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

2026-04-02

AI Summary

This study addresses the limitations of current bias evaluations for large language models, which predominantly rely on single-task benchmarks and fail to capture systemic biases across diverse tasks—particularly along underexamined axes such as caste, language, and region. To remedy this, the authors introduce the concept of task-dependent bias and develop a hierarchical evaluation framework encompassing nine bias categories and seven task types. Using approximately 45,000 structured prompts, they conduct a multidimensional audit of seven mainstream models. Their findings reveal that bias manifestations vary significantly across tasks (with stereotype scores differing by up to 0.43), that less-studied bias dimensions exhibit the strongest stereotypical associations, and that current safety alignment techniques suppress only explicit negative stereotypes while permitting implicit positive preferences, thereby masking rather than eliminating representational harms.

Technology Category

• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy
• Machine Learning: Ethics, Bias, and Fairness
• Philosophy and Ethics of AI: Bias, Fairness & Equity

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Social Networks and Social Media: Fairness and bias in social network and social media analysis

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.

Problem

• bias evaluation
• task-dependent stereotyping
• LLM alignment
• representational harm
• under-studied bias axes

Innovation

• task-dependent bias
• hierarchical bias taxonomy
• implicit vs. explicit stereotyping
• LLM alignment asymmetry
• under-studied bias axes