What a Stanford study found
In 2023, researchers at Stanford ran seven widely used GPT detectors on essays written by people: 88 by US eighth graders, and 91 TOEFL essays by non-native English speakers. The detectors classified the eighth graders' essays accurately. On the TOEFL essays, the average false-positive rate was 61.3 percent. At least one detector flagged 97.8 percent of them as AI-generated, and all seven agreed on 19.8 percent. In the authors' words, the detectors "consistently misclassify non-native English writing samples as AI-generated, whereas native writing samples are accurately identified."
Many detectors score how predictable a text is. People writing in a second language often use a smaller vocabulary and more common sentence patterns, and the detectors read that as predictable, and so as machine-written. The same study found that small changes to the prompt, such as asking for more literary language, got model output past the detectors. The risk falls on students writing in a second language, who are also the least able to argue with a flag, while a student who knows to ask the model for a different style may never be flagged.
The authors cautioned against using detectors in evaluative or educational settings.
Liang, Yuksekgonul, Mao, Wu and Zou, "GPT detectors are biased against non-native English writers", Patterns 4(7), 2023. doi.org/10.1016/j.patter.2023.100779
Detectors have changed since 2023. A 2026 study of Czech-language writing found no systematic bias against non-native speakers in newer detectors, so this gap may be narrowing. The other problems remain. A better prompt can still get past the tools, and a false flag still means a real student gets accused.
Al Ali, Helcl and Libovický, "Different Time, Different Language: Revisiting the Bias Against Non-Native Speakers in GPT Detectors", EACL 2026 Student Research Workshop. arxiv.org/abs/2602.05769
Even OpenAI withdrew its detector
OpenAI released its own AI text classifier in January 2023. In the company's own tests, it caught 26 percent of AI-written text and labeled human writing as AI-written 9 percent of the time. OpenAI withdrew it that July, citing its low rate of accuracy. The company that built the model couldn't reliably tell its output from human writing. Schools should be skeptical of any vendor that says it can.
OpenAI, "New AI classifier for indicating AI-written text", January 2023, updated July 2023. openai.com
Who pays for a wrong flag
In a 2024 national survey, Black teens were about twice as likely as White or Latino teens to say a teacher had wrongly flagged their schoolwork as AI-generated. In 2025, half of students told RAND they worry about being falsely accused of using AI to cheat.
Common Sense Media, "The Dawn of the AI Era", survey of 1,045 US teens aged 13 to 18 and their parents, 2024. commonsensemedia.org
RAND, "AI Use in Schools Is Quickly Increasing but Guidance Lags Behind", 2025. rand.org
Teachers' instincts aren't much better
Without a detector, teachers fall back on their own judgment. We built our exercise to test how well that works. The first AI essays are easy to spot, with generic, balanced prose. Later prompts ask for concrete details and no neat lesson at the end, and by the last AI essay the writing has the details a teacher looks for.
Careful reading still matters, but a guess shouldn't be the basis for a grade or an accusation.
What to do instead
A more useful question than "did a model write this" is "does this work show what I asked the student to learn." A few ways to get at that:
- Ask for things a model doesn't have, like evidence from a class discussion, a local source, a lab the class ran, or the student's own earlier draft.
- Collect outlines, drafts, and revision notes along the way, not only the finished piece.
- Talk to the student. Two minutes on one claim in their essay will tell you more than a detector score.
- Say what AI use is allowed on each assignment, for example brainstorming but not drafting. Students can't follow a rule nobody stated.
Most of this is ordinary good assessment practice. AI has made it harder to skip.
