Scholarlink

Large Language Models (LLMs) are essential for knowledge generation in science, business, governance, and education. However, multi-level feedback loops—spanning user-AI interaction, algorithmic curation, and training data feedback—exacerbate Bias, Misinformation, and Errors (BME), driving model drift and information quality decay. This paper introduces three novel metrics—Bias Amplification Rate (BAR), Echo Chamber Propagation Index (ECPI), and Information Quality Decay (IQD)—to quantify and track bias propagation. Simulations reveal evolving risks across iterative updates. We emphasize the need for lifecycle-wide governance incorporating real-time bias detection, algorithmic fairness, and human-in-the-loop verification to preserve long-term reliability, neutrality, and accuracy of LLM outputs.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Echo Chamber Dynamics in LLMs: Mitigating Bias and Model Drift

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Dale Rutherford,
  • Ningning Wu

摘要

Large Language Models (LLMs) are essential for knowledge generation in science, business, governance, and education. However, multi-level feedback loops—spanning user-AI interaction, algorithmic curation, and training data feedback—exacerbate Bias, Misinformation, and Errors (BME), driving model drift and information quality decay. This paper introduces three novel metrics—Bias Amplification Rate (BAR), Echo Chamber Propagation Index (ECPI), and Information Quality Decay (IQD)—to quantify and track bias propagation. Simulations reveal evolving risks across iterative updates. We emphasize the need for lifecycle-wide governance incorporating real-time bias detection, algorithmic fairness, and human-in-the-loop verification to preserve long-term reliability, neutrality, and accuracy of LLM outputs.