Are AI checkers fair when English is a second language?

For teachers who mark written work

If you act on an AI checker’s number alone, the students most likely to pay for it are the ones who learned English at school and write it with great care. Their essays can come back marked as AI when every sentence is their own, and a wrong charge can cost them a grade, their place in the course and their trust in you. So if you have been asking whether AI detectors flag international students, what matters is not where a student comes from but whether English is a language they learned. For students whose first language is not English, the answer is yes: on every checker, ours included, their writing is the most widely known case of human work wrongly marked as AI. A checker is still of some use. But on such an essay its result deserves less of your weight, and your talk with the student deserves more.

What the 2023 study tested

In 2023, researchers at Stanford University put seven widely used AI detectors to a test. They gave them two sets of essays. One set held 91 essays from TOEFL, an English test, written by people whose first language is not English. The other held 88 essays by eighth-grade students in the United States.

On the test essays, the detectors were wrong more than half the time: on average they marked 61.3% of them as AI. The essays by the United States students were classified accurately.

It looks worse when you count how often they agreed. For 19.8% of the test essays, all seven detectors agreed in marking them as AI. At least one detector marked 97.8% of them. Almost every essay in that set, then, was called AI by one detector or another.

What one study can and cannot settle

The study shows a clear gap between two groups of writers on seven detectors, as those detectors were in 2023. It can’t tell you how the checker your school uses today would score the same essays. Our own view, which we hold for every checker and for ScanForAI as well, is that formal writing, and writing by people using English as their second language, still score higher than relaxed writing by people using their first language. You can read how far our own results can be trusted on a page of its own.

The study does leave you a fair question to ask. Before your school renews its checker, someone can write to the seller and ask how the tool scores formal essays written entirely by students whose first language isn’t English, next to relaxed essays by students who grew up with English. A seller that can’t answer in plain words has told you how far to rely on its numbers for those students.

Some questions no study can answer for you. Did this student use AI for this essay? Research on groups of writers describes patterns across many texts, and it has nothing to say about the one essay on your desk. Has a newer checker fixed the problem for good? One test could not show that. And even a checker with very little bias would still not prove who wrote anything, because no checker can.

Why second-language writing looks regular

English learned at school often follows set patterns. The essays tend to have sentences of similar length, formal words, few forms like “don’t” or “it’s”, and phrases a reader can see coming. AI text often has the same features on the surface, for quite different reasons.

Some checkers measure writing with numbers, and to such a checker careful, even and formal writing sits very close to machine-written text. The checker only has the sentences on the screen. It can’t tell whether their regular shape came from years of grammar lessons or from a prompt typed into a chatbot, and both can produce sentences that look much alike.

Then there’s the task. A lab report or a literature review asks everyone in the class for formal English, so when a second-language student writes one, the task and their habits both push it toward the style checkers score higher.

So a wrong flag here doesn’t mean the student was careless or dishonest. It comes from what the checker measures. The students at most risk can be the ones who worked hardest to write correctly, so when an essay seems almost too tidy, it may simply be written the way its author was taught.

Extra care before you act on a score

With students who write in a learned language, a high result counts for even less than usual, and the conversation counts for even more. Since no checker can prove authorship, a score alone should never be the proof, and it can’t carry a grade or a penalty by itself either. Before you take a case further, it’s worth going through these checks.

  • Find out which checker produced the result and what its number stands for, since each checker means something different by its score.
  • Read the confidence the checker reports, if it reports one, and the reason given with it. Every ScanForAI result shows a confidence of high, moderate or low with a one-line reason.
  • Compare the essay with writing the same student produced in class, under time pressure, or earlier in the course.
  • Ask whether the task itself calls for formal English, as a report or a review does.
  • Ask the student for their drafts, sources and planning notes, and for a short account of how they built the essay.

We explain how checkers read second-language writing, ours included, and why the bias stays on every one of them. If you scan a sample on ScanForAI, choose a text type that fits the work and use the marked lines, which come with their reasons, to open a talk with the student, never to decide the case.

Turn each marked passage into a question for that talk, and ask about drafts, sources and how the work grew. If an essay has already come back with a high result, the steps for a teacher whose student’s essay was flagged put the next moves in order. For the meeting itself, read the guide to talking with a student you think used AI, which covers how to open, what to ask and how to listen.

For students whose first language is not English, two things make the meeting fairer. Send your questions in writing a day ahead, so they can prepare in the language they think in. And don’t hold their spoken English up against their written English. At a desk, with time, a dictionary and a chance to fix mistakes, many people write far more carefully than they speak, above all in a language they learned. Someone who stumbles over a word in the meeting may still have chosen it well on paper.

What to tell your students

Students who learned English at school may already have heard that checkers are hard on them, and may wonder about it before your first assignment is even set. So bring it up yourself in the first week, while nobody’s work has been scanned yet. If your course uses a checker, you could say something like this: “Some essays in this course will be scanned by an AI checker. A result that raises a question leads to a talk with you, and nothing is decided before that talk. Write as carefully and formally as you like, because that will never count against you on its own.” If your course doesn’t use one, say that too.

While you have their attention, ask them to write each essay in a file that keeps its own history. A Google Docs document does this, and so does a Microsoft Word file stored on OneDrive. Plans, outlines and earlier drafts are worth keeping too, with any notes they made in their first language. Nobody expects to need them. But if a question ever comes up about a piece of work, those files will support the student better than anything they can say in a meeting.

A student may ask you whether they should start writing differently because of all this. The honest answer is no. You’re marking their English as they learned it, and a text bent out of shape for a tool wouldn’t show you anything more about who made it.

Put a line on the course page, too, for any student who feels a result was used unfairly. Your own name and email go first, then the course leader’s, then the student support office. A worried student will write to an address they can already see far sooner than they will go looking for one.

The seven detectors in the 2023 study were widely used ones, and on average they marked most of those test essays as AI. Would the checker in your own course do any better with the students you teach, and how would you find out?

Sources

  1. GPT detectors are biased against non-native English writers. Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu and James Zou, Patterns 4(7), 2023.
Questions

Frequently asked questions

Should I scan essays by second-language students less often, or not at all?

No. Scanning some students’ work and not others would single them out, and that is unfair in its own way. Keep one rule for the whole class. What changes is what you read next to the result: for these students, the pieces they wrote in class and the drafts they kept will often tell you more than the number does.

Should I mention the bias when I meet the student?

Yes, near the start, before you get to the essay itself. Explain that checkers tend to score careful, formal English as more machine-like whoever wrote it, and that you won’t decide anything until you’ve heard them out. Otherwise they may sit through the whole meeting waiting for the accusation, when you’d much rather hear how they wrote it.

Does a low result mean a second-language essay is fine?

Not on its own. A low result only means the checker found little to mark. Suppose the essay reads quite unlike anything the student has handed in before. The low number doesn’t answer that, just as a high number wouldn’t have proved anything either.

What should I keep from each case from now on?

Keep a folder for every case where a checker’s result played any role at all, and name it so you can find it again, for instance with the course code and the date. It should hold the essay exactly as it was handed in, the full result, the checker’s name and the day of the scan. If you used ScanForAI, also note which text type you chose and the confidence with its one-line reason. Write your notes from the conversation the same day, while you still remember it: what you asked, what the student said, which drafts they showed you. Add copies of those drafts, then a sentence or two on what your decision rested on. Will anyone ever read it? Perhaps only you, at the end of term, when the folders side by side show whether results on second-language writing have been steering your doubts. But if a student appeals, a colleague or a panel can follow your reasoning from the folder, and you won’t have to rely on memory.

This guide is general information for students and teachers. It is not legal advice, and your school’s own rules come first. A score from any AI checker is a signal, never proof.

Cite this guide

ScanForAI team (2026). Are AI checkers fair when English is a second language? ScanForAI. https://scanforai.com/guides/are-ai-checkers-fair-in-a-second-language

You may share and adapt this guide if you credit ScanForAI and link to it. See Reuse and citation.

AI detection is not 100% accurate. No detector, including ScanForAI, can prove if a text was written by a person or by AI. Every result is an estimate based on writing patterns. It can be wrong in both directions. Never use it as the only reason for an important decision about a person, such as a school decision, a punishment or a hiring choice. Never use it to accuse someone.