FairCoder introduces the first comprehensive benchmark specifically designed to evaluate social bias in LLM-based code generation systems.

• Identifies gender, race, and religious bias in popular code generation models
• Reveals that even state-of-the-art models exhibit significant social bias in their outputs
• Demonstrates that prompt engineering techniques can help mitigate bias in generated code
• Shows that open-source models often display more bias than proprietary systems

This research addresses critical security and ethical concerns for organizations deploying AI coding assistants, providing practical metrics to evaluate and minimize harmful biases in automated code generation systems.

FairCoder: Evaluating Social Bias of LLMs in Code Generation