BIGbench provides a unified framework to assess multi-dimensional social biases in text-to-image generative AI, addressing a critical gap in AI ethics evaluation.

• Differentiates between representational and allocational biases in image generation
• Offers a systematic methodology for bias measurement beyond simplistic approaches
• Evaluates bias across multiple social dimensions (gender, race, age, etc.)
• Provides actionable insights for developing fairer AI systems

Security Implications: Identifying and mitigating biases in text-to-image models is essential for preventing harmful stereotypes and ensuring safe deployment in commercial applications.

BIGbench: A Unified Benchmark for Evaluating Multi-dimensional Social Biases in Text-to-Image Models