AI Detectors and Non-Native English Writers: The Evidence

👁 130 reads

If English is an additional language for you, an AI accusation can create a difficult double bind: write clearly, then explain why your clear writing looks suspicious. Research has documented unequal false-positive rates in some English datasets. It does not establish that every detector treats every multilingual writer unfairly.

The often-quoted Stanford study from 2023 is an important starting point. Newer work helps define where its conclusions travel and where they do not. Students, instructors and authors need both parts of that evidence, especially when a detector score could affect a grade or publication decision.

Sources checked October 14, 2026, by the Quantslant editorial team. This is a research-based guide, not a new detector experiment. Our article on the French novel and AI accusation controversy covers the separate news story.

What the Stanford study found

Liang and colleagues’ study, first released April 6, 2023, evaluated seven detectors on 91 TOEFL essays and 88 US eighth-grade essays. The TOEFL set had an average false-positive rate of 61.22%, with much stronger results on the comparison set. These were specific tools and samples, not a current estimate for every English learner.

The groups also differed in their writing context. That matters when interpreting the size and cause of an observed difference. The finding supports concern and further evaluation; it does not assign a personal risk percentage to every international student.

When reading a headline about “61% of non-native writing,” ask what the denominator actually was. Were researchers counting essays, passages, tool decisions or people? Without that distinction, a useful research result can turn into a claim much broader than the experiment.

Why predictable language can be mistaken for generated writing

Stanford HAI’s May 15, 2023 explanation connected the issue to linguistic predictability, including perplexity. More constrained vocabulary and sentence patterns can overlap with features associated with generated text. That provides a possible mechanism for the observed disparity, not an individual authorship test.

Think of a writer choosing a familiar phrase because it communicates the intended meaning precisely. That choice tells you little about where the argument came from. A conventional introduction can lead into original reasoning, while an unusual metaphor does not establish an independent writing process.

Perplexity measures predictability relative to a reference language model. It is not a measure of intelligence, honesty or the quality of an idea. Modern detection systems also should not all be reduced to a single perplexity rule.

What newer research changes

Al Ali, Helcl and Libovický’s March 2026 study revisited the issue in Czech and found no systematic non-native disadvantage in its tested conditions. It also demonstrated that the evaluated approaches need not depend on perplexity. The findings limit universal claims about language background and detection.

That is not a reason to dismiss the earlier English evidence. It is a reason to specify the language, dataset, detector and date whenever making a claim. Better performance in one setting does not guarantee equal treatment in another.

The practical position is therefore narrower than “all detectors are biased” or “the problem has been fixed.” A particular system needs relevant evidence for the population and task where it will be used. Its overall score cannot answer every fairness question.

Multilingual detection is a separate research problem

MULTITuDE, published in December 2023, introduced 74,081 human and generated texts across eleven languages to study multilingual detection. It examined transfer across languages and generators. That is a different research question from estimating how often a particular university’s multilingual students will be falsely accused.

Do not treat “supports English” as proof that a detector has been validated equally across all forms of English writing. Equally, do not infer poor performance solely from a student’s nationality. The necessary question concerns the actual evaluation evidence, not a presumed identity category.

A multilingual writer may be highly proficient in English, may have learned academic writing in several countries, or may use different languages at different drafting stages. For a review of an individual submission, ask about that process rather than guessing it from a name or affiliation.

Proofreading, translation and generation need different descriptions

Imagine three hypothetical submissions. One student writes an essay independently in English. Another writes the analysis and uses a tool to revise the language. A third asks a system to supply the analysis itself. A similar detector score would not make those processes equivalent.

Victor Angelier’s September 18, 2026 review distinguishes genuine false positives from boundary cases involving AI translation or polishing. For these mixed workflows, the relevant questions include what changed, which contribution remained the writer’s, and what the rules allowed.

We recommend describing assistance by action rather than by product name. “Checked spelling” is different from “accepted rewritten paragraphs.” “Translated my own draft” is different from “generated supporting arguments.” An accurate description helps a reviewer identify the issue without assuming that every tool feature is interchangeable.

For authors, our guide to writing contest and publisher AI policies shows why permission cannot be assumed. What a classroom accepts may differ from what a journal, contest or book publisher requires.

What students should ask before using language assistance

Ask for guidance early and make the question specific. Instead of “Can I use AI?”, describe the intended function, the material it will see and the output you plan to retain. An instructor can then respond to a defined workflow.

• May I use a spelling or grammar checker on my own draft?
• May I accept replacement sentences, or only review flagged errors?
• Does the assignment permit translation of my own notes or prose?
• What disclosure is required when submitting the work?
• Should I retain an earlier version or particular process materials?

These are suggested questions, not statements of universal university policy. Save the response with the assignment instructions. If the task is intended to assess unaided language production, clarify how that purpose affects the allowed support.

Keep a useful record without turning writing into surveillance

You do not need to manufacture an elaborate archive. Our suggested approach is to retain the ordinary materials that explain your work: an outline, source notes, a draft before major revision and comments received during review.

Where you move between languages, identify those stages. For example, keep the notes in their original language alongside the English draft, if that reflects your actual process. Do not retrospectively create a supposed original to strengthen a defense.

Above all, do not add mistakes or replace precise words merely to change a score. A detector should not become your style guide. Your purpose is to write clearly, meet the assignment’s requirements and give an honest account of the process.

How to handle an accusation fairly

If your work is flagged, request the report and identify the specific alleged violation. Preserve the submitted file. Then connect your available records to the passages in question, including any authorized language assistance.

For a meeting, prepare to explain decisions: why a source mattered, how a paragraph changed or why you chose a particular example. If speaking spontaneously is difficult, ask whether you can supplement the discussion with written explanations and existing records.

Our guide to AI detector false positives and student appeals discusses documented disputes. Treat those examples as guidance about evidence and procedure, not as automatic proof that your own detector result is wrong.

What instructors and editors can do differently

We recommend separating three assessments: the quality of the ideas, the quality of the language and compliance with the permitted process. Their relative importance may change with the task, but the criteria should be clear before submission.

• Explain allowed functions with examples rather than relying only on brand names.
• Use a short disclosure format that writers can complete accurately.
• Request intermediate work when it supports learning or editorial review.
• Give the writer a meaningful opportunity to address the concern.
• Do not equate an accent, nervousness or slower speech with lack of authorship.

This approach leaves room to investigate genuine misconduct. It simply keeps the investigation focused on relevant evidence. A writer’s language background can provide context, but it should not become a substitute for establishing what happened.

Related: my test of the Orélien novel in three detectors, what to do if you are falsely accused, AI Detector False Positives: Cases and Student Appeals.

FAQ

Are AI detectors biased against every non-native writer?

The evidence does not support that universal statement. Some English datasets show substantial disparities, while newer findings demonstrate that the pattern does not automatically extend to every language.

Is the Stanford figure a current false-positive rate?

No. It belongs to the study’s particular samples and tools. It should not be presented as the present-day error rate for all detectors or all international students.

Does simple English prove that a text is AI-generated?

No. Vocabulary and sentence structure do not independently establish authorship. Evaluate the actual writing process and the evidence relevant to the task.

Is AI proofreading automatically academic misconduct?

It depends on the rules and the assistance involved. Identifying errors, rewriting prose and generating reasoning require different descriptions. Get clarification before relying on an ambiguous permission.

Should I disclose that English is an additional language?

You can explain your language background where it helps clarify the writing process. It supplies context; it should not replace evidence about the submitted work or become a basis for suspicion.