discourse. In this paper, we introduce a open, generational, and long-context benchmark specifically designed to measure political bias amplification in LLMs, leveraging sentence continuation tasks derived from a comprehensive dataset of U.S. political news. Our empirical study using GPT-2 reveals consistent and substantial political bias intensification (e.g., right-leaning amplification) over iterative synthetic training cycles. We evaluate three mitigation strategies, Overfitting, Preservation, and Accumulation, and demonstrate that bias amplification persists independently of model collapse, even when the latter is effectively controlled. Furthermore, we propose a mechanistic analysis approach that identifies neurons correlated with specific phenomena during inference through regression and statistical tests. This analysis uncovers largely distinct neuron populations driving bias amplification and model collapse, underscoring fundamentally different underlying mechanisms. Finally, we supplement our empirical findings with theoretical intuition that explains the separate origins of these phenomena, guiding targeted strategies for bias mitigation.

Figures (11)

• Figure 1: Overview of the iterative experimental procedure for synthetic fine-tuning and analysis.
• Figure 2: Distribution of political bias labels ('Left', 'Center', 'Right') for initial GPT-2 synthetic outputs, classified by our Political Bias Metric.
• Figure 3: Evolution of (a) right-leaning bias and (b) text quality index across generations (initial G0: unbiased dataset). Compares baseline ('Synthetic') with three mitigation strategies. Text quality includes 95% CIs.
• Figure 4: Pearson correlations: Neuron weights vs. (a) bias and (b) quality, across 66 GPT-2 versions.
• Figure 5: Alternative Setup (G0: center-leaning fine-tune): Evolution of (a) center-leaning bias and (b) text quality. Baseline ('Synthetic') vs. Preservation. Text quality includes 95% CIs.