Seonglae ChoAI Bias Benchmarks
AI Bias Notion
Clearly defining "what constitutes bias" across 5 dimensions, enabling control in both training and monitoring. During GPT-5 training, a self-evaluation model (LLM grader) scores responses → reflected as a bias loss term. This is not simply "data cleansing" but a structural improvement that incorporates bias itself as a learning objective
