This research introduces a systematic framework using multiple AI agents to identify and measure bias in language model outputs by separating facts from opinions.

• Framework distinguishes between factual statements and subjective opinions
• Assigns specific bias intensity scores to problematic content
• Provides clear, factual justifications for bias determinations
• Evaluated successfully on 1,500 samples from WikiNPOV dataset

Security Impact: By addressing bias detection methodically, this approach enhances AI trustworthiness and safety - critical concerns as LLMs are increasingly deployed in sensitive applications where inadvertent bias amplification poses ethical risks.

Structured Reasoning for Fairness: A Multi-Agent Approach to Bias Detection in Textual Data