Fixed by AI, Flagged as AI: A Matched Audit of AI-Text Detectors on 12,132 Student Essays

Abstract

A 2023 Stanford study (Liang et al., 2023) reported that AI-text detectors flag most essays by non-native English writers as AI-generated: seven detectors labeled 61% of 91 TOEFL essays as AI-written while almost never flagging essays by US eighth graders. But the two sets differ in more than the writer's first language: the TOEFL essays are short test answers on different topics. As a high school student applying to US universities in my second language, I wanted to know whether essays like mine would be judged unfairly, so I repeated the test with a fairer comparison: 12,132 student essays, including 8,244 in which English learners and native speakers in US schools answered the same prompts, 2,989 AI-written essays on those prompts, and 9 open detectors released between 2019 and 2025. On the Stanford TOEFL essays the result holds: trained detectors flag 43%–97% of them. On matched essays it does not: 7 of 9 detectors flag English learners significantly less often than native speakers, and none flags them more often. What does trigger false flags is the text itself. Short essays are flagged more often, and so are essays touched by a language model: when an AI tool corrected only the grammar of students' essays, leaving most words unchanged, every detector flagged more of them, up to 57 times more for English learners' essays (Mozilla's Fakespot detector: 0.5% → 28.5%). Fixing spelling with a dictionary changed almost nothing, and the rewording that the Stanford study found to reduce false flags now increases them for every detector. In these data, detectors do not penalize non-native writers; they penalize text that is short, formulaic, or polished by AI. A flag on a carefully edited essay, such as a college application, says little about who actually wrote it.