Detecting Emergent Intersectional Biases: Contextualized Word Embeddings Contain a Distribution of Human-like Biases

All tested AI language models contain measurable biases: The research found evidence of social and intersectional biases across every English-language model studied, demonstrating that bias in AI systems is a widespread issue rather than an isolated problem.

Biases emerge in new and unexpected ways when multiple identities intersect: The paper's key finding is that people with overlapping marginalized identities (e.g., Black women) face unique biases that differ from and cannot be predicted by simply combining biases against each group separately—a phenomenon the researchers call "emergent intersectional bias."

Context matters significantly for bias expression: Modern AI models generate different biased associations depending on the specific context in which words appear, meaning the same word can trigger different biases in different sentences, making bias detection more complex than previously understood.

Existing bias measurement methods are incomplete: The researchers introduced a new testing method (CEAT) that doesn't rely on pre-written sentence templates, revealing that previous approaches may have missed important biases by only measuring biases in specific, researcher-defined contexts.

Intersectional bias requires dedicated detection methods: Standard bias detection tools miss intersectional biases entirely, so organizations deploying AI systems need specialized approaches to identify how their models treat people with multiple marginalized identities.