AI Services Quietly Send Some Kinds of English to Weaker Models

Big AI companies don't run every question through their smartest, most expensive model. Instead, they use a quick shortcut: estimate how hard your question is, send easy ones to small cheap models, and hard ones to big powerful ones. It's like a hospital triage nurse deciding who sees the specialist. The system only works if the triage is fair.

A new paper from researcher Simran Koul shows it isn't. The shortcut leans heavily on one signal — how long your question is. Longer questions look harder. But different styles of English use different numbers of small filler words like "the," "is," and "of." African American English and the English written by second-language learners tend to drop those words. So the same question, written in those styles, looks shorter, gets labeled "easy," and is handed to a weaker model.

The researcher tested this on 37,704 real sentences written by language learners, plus a controlled set where the same meaning was written in two different styles. The pattern held: non-standard versions were consistently pushed to lower-capacity models.

But here's the twist, and it's the more important part. When the team actually measured answer quality on a ladder of models — phone-sized, laptop-sized, and a top cloud model — they found that *every* tier, including the best one, answered non-standard English less accurately. The routing decision itself added only a small extra penalty. In other words, the AI is already worse at these users' English. The cost-cutting shortcut just piles on top.

For anyone who writes or speaks in a dialect, texts in shorthand, or learned English as a second language, the practical takeaway is uncomfortable: you may be getting a cheaper, weaker AI than the person next to you asking the exact same thing — and even the premium version may not fully understand you. The researcher argues this quietly widens a gap for the users AI already serves worst.