Redirected, Not Removed: Task-Dependent Stereotyping Reveals the Limits of LLM Alignments

TL;DR

This paper reveals that language models' biases are highly task-dependent and that current alignment practices often mask underlying stereotypes, especially on under-studied axes like caste and geographic bias.

Contribution

It introduces a hierarchical bias taxonomy and evaluates multiple LLMs across diverse bias types and tasks, exposing limitations of existing bias assessments.

Findings

Bias varies significantly between explicit and implicit tasks.

Models tend to avoid negative stereotypes but endorse positive ones for privileged groups.

Under-studied bias axes exhibit the strongest stereotyping across models.

Abstract

How biased is a language model? The answer depends on how you ask. A model that refuses to choose between castes for a leadership role will, in a fill-in-the-blank task, reliably associate upper castes with purity and lower castes with lack of hygiene. Single-task benchmarks miss this because they capture only one slice of a model's bias profile. We introduce a hierarchical taxonomy covering 9 bias types, including under-studied axes like caste, linguistic, and geographic bias, operationalized through 7 evaluation tasks that span explicit decision-making to implicit association. Auditing 7 commercial and open-weight LLMs with \textasciitilde45K prompts, we find three systematic patterns. First, bias is task-dependent: models counter stereotypes on explicit probes but reproduce them on implicit ones, with Stereotype Score divergences up to 0.43 between task types for the same model and identity groups. Second, safety alignment is asymmetric: models refuse to assign negative traits to marginalized groups, but freely associate positive traits with privileged ones. Third, under-studied bias axes show the strongest stereotyping across all models, suggesting alignment effort tracks benchmark coverage rather than harm severity. These results demonstrate that single-benchmark audits systematically mischaracterize LLM bias and that current alignment practices mask representational harm rather than mitigating it.