AI Tools for English Teachers

AI Detectors for Teachers: What They Catch and What They Get Wrong

I ianadmin · 20 November 2026 · 9 min read

AI detectors cannot prove a student used AI. They produce a probability score, they misfire on perfectly honest writing, and they misfire hardest on writing by non-native English speakers. For English language teachers that makes them unusable as evidence. They have a role as a prompt to look more closely, and no role in an accusation.

Key takeaways

  • Every AI detector outputs a likelihood, not a finding. There is no technology that can prove authorship of a piece of text.
  • False positives are common. Reporting in Nature puts GPTZero’s false-positive rate on genuinely human writing at an estimated 16%.
  • The problem is far worse for language learners. A Stanford-linked study found common detectors wrongly flagged over 61% of TOEFL essays written by non-native English speakers.
  • That happens because detectors reward unpredictable, varied language, and learners writing in a second language produce exactly the opposite.
  • Turnitin, the tool most institutions actually use, reports its own false-positive rate and advises that its score is not grounds for an allegation.
  • Process evidence beats detection. Drafts, version history, in-class writing and a two-minute conversation tell you more than any percentage.

What an AI detector actually does

An AI detector does not look for a watermark or a fingerprint. It measures how predictable the writing is.

Language models choose each word from a probability distribution, and they tend to choose likely words. Detectors measure two related things: perplexity, meaning how surprising the word choices are, and burstiness, meaning how much sentence length and structure vary across a piece. Text that is smooth, even and predictable scores as probably AI. Text that is uneven, surprising and varied scores as probably human.

That is the whole mechanism, and once you see it the central problem becomes obvious. Predictable writing is not evidence of a machine. It is evidence of predictable writing.

Teacher marking student writing at a laptop while considering AI detection results

Why detectors fail English language learners specifically

This is the part most guidance for teachers leaves out, and it is the part that matters most if you teach English.

A B1 learner writing an essay draws on a smaller vocabulary, reaches for familiar sentence patterns, and often works from structures practised in class. The result is writing that is grammatically careful, tonally even and lexically repetitive. To a detector measuring perplexity and burstiness, that profile is close to indistinguishable from generated text.

The evidence bears this out. A Stanford-linked study found that widely used detectors falsely flagged more than 61% of TOEFL essays written by non-native English speakers as AI-generated, while performing far better on writing by native speakers. In other words, the tool is most likely to accuse exactly the students least able to defend themselves, and most likely to be penalised for it.

There is a second, quieter unfairness. Learners who use legitimate support, a grammar checker, a translation tool for a single phrase, or the editing suggestions built into their word processor, will often push their text toward the smooth, predictable profile detectors punish. Good practice gets flagged as cheating.

What the main detectors claim, and what to make of it

Figures below are as published by each provider or reported in the sources named. Treat any accuracy claim from a vendor as a marketing number until independently tested.

Tool Where teachers meet it What it reports The catch for ELT
Turnitin Built into university and college submission systems A percentage of text it judges AI-generated Turnitin itself advises the score is not proof and should not trigger an allegation on its own
GPTZero Standalone, widely used by individual teachers Sentence-level probability highlighting Estimated 16% false-positive rate on human writing, per reporting in Nature
QuillBot and Scribbr Free web checkers students use too A percentage plus highlighted passages Free tiers are heavily length-limited and accuracy is not independently audited
Winston AI Paid, marketed to educators Score plus a readability profile High vendor-claimed accuracy, limited independent verification
Grammarly authorship Inside the writing tool itself Tracks how text entered the document Process data rather than detection, but only if the student drafted there

Compiled from published vendor documentation and the research named above. Verify current figures with each provider before relying on them.

Reading a score without drawing the wrong conclusion

If you do run a detector, the number is a question, not an answer. Three habits make it safer.

Never treat a percentage as a verdict. A 90% score means the text is statistically similar to generated writing. It does not mean there is a 90% chance the student cheated. Those are different claims and conflating them is how teachers end up in unwinnable meetings.

Check it against what you already know. You have months of that learner’s work. A sudden jump in register, vocabulary well beyond their speaking level, or a flawless piece from a student who struggles in class tells you more than any tool. So does the reverse: a flagged essay that reads exactly like everything else they have written is almost certainly theirs.

Never paste student work into a free web checker without thinking about privacy. You are uploading a minor’s work to a third party in many cases. Check what your institution permits.

What actually works instead

The honest answer is that detection is the weakest tool available and assessment design is the strongest. Four approaches that hold up.

Collect process, not just product. Ask for an outline, a first draft and a final version. Version history in a shared document shows writing happening over time. A student who produces a finished essay with no trace of process has a question to answer that no detector score can raise as cleanly.

Anchor some writing in class. One timed, unaided piece early in a course gives you a baseline for every later submission. It is also the single most useful diagnostic you will get, detection aside.

Set tasks AI handles badly. Writing that draws on a specific class discussion, a local context, a personal experience or a text only your group has read is far harder to generate convincingly. This is a better use of your planning time than policing.

Talk to the student. Two minutes asking someone to explain a choice they made in their own essay resolves most cases, in either direction. It is also fairer, because it gives the learner a chance to demonstrate authorship rather than disprove a number.

Student writing by hand in class as part of an unaided baseline writing task

If you think a student has used AI

Treat it as a conversation about learning, not an investigation. Open with curiosity rather than accusation: ask how they approached the task, what they found difficult, why they made a particular choice. Keep a detector score out of the conversation entirely. Bring process evidence if you have it. Follow your institution’s academic integrity procedure rather than improvising, and remember that a false accusation against a language learner can end a course, a visa or a career.

The asymmetry matters. Missing a case of AI use costs a grade. Wrongly accusing a learner costs far more, and the research says you are most likely to do it to the students already at the greatest disadvantage.

Where this is heading

Detection is unlikely to get decisively better. As models improve, generated text becomes more varied and harder to distinguish, while the detectors chase a moving target. Institutions are already shifting emphasis from catching AI use toward designing assessment that tolerates it, and major assessment bodies have moved the same way: Cambridge University Press and Assessment’s principles for AI-powered marking insist on humans remaining in the loop for any high-stakes judgement.

For English language teaching specifically, the useful question is no longer whether a student used AI. It is whether they can do the thing unaided when it counts, and whether your assessment design actually tells you that.

The rest of this guide

Read next

This guide is part of our wider work on AI in the classroom. For the full picture, see our guide to AI in English language teaching, our roundup of the best AI tools for English teachers, and the research behind these figures in our report on the future of AI in English language teaching.

If marking is where your hours go, our free lesson plan generator and the wider tool suite are built to give that time back legitimately.

Frequently asked questions

Can AI detectors actually prove a student used AI?

No. Every detector produces a probability score based on how predictable the writing is, not evidence of authorship. No current technology can prove a specific person did or did not write a specific text, which is why detector output should never be the basis of an academic integrity allegation.

Why do AI detectors flag non-native English speakers so often?

Detectors measure how varied and unpredictable writing is. Language learners typically use a smaller vocabulary and more familiar sentence patterns, producing even, predictable prose that resembles generated text. A Stanford-linked study found detectors wrongly flagged over 61% of TOEFL essays by non-native speakers.

What AI detector do most teachers use?

In universities and colleges it is usually Turnitin, because it is built into the submission system rather than chosen by the teacher. Individual teachers more often use GPTZero or free web checkers such as QuillBot and Scribbr. Availability drives that choice more than accuracy does.

Is Turnitin’s AI score reliable?

Turnitin publishes its own false-positive rate and advises that the score alone is not proof of misconduct and should not trigger an allegation by itself. Treat it as a signal to look at process evidence and speak to the student, not as a finding.

What should I do instead of using a detector?

Collect process evidence. Ask for outlines and drafts, use document version history, set one unaided in-class writing task early to establish a baseline, and design tasks around class-specific content AI cannot easily generate. A short conversation with the student resolves most cases either way.

Can students get around AI detectors?

Yes, easily, which is another reason not to rely on them. Paraphrasing tools, manual editing and simple prompting changes all reduce detection scores. A detector mainly catches unedited output, which means it disproportionately catches the least sophisticated attempts rather than the most serious ones.

Quick favour: how do you actually use AI in your teaching?

We are running the first real study of AI use among TEFL teachers. 12 questions, about 3 minutes, fully anonymous. Leave your email at the end and you will get the report before anyone else.

Take the 3-minute survey

I

Written by ianadmin

Part of the TEFL Institute Group team behind tefl.ai.

Shopping Basket
Accredited & Approved
Highfield Qualifications, Approved Centre 21335 Approved CentreNumber: 21335 OTCAC — Accreditation Number 20260330 OTCAC AccreditedNumber: 20260330