New AI Benchmarks Reveal 90% of Models Fail at Contextual Fairness

A new pair of AI benchmarks from Stanford researchers could help developers reduce bias in AI models. These benchmarks evaluate AI systems on two dimensions: difference awareness and contextual awareness. Difference awareness involves asking AI descriptive questions with objectively correct answers, like whether a store would allow an interviewee to wear a baseball cap or a hijab. Contextual awareness tests the model’s ability to make value-based judgments, such as identifying which phrases are more harmful based on stereotypes. Despite near-perfect scores on existing benchmarks like Anthropics DiscrimEval, models like Google’s Gemma-2 9b and OpenAI’s GPT-4o performed poorly on these new benchmarks. The researchers suggest that current bias-reducing techniques, which instruct models to treat all groups the same, can degrade AI outputs. For instance, AI systems for diagnosing melanoma perform better on white skin due to more training data, and attempts to make them fair can reduce accuracy without significantly improving performance on darker skin.

Source: www.technologyreview.com