Short answer

AI detectors don't know who wrote a text. They estimate how predictable it is. Plain, formal or non-native English often looks predictable, so it gets flagged: a 2026 study found commercial detectors flagged 9 to 15% of genuine 2023 to 2025 abstracts, and a 2023 Stanford study found detectors labeled 61% of TOEFL essays as AI. Humanizer tools beat detectors by making text worse, and the same 2026 study concluded detector scores shouldn't be used alone as evidence of misconduct.

Searches for GPTZero, Turnitin's AI checker, ZeroGPT and "AI humanizer" are all climbing worldwide on Google Trends this month, with tools like Clever AI Humanizer among the breakout queries. Two groups are doing most of that searching: people trying to hide AI use, and people who wrote every word themselves and got flagged anyway.

This piece is for the second group, and for teachers and editors who have to read the scores. You'll see what detectors actually measure, try a pattern highlighter on your own text, check your false-positive risk, and get a short plan for what to do if you're accused of something you didn't do.

What an AI detector actually measures

Most detectors combine some version of three signals:

  • Perplexity: how surprised a language model is by each next word. AI text tends to pick likely words, so it scores low. So does careful, simple, rule-following human writing.
  • Burstiness: how much sentence length and structure vary. People tend to mix long and short sentences; models tend to keep an even rhythm.
  • A trained classifier: a model trained on examples labeled human or AI, which learns surface patterns such as favorite words and phrasing.

None of these sees authorship. They see style. That's why the 2026 study by Karr, Khvatskii, Hua and Chawla found that high scores tracked long words and academic vocabulary density, not intent.

Diagram of how AI detectors score text: perplexity, burstiness and a trained classifier combine into a probability, with notes on why formal and non-native writing scores as AI
Three signals, one probability. Nothing in the pipeline knows who typed the words.

Try it: the pattern highlighter

Paste a paragraph you wrote. This tool highlights words and phrases that people associate with AI writing, and measures how much your sentence length varies. It is not a detector and doesn't guess authorship. It shows you the surface patterns detectors and suspicious readers react to. Your text stays in your browser.

Pattern highlighter

What a suspicious reader would notice

Paste text and press the button. The sample above is deliberately full of patterns.

If the sample lights up and your own writing doesn't, good. If your own writing lights up, the fix isn't a humanizer. It's writing the way you talk: concrete nouns, your own examples, sentences of different lengths, and fewer stock transitions like "moreover" and "ultimately."

How often detectors get it wrong

Detector companies tend to advertise false-positive rates below 1%. Independent tests keep finding more.

Bar chart of reported AI detector false positives: 61 percent of TOEFL essays in a 2023 Stanford study, 9 to 15 percent of real academic abstracts in a 2026 study, compared with vendor claims under 1 percent
Different studies, different texts, different detectors. The pattern is consistent: formal and non-native writing gets flagged far more than vendors suggest.
  • Stanford, 2023: Weixin Liang, James Zou and colleagues ran seven detectors on 91 TOEFL essays written by people. The average false-positive rate was 61.22%, and 89 of the 91 essays were flagged by at least one detector. Essays by US eighth-graders were classified almost perfectly. The authors blamed perplexity: non-native writers use a narrower vocabulary, which reads as predictable.
  • Karr, Khvatskii, Hua and Chawla, 2026 (preprint): commercial detectors flagged 9 to 15% of unmodified 2023 to 2025 abstracts, with non-STEM fields flagged far more than STEM. Abstracts that people had only lightly refined with AI were flagged 38 to 80% of the time.
  • Turnitin's own position: the company describes its AI score as a starting point for a conversation between instructor and student, not proof, and acknowledges it can produce false positives.

Check your false-positive risk

Tick anything that applies to the text in question. The more boxes, the more likely a detector flags genuine work.

Risk check

Will a detector misread you?

Three or more: keep your drafts and version history from now on. You may need them.

Why humanizers aren't the answer

Humanizer tools rewrite text to raise perplexity and burstiness: unusual word choices, shuffled sentence lengths, swapped synonyms. Against detectors, they work. The 2026 study found that after one popular humanizer, under 4% of AI rewrites were still flagged.

That result is the problem, not the solution. The study's authors drew the uncomfortable conclusion directly: honest AI editing carries a higher risk of sanction than humanizer-assisted evasion. Meanwhile, the humanized text usually reads worse, with odd synonyms and broken rhythm that a human reader notices even when software doesn't. And if your institution bans undisclosed AI use, a humanizer doesn't change what you did; it only hides it.

A detector score measures style. A humanizer changes style. Neither one tells you who did the thinking.

Myth or fact?

6 questions

Do you know how detectors really work?

"AI detectors can tell who wrote a text."

"Non-native English writers get flagged more often."

"A 0% AI score proves a human wrote it."

"Lightly editing your own text with AI can get it flagged."

"Turnitin says its AI score is proof of misconduct."

"Your draft history can help if you're falsely accused."

Falsely accused? Do this, in this order

  1. Stay calm and ask for the evidence in writing: which tool, which score, which passages.
  2. Collect your process: Google Docs or Word version history, earlier drafts, notes, outlines, browser history of your research, and timestamps.
  3. Offer to explain your work out loud. Walk through your argument and sources. Someone who did the thinking can do this easily.
  4. Point to the research: the Stanford study on non-native writers and the 2026 study's conclusion that detector scores shouldn't be used alone as evidence.
  5. Ask what the policy actually says. Many schools allow some AI help, such as grammar checks, if you disclose it.

For teachers and editors

Treat a detector score like a smoke alarm, not a verdict. Look at the writing process instead: drafts, a short conversation about the work, or an in-class writing sample. Be extra careful with students who write in a second language, since the evidence says they bear most of the false positives.

FAQ

Is GPTZero accurate?

It catches unedited AI text often, but independent tests find meaningful false positives on human writing, especially formal and non-native English. Use it as one signal, never as proof.

Can Turnitin detect humanized AI text?

Not reliably. The 2026 study found that humanized AI rewrites were detected less than 4% of the time by commercial detectors.

Why does my own writing get flagged as AI?

Usually because it's formal, uses common academic phrases, keeps an even sentence rhythm, or was polished with a grammar tool. Detectors read those as predictable.

Should I use an AI humanizer?

If you wrote it yourself, you don't need one; keep your drafts instead. If you used AI where it isn't allowed, a humanizer hides it without fixing it, and it often makes the writing worse.

Sources

Last checked November 10, 2026.