The "MIT Study" on AI Detectors: What It Actually Is

The topic slug points at MIT, but the peer-reviewed study most often cited for the claim that AI text detectors are unreliable on non-native English writing is not from MIT. It is the Stanford paper by James Zou and colleagues (Liang et al., 2023), published in Patterns. If you are looking for research to cite in an AI-accusation defense, using the correct attribution matters: institutions check citations.

The attribution problem

Search results and social media posts frequently refer to a "MIT study" showing AI detectors are biased against non-native English speakers. From the public record, no such MIT study exists. The paper that produced the widely cited finding is Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., and Zou, J. (2023), "GPT detectors are biased against non-native English writers," Patterns, 4(7), 100779, published by Cell Press. All five authors were affiliated with Stanford University at the time of publication.

Citing the study to the wrong institution is a small error with a real cost. Academic integrity panels read response letters carefully. A misattributed source invites the reviewer to question everything else you write. For a full walkthrough of the correct citation, see our Stanford HAI 2023 study breakdown.

What the Stanford study actually found

Liang and colleagues ran a set of TOEFL essays written by non-native English speakers, alongside essays written by native English speakers, through seven widely used GPT detectors. The disparity was severe. In one of the paper's headline results, more than half of the non-native TOEFL essays were misclassified as AI-generated by the tested detectors, while native English essays were flagged far less often.

The authors also showed that simple prompt engineering (asking a language model to rewrite the non-native essays in more elevated English) collapsed the detectors' ability to flag actual AI text. That finding matters because it demonstrates that the signals detectors treat as "AI-like" are not markers of machine generation. They are markers of low lexical variety and predictable phrasing, features that overlap heavily with second-language writing.

Why the detectors get it wrong

GPT detectors like GPTZero, Originality.ai, and Turnitin's AI indicator rely on statistical properties of the text: perplexity (how surprising each word is given the preceding context) and burstiness (variation in sentence-level perplexity). Non-native writers tend to use a smaller vocabulary and more common phrasings, which produces lower perplexity. The detector reads low perplexity as machine-generated, even when the writing is entirely human.

The Stanford paper's contribution was demonstrating this failure mode at scale, on real student writing, with a clear statistical result. For a broader view of what the peer-reviewed literature has concluded about detector reliability, see the 2023 research on detector accuracy.

How to use the research in your defense

The Stanford finding is most powerful when it maps directly to your situation. English as a second or additional language, TOEFL or IELTS background, or writing shaped by formal ESL instruction all connect the research to the specific accusation you are facing. Your response letter should:

• Name the detector that flagged you and cite the Liang et al. finding as evidence of its known failure mode on non-native writing
• State your language background plainly and give context (years of English instruction, first language, standardized test history)
• Attach process evidence: version history from Google Docs or Word, drafts, research notes, timestamps
• Reference your institution's evidentiary standard and ask whether a detector score alone meets it

The procedural rights FAQ covers what you are entitled to request before a hearing, including the specific detector, threshold, and human review notes.

What the study does not prove

The paper does not prove that any particular student did not use AI. It shows that the class of tools being used to make that determination has a known and severe failure mode when applied to non-native English writing. That is a different claim, and a stronger one for your purposes: it shifts the argument away from "trust me" and toward "the evidence against me is a tool that peer-reviewed research shows misclassifies writing like mine."

If you are preparing a written response and English is not your first language, NotBot generates a personalized defense package that cites the Stanford research by correct attribution, incorporates your language background, and addresses the specific detector used against you. Ready in about a minute.