Exploring Subjectivity for more Human-Centric Assessment of Social Biases in Large Language Models

An essential aspect of evaluating LLMs is identifying potential biases. This is especially relevant considering the substantial evidence that LLMs can replicate human social biases in their text outputs and further influence stakeholders, potentially amplifying harm to already marginalized individuals and communities. Therefore, recent efforts in bias detection invested in automated benchmarks and objective metrics such as accuracy (i.e., an LLMs output is compared against a predefined ground truth). Nonetheless, social biases can be nuanced, oftentimes subjective and context-dependent, where a situation is open to interpretation and there is no ground truth. While these situations can be difficult for automated evaluation systems to identify, human evaluators could potentially pick up on these nuances. In this paper, we discuss the role of human evaluation and subjective interpretation to augment automated processes when identifying biases in LLMs as part of a human-centred approach to evaluate these models.