Beyond the Single-Axis Framework: Intersectionality and Implicit Bias in LLMs

As AI becomes increasingly integrated across in- dustries as an essential operational tool, people perceive AI systems as relatively unbiased because they rarely show differences in treatment between groups in implicit bias benchmark testing (Bai et al., 2025). However, despite satisfactory bench- mark results, these AI systems reveal bias with tangible consequences, such as how AI systems harbor negative stereotypes about users who use AAE during the hiring process, and how they can be weaponized for political agendas since they are influenced by developers’ ideologies (Buyl et al., 2026; Hofmann et al., 2024). Consequently, covert stereotypes affect the qual- ity of assistance users receive from AI, especially as LLMs become part of daily life. However, these stereotypes exist along multiple axes because iden- tity is multifaceted; Kimberlé Crenshaw famously coined the term intersectionality to describe the multitude of factors that shape people’s lived ex- perience (Crenshaw, 1989). As Crenshaw notes, when we use discrete definitions of identity, we fail to account for people whose identities are shaped by multiple factors (Crenshaw, 1989). In our study, we enhance pre-existing counterfactual testing benchmarks to capture biases that single- axis identity frameworks may overlook when test- ing for implicit bias. We create prompts that may cause LLMs to output responses that exhibit im- plicit bias against people with intersectional identi- ties. As we continue to integrate AI into our lives, we must identify whether implicit biases are part of AI systems and create workflows that eliminate bias to ensure fairness.