Most of the world writes English as a second language. When those writers turn in an essay, a cover letter, or a journal submission, an automated detector increasingly stands between their words and the reader. The promise is neutrality: a number that says “human” or “AI,” the same for everyone. The reality is that the number is not the same for everyone. It is systematically harsher on people who learned English later.
What a detector actually measures
Most AI-text detectors don’t understand a sentence the way a person does. They score surface statistics — how predictable the next word is (often called perplexity), how much that predictability varies (burstiness), how common the vocabulary is, how uniform the sentence structure is. Text that is smooth, common, and even-paced reads as “machine-like.” Text that is surprising and uneven reads as “human.”
Here is the problem. Skilled second-language writing tends to be clear, regular, and built from high-frequency words — exactly the qualities a careful learner is taught to aim for. The same traits that make ESL prose clear make it look synthetic to a perplexity-based detector. The detector isn’t catching a machine. It’s catching an accent.
The mathematical layer: why the false positives don’t go away
A detector picks a threshold: above it, “AI”; below it, “human.” Move the threshold to catch more AI text, and you sweep up more human text with it. Move it the other way to spare human text, and more AI text slips through. Every detector lives on this trade-off curve.
Garland (2026) makes the structural point: because second-language English overlaps the statistical region the detector associates with machine text, there is no threshold that both clears ESL writing and still flags AI output. Reducing false positives for native writers pushes them up for non-native writers. It’s a property of where the two distributions sit, not a tuning mistake — which is why “we improved our model” doesn’t dissolve it. The bias is in the method, so it survives better models.
The empirical layer: the studies agree
In 2023, researchers at Stanford (Liang and colleagues) ran a set of widely used detectors over essays written by non-native English speakers — TOEFL essays — and over essays by native US students. The detectors flagged a majority of the non-native essays as AI-generated while clearing nearly all of the native ones. Same task, sharply different verdicts, split along first-language lines.
Follow-up work has reproduced the direction of the effect. A 2025 study out of Maryland, and others across different detectors and datasets, keep landing on the same finding: non-native writing draws disproportionate false positives. No single paper is the whole argument — but a stack of them, using different detectors and different corpora, pointing the same way, is hard to wave off as noise.
The real-world layer: the rate the brochure leaves out
Detector vendors typically advertise a false-positive rate under 1%. On second-language English the published figures are of a different order: in the Stanford study, seven detectors misclassified 61.22% of TOEFL essays written by non-native speakers as AI-generated, all seven agreed on 19.8% of them, and 97.8% were flagged by at least one (Liang et al., 2023). Turnitin itself states an uncertainty band of roughly ±15 percentage points and that its indicator should not be the sole basis for punitive action. That gap is the whole story. The difference between the advertised number and the real one is measured in people who have to defend work they did honestly.
And the burden lands unevenly. A native speaker who gets flagged is an anomaly the system treats as a glitch. A non-native speaker who gets flagged is fighting a prior — the quiet assumption that their fluent English “came from somewhere.” Same flag, heavier cost.
The human layer: flagxiety
Numbers undersell what this feels like. Students report anxiety about being falsely accused by AI detection — a feeling colloquially called “flagxiety” — and it is reported more often by international students, who are also the group the published false-positive figures describe. The damage starts before any accusation. Writers begin sanding down their own clearest sentences, adding deliberate roughness, second-guessing the very fluency they worked years to earn, all to look “human enough” for a machine. That is a tax on honesty, paid in advance, by the people who can least afford it.
What Writence does about it
You cannot prove a negative to a black box. Arguing with a detector’s score is a losing game — there’s no transcript to point to, no record of how the words actually came to be. So we built the inverse of a detector.
As you write in our editor, with your consent, it records an append-only chain of edits — every meaningful change, in order — and signs it with ed25519 cryptographic keys. The result is an Authorship Certificate: a publicly verifiable record of the writing process that anyone can check independently, without an account and without trusting us.
We are deliberate about what this is and isn’t. The certificate documents the process by which a piece of writing was made. It is not a legal verdict, it does not prove innocence, and it is not a guarantee against any detector. We will never tell you it makes your text “pass.” What it offers is fairer ground: instead of being asked to disprove a number, you can show the work.
The honest bottom line
Detectors aren’t worthless, and the people building them aren’t villains. The narrow, well-supported claim is this: on second-language English, current detectors are wrong often enough — for reasons baked into how they work — that their output should never stand alone as proof. The fair response isn’t a better accusation engine. It’s giving writers a checkable record of their own process. That’s the company we set out to build.
References
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779. doi:10.1016/j.patter.2023.100779 — 61.22% of TOEFL essays by non-native writers misclassified as AI-generated; 19.8% flagged unanimously by all seven detectors; 97.8% flagged by at least one.
Turnitin. AI writing detection guidance for administrators and educators. The company states an uncertainty band of roughly ±15 percentage points depending on text length, and that the AI indicator “should not be the sole basis for punitive action.”
Vanderbilt University, Office of the Provost (2023). Guidance on AI detection and why we’re disabling Turnitin’s AI detector. One of the first US universities to switch the detector off institution-wide.
Office of the Independent Adjudicator for Higher Education (UK, 2025). Published casework on academic-misconduct complaints involving AI-detection evidence.
Garland (2026) on the structural trade-off between false positives on native and non-native writing. Cited from the secondary literature; we have not independently reproduced the result.
Sources are summarized for a general audience; study findings and reported false-positive ranges vary by detector, dataset, and methodology. This article describes research and design intent — it is not a guarantee about any third-party detection product.