The African Woman is Rhythmic and Soulful: Evaluation of Open-ended Generation for Implicit Biases

This study investigates the subtle and often concealed biases present in Large Language Models (LLMs), which, despite passing explicit bias tests, can still exhibit implicit biases akin to those observed in humans who profess egalitarian beliefs yet demonstrate underlying prejudices. The challenge of measuring such biases is exacerbated as LLMs become increasingly proprietary, restricting access to their internal mechanisms such as embeddings, which are crucial for applying traditional bias measures. To tackle these issues, this study introduces innovative measures of bias inspired by psychological methodologies: the LLM Implicit Association Test (IAT) Bias and the LLM Decision Bias. The LLM IAT Bias is a prompt-based method designed to unearth implicit biases by simulating the well-known psychological IAT but adapted for use with LLMs. The LLM Decision Bias measure is developed to detect subtle discrimination in decision-making tasks, focusing on how LLMs choose between individuals in various scenarios. Open-ended generation is also utilised through thematic analysis of word generations and storytelling. The experiments revealed biases across gender and racial domains, from discriminatory categorisations to exoticisation. Our findings indicate that the prompt-based measure of implicit bias not only correlates with traditional embedding-based methods but also more effectively predicts downstream behaviors, which are crucially measured by the LLM Decision Bias. This relationship underscores the importance of relative, rather than absolute, evaluations in assessing implicit biases, reflecting psychological insights into human bias assessment. This research contributes to the broader understanding of AI ethics and provides suggestions for continually assessing and mitigating biases in advanced AI systems, emphasising the need for more qualitative and downstream focus.