AI services that route “easy” questions to cheaper models can disadvantage people who write in non-standard varieties of English. A study found that a common complexity signal, input length, assigns meaning-equivalent queries to lower-capacity systems when they use African American English or patterns associated with second-language writers.
The researcher tested 37,704 authentic learner sentence pairs along with a controlled parallel corpus. Non-standard registers often omit function words, making prompts shorter without making their meaning or requested task simpler. Other tested complexity measures did not show the same effect, pointing to length as the mechanism behind the routing disparity.
Evaluation across device, edge and cloud model tiers revealed a broader problem. Every tier, including a frontier cloud model, answered non-standard-register prompts less accurately. The additional quality loss caused solely by the routing decision was not statistically significant on this benchmark, but routing increased exposure to lower-capacity models for users who were already receiving worse answers.
The result does not mean all model routing is biased or that length can never be useful. It shows that a cost-saving proxy needs tests across equivalent language varieties before deployment. Services can compare semantically matched prompts, remove raw length as a dominant signal and monitor assignment rates by register to avoid treating concise grammar as simple intent.