The short answer

  • AI detectors are probability estimates, not proof. No detector can prove a text was machine-written; even Turnitin says its scores below 20% are less reliable and marks them with an asterisk.
  • They are badly biased against non-native English writers. In the landmark Stanford study, seven detectors flagged human-written TOEFL essays as AI at an average rate of 61.22% — while classifying essays by US 8th-graders “near-perfectly.”
  • Even the best-case numbers still break at classroom scale. A ~1% false-positive rate across 150 students × 10 assignments is roughly 15 wrongly accused submissions a term.
  • OpenAI withdrew its own detector in July 2023 for low accuracy — the company that built the model could not reliably detect the model.
  • Use detector output as a prompt to talk, never as evidence. Process artifacts and conversation beat scores every time.
A printed student essay on a desk with a magnifying glass resting on it
Detector scores invite scrutiny. They don’t settle it.

Every school year since 2023 has opened with the same hope: that some tool will simply tell teachers which homework was written by a machine. Every year, the evidence says the tool doesn’t exist. This page is a plain reading of what researchers have actually found — because the gap between the marketing and the data is where real students get hurt.

How do AI detectors decide?

Most detectors don’t look for AI. They look for predictability. A language model, by design, tends to choose likely next words. So detectors measure two statistical properties of text:

  • Perplexity — how surprising the word choices are. Low perplexity (very predictable) reads as machine-like.
  • Burstiness — how much sentence length and rhythm vary. Humans write unevenly: a long, winding sentence, then a short one. Models tend to be smoother.

That’s the whole trick. A detector isn’t detecting authorship; it’s scoring how statistically ordinary your prose is. Once you understand that, every one of its failures becomes predictable — including the worst one.

Why do detectors flag non-native English writers?

Because writing in a second language produces exactly the statistical signature detectors treat as machine-generated. A student working in their second language reaches for common vocabulary and clean, simple structures. That is low perplexity. The detector doesn’t know why the text is predictable — it just knows it is.

The scale of this is not a rounding error. In the Stanford study led by James Zou’s group, researchers ran seven widely-used GPT detectors over 91 TOEFL essays written by non-native English speakers with no AI involvement at all:

FindingResult
Average false-positive rate across 7 detectors61.22%
Essays flagged by at least one detector89 of 91 (97.8%)
Essays unanimously misclassified by all seven18 of 91 (19.8%)
US 8th-grade essays (comparison set)Classified “near-perfectly” (the summary gives no single rate)
Human-written essays wrongly flagged as AI (Stanford, 7 detectors)
Non-native writers (average)61.2%Flagged by ≥1 detector97.8%Misflagged by all 719.8%

Source: Liang et al., Patterns (2023), via Stanford HAI

Set those bars against the comparison group: the same seven tools classified US 8th-grade essays near-perfectly. The error lands almost entirely on one group of students, and the researchers put their conclusion in the paper’s title: GPT detectors are biased against non-native English writers.

The number that should end the argument A detector that is right 99% of the time still gets one submission wrong in a hundred. A teacher with 150 students collecting ten assignments a term processes 1,500 documents — roughly 15 false accusations a term, in the vendor’s own best-case scenario. Accuracy that sounds excellent in a marketing deck is a disciplinary crisis in a gradebook.

What do the vendors themselves claim?

Vendor numbers are best-case numbers, and the vendors say so in the fine print. In May 2023 Turnitin said its document-level false-positive rate was under 1% — but only for documents with over 20% detected AI writing, validated on 800,000 pre-ChatGPT academic samples. That qualifier rarely survives into staffroom conversation. The same update put the sentence-level false-positive rate at about 4%, added an asterisk to scores under 20%, and raised the minimum length from 150 to 300 words. Turnitin had initially claimed 98% confidence, and reported a “higher incidence of false positives” below the 20% line (K-12 Dive, June 2023).

The single most telling data point isn’t from a critic. In July 2023, OpenAI withdrew its own AI Text Classifier, citing its low rate of accuracy. On OpenAI’s own challenge set it had caught only 26% of AI-written text while mislabeling 9% of human text as AI. The organisation with the deepest possible knowledge of how its model writes could not build a reliable detector for it. Nothing since has changed the underlying maths.

Where do detectors fail hardest?

Any edit blurs the signal. Detection degrades sharply once AI text is paraphrased, restructured, or run through another tool — precisely what a student who intends to cheat will do. Perkins et al. (2024) tested six major detectors on 805 samples: baseline accuracy was 39.5%, and it dropped to 17.4% on adversarially manipulated text. Weber-Wulff et al. (2023), testing 14 tools, called them “neither accurate nor reliable.” The result is a tool that is toughest on the honest and easiest on the deliberate.

Known failure zones:

  • Non-native English writing — the flagship failure (see above).
  • Edited or paraphrased AI text — accuracy drops substantially.
  • Formulaic genres — lab reports, legal summaries, and standard five-paragraph essays are predictable by design.
  • Neurodivergent and heavily-scaffolded writing — students taught rigid structures produce structured prose.
  • Very short submissions — less text, weaker signal, noisier score.
  • Deliberate evasion tools — an entire industry exists to defeat detectors, and it largely works.

The pattern is uncomfortable: detectors are least reliable exactly where the stakes are highest.

What should a teacher do with a flagged paper?

Treat the score as a reason to look, not a reason to conclude. A flag is a prompt for a conversation — one that a student who did the work can pass easily and one that a student who didn’t, generally can’t.

SymptomLikely causeWhat to do
High score, student writes this way in class tooFalse positive — plain, predictable styleCompare with known in-class writing; drop it
High score, non-native English writerDocumented detector biasDo not act on the score alone
High score, no drafts or version historyPossible AI use — or a student who drafts elsewhereAsk to see process; ask about the argument
High score, work is unlike the student’s usual voiceWorth a conversationAsk them to explain their reasoning aloud
Low score, but the work feels offParaphrased AI or a ghostwriterThe detector can’t help you here

The most reliable instrument in the room is still a teacher who knows what a student sounds like.

A better question than 'did you use AI?' Ask: “Walk me through how you got here.” A student who did the thinking can talk you through the reasoning, the dead ends, and why they cut the second paragraph. A student who outsourced it can’t. No score is as informative as ninety seconds of that conversation — and unlike a detector, it teaches something.

What should a student do if they’re wrongly accused?

Produce process, not protest. Insisting “I didn’t use AI” is unfalsifiable; showing how the work happened is not.

Keep and be ready to show:

  1. Version history — Google Docs and Word both keep it automatically. This is the single strongest artifact you have.
  2. Drafts, outlines, and notes — even messy photos of handwritten planning.
  3. Sources you consulted — browser history, library records, annotated PDFs.
  4. Your own understanding — be able to explain your argument, your evidence, and what you’d change.

Then ask two fair questions: which specific passages were flagged? and what does school policy say about detector scores as evidence? Most policies are more cautious than the accusation implies.

So what actually works?

Design, not surveillance. The schools handling this best have mostly stopped trying to win an arms race they can’t win, and changed what they ask for instead: work that references a specific class discussion, in-class writing samples that establish a voice, staged assignments where the outline and draft are graded, and oral check-ins where students defend their reasoning.

None of that is free — it costs teacher time, which is the scarcest resource in the building. But it’s the only approach whose failure mode isn’t a wrongly-accused sixteen-year-old.

If you’re building that structure, start with designing AI-proof assignments and our classroom AI policy template. Parents worried about the other side of this should read how to spot AI-written homework — with the same caveats that apply here.

An honest caveat about this page

Detection research moves fast, and every number here has a shelf life. Vendors update models, and newer evidence is mixed: a 2025 University of Chicago working paper (Jabarian & Imas, not peer-reviewed) found commercial detectors outperformed open-source ones, with Pangram showing near-zero error rates on unedited text in its corpus. Edited and mixed text remains the hard case. It is entirely possible that detectors get meaningfully better than what’s described above.

What is unlikely to change is the structure of the problem: a detector infers from statistical patterns, patterns vary by writer, and some humans write predictably. Until a tool can distinguish “predictable because machine-written” from “predictable because the writer learned English at fourteen,” the false-positive floor stays where it is. Weigh the evidence accordingly — and keep the burden of proof where it belongs.

For the wider picture of AI and schoolwork, start with our complete guide to AI homework help.

Sources

  • Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., Zou, J. — GPT detectors are biased against non-native English writers, Patterns (2023). The 91-essay TOEFL study: 61.22% average misclassification, 18/91 flagged by all seven, 89/91 by at least one. arXiv 2304.02819 · Stanford HAI summary
  • Turnitin — Annie Chechitelli, AI writing detection update (May 23, 2023): under 1% document-level false positives above the 20% threshold, about 4% sentence-level. Turnitin blog · K-12 Dive, June 7, 2023
  • OpenAI — AI Text Classifier: 26% of AI text caught, 9% of human text mislabeled; withdrawn July 2023 for low accuracy. Original announcement · Decrypt report
  • Perkins, M. et al. — GenAI Detection Tools, Adversarial Techniques and Implications for Inclusivity in Higher Education (2024). arXiv 2403.19148
  • Weber-Wulff, D. et al. — Testing of Detection Tools for AI-Generated Text, International Journal for Educational Integrity (2023). arXiv 2306.15666
  • Jabarian, B. & Imas, A. — Artificial Writing and Automated Detection, BFI Working Paper 2025-116 (not peer-reviewed). PDF

Figures reflect the cited studies and vendor statements as of October 2026. If you spot an error, tell us and we’ll correct it in place.

Related: do AI detectors flag honest students? (6 studies); AI policies in 32 large US school districts. All of our AI guides are collected in the AI & School hub.

Frequently asked questions

How accurate are AI detectors in 2026?
Vendors claim high accuracy — Turnitin initially claimed 98% confidence and says its document-level false-positive rate is under 1%, but only for documents with over 20% AI writing; its sentence-level false-positive rate is about 4%. Independent research is less flattering: a 2024 study of six detectors found 39.5% baseline accuracy, falling to 17.4% on deliberately manipulated text. The honest summary: detectors are decent at spotting raw, unedited AI output and unreliable on everything else.
Can an AI detector be wrong?
Yes, in both directions. It can clear AI-written text (false negative) and it can flag human writing (false positive). The Stanford study found seven detectors misclassified 61.22% of human-written TOEFL essays on average, and all seven unanimously misclassified 18 of the 91 (about a fifth).
Why do AI detectors flag non-native English speakers?
Most detectors score text by how predictable it is. Writers working in a second language tend to use more common vocabulary and simpler sentence structures, which reads as statistically predictable — the same signal detectors treat as machine-like. The bias is a direct consequence of how the measurement works.
Can a teacher fail a student based on an AI detector score?
Schools vary, but a score alone is a weak basis for an academic-integrity case, and Turnitin itself reports a higher incidence of false positives when scores are below 20%. Most institutions require corroborating evidence — drafts, version history, or a conversation with the student.
What should a student do if they are falsely accused?
Stay calm and produce process evidence: draft history, version history in Google Docs or Word, notes, outlines, browser history, and your ability to explain your own argument out loud. Ask which specific passages were flagged and what the school’s policy says about detector evidence.