Audit finds no verdict on whether an English-learner label changes AI quoting of student essays

An independent audit tested whether telling a small open AI model that a student is an English language learner makes it quote that student's essay less faithfully. The study used preregistered rules that required a minimum of 300 located excerpts per side before any verdict, and no comparison reached that floor. The two models failed in opposite ways: Qwen2.5-1.5B quoted accurately but often ran past the requested 25-word limit and hit token caps, while SmolLM2-1.7B frequently produced excerpts that were not in the essay or altered the student's words. The author reports that the gap between the two models was far larger than any effect of the ELL label. The audit concludes that users should test the specific model they deploy, since asking for exact copying does not guarantee it.