We tested GPTZero on 200 student texts. Learners weren't flagged. AI-polished writing was.

Most articles about AI detectors and non-native writers, including ours, lean on one study from 2023. Detectors have changed a lot since then, so on 8 October 2026 we ran our own test: we sent student writing that was certainly written by people, because it was written before ChatGPT existed, to GPTZero's API, along with AI written texts and edited versions of the student texts. We did not expect the result.

The results in one table

In short: GPTZero did not flag a single human-written text, whether the writer was a learner or not, and whatever its length. It caught every AI-written control. And it labelled every learner text that had been fully polished by AI as AI-assisted, even though the ideas, examples and structure were the student's own.

Test 1: English learners vs native speakers

We took 80 essays by students recorded as English language learners and 80 essays by other students from the PERSUADE 2.0 corpus, matched on writing prompt so the topic couldn't explain a difference. The essays were written by US students in grades 6–12 before generative AI was widely available. We scored the first 300 words or so of each essay, cut at a sentence boundary.

All 160 were classified as human, with an average AI probability of 0.0% in both groups. The learners' essays received lower writing scores on average (2.8 vs 4.0 on the corpus's 6-point scale), so weaker or less fluent English did not lead to a flag either.

Test 2: short texts

Detectors are often said to struggle with short texts, so we took 40 passages that learners wrote in their own questions on English Language Learners Stack Exchange before ChatGPT was released, with quotations, code and example sentences removed. We scored the first 50, 100 and 150 words of each, and did the same with 10 texts written by an AI model in a typical assistant style.

Every learner passage was classified as human at every length, including passages of around 40 words. Every AI text was classified as AI at every length. In this test, length made no difference.

Test 3: what happens when you edit your writing

Many people who write in a second language run their text through a grammar checker or ask an AI to improve it. We took 19 of the learner passages and made three edited versions of each:

• Automatic grammar fixes: every suggestion from a rule-based checker (LanguageTool) accepted.
• Light AI correction: grammar, articles and word choice fixed, the writer's sentences kept.
• Full AI polish: rewritten to sound fluent and academic, the way an assistant responds to “improve my English”. Same ideas and examples, new wording.

Automatic fixes and the light correction stayed human in all 19 cases. The full polish was classified as mixed (AI-assisted) in all 19, with an average mixed probability of 100% and nearly every sentence highlighted as AI.

What this test can't tell you

• One detector, one date. We tested GPTZero's API model version 2026-10-07-base. Turnitin, Copyleaks, ZeroGPT and others may behave differently, and detectors change often.
• Possibly familiar texts. Both sources have been public for years and may be in detectors' training data as human examples.
• Small samples. Zero flags in 299 human scans doesn't mean a false positive rate of zero; it means it was rare enough not to show up here.
• Our own edits. The AI corrections and polishes were produced with an AI model by us, to a fixed instruction. Other tools and instructions will produce different text.
• It doesn't overturn the 2023 study. Liang et al. tested different detectors, at a different time, on TOEFL essays. Both results can be true.

What it means for your writing

1. Your own English is not the problem. In our test, imperfect learner writing was read as human every time.
2. Fix errors, don't replace your voice. Correcting grammar kept texts human; letting an AI rewrite everything did not. If you use a tool, use one that changes only what is wrong and shows you every change.
3. Keep your process evidence. Version history and drafts settle disputes that no score can. Our guide for students accused of using AI explains what to gather.

ESL Humanizer is built on that idea: it fixes the English in text you wrote, keeps your wording where it already works, shows each change, and reports how a detector reads the result before you submit.

Method and data

• Detector: GPTZero API (/v2/predict/text), model version 2026-10-07-base, 8 October 2026. A text counts as flagged when GPTZero's predicted class is “ai” or “mixed”.
• Test 1: 80 essays by students marked as English language learners and 80 by other students, drawn at random from essays of 150+ words in the PERSUADE 2.0 training set and matched by prompt; first ~300 words.
• Test 2: 40 random passages of 155+ words from English Language Learners Stack Exchange questions posted before 30 November 2022, with quotations, code, lists and links removed; scored at ~50, ~100 and ~150 words, cut at sentence boundaries. One canonical post by a native-speaking moderator was excluded from Test 3.
• Test 3: 19 of the Test 2 passages at ~150 words, in four versions, all scored in the same run.
• Data: every score, with the source ID of each text (CSV). We don't republish the texts themselves; the PERSUADE essays are available from their authors.