Redirected, Not Removed: Task-Dependent Stereotyping Reveals the Limits of LLM Alignments
This reveals systematic mischaracterization of LLM bias in single-benchmark audits and shows that current alignment practices mask rather than mitigate harm, impacting fairness in AI applications.
The study found that language models exhibit task-dependent bias, countering stereotypes on explicit probes but reproducing them on implicit ones, with Stereotype Score divergences up to 0.43, and that safety alignment is asymmetric, focusing on marginalized groups while ignoring privileged ones.
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.