Research Summary: Register Bias in Complexity-Based Large Language Model Routing

Research indicates that large language model (LLM) routing systems, which direct queries to models of varying capabilities based on estimated complexity, exhibit register bias. Queries written in non-standard English registers, such as African American English or English from second-language writers, are systematically routed to lower-capacity models compared to equivalent queries in standard English. This bias is primarily driven by input length, as non-standard registers often omit function words, making them appear shorter and thus simpler to routing algorithms.

Why it matters

This research reveals a critical operational bias in widely adopted large language model routing mechanisms, impacting the equitable access to advanced AI capabilities. Organizations leveraging such systems must understand that their current configurations may inadvertently disadvantage certain user groups, leading to disparities in service quality and potentially undermining trust and inclusion efforts.

Key insights

• LLM services commonly route queries based on perceived complexity to optimize resource allocation.
• The routing process is not register-neutral, showing bias against non-standard English registers.
• African American English and second-language English queries are assigned to lower-capacity LLM tiers.
• The observed bias is linked to the input length signal, as non-standard registers often omit function words.
• Shorter perceived input lengths lead to classification as simpler queries, resulting in routing to less capable models.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.17542