Why AI Detectors Flag Human Writing

Loch Ness · HUMAN WRITING · AI detector false positives

Part of the guide: Do AI Detectors Actually Work? What the Research Says

By Camille G., founder of Loch Ness · Published September 29, 2026 · Last updated September 29, 2026

Short answer: most detectors don’t recognize AI at all. They measure how predictable a text is, and plenty of real people write predictably. People writing in a second language, students following a template, anyone who writes plainly on purpose: they’re the ones who get flagged. Here’s how it happens, and why it isn’t your fault.

Human Origin Label · For writers A detector flagged you? Your process is the answer. Show your drafts and version history once, live or in a screen recording, and get a public page that says what was checked. Free during the pilot. Get verified, it’s free → Not ready for a review? Get the free self-declared badge →

How most detectors decide

Language models tend to choose likely words. So many detectors score a text on how surprising its word choices are, a measure researchers call perplexity. Low surprise reads as “machine-like”. The trouble is that this measures style, not authorship. In the Stanford study of seven popular detectors, the human essays that every single detector flagged were exactly the ones with the lowest perplexity (Liang et al., 2023).

Who gets flagged most

• People writing in a second language. The same study ran 91 TOEFL essays through seven detectors: on average, 61% were wrongly labeled as AI-generated, and 89 of the 91 were flagged by at least one detector. US eighth-graders’ essays were classified almost perfectly (Patterns).
• Short texts. OpenAI said its own classifier was unreliable on short inputs and got better as texts got longer, before withdrawing it for low accuracy (OpenAI). A cover letter or a product description gives a detector very little to work with.
• Formulaic writing. Templates, standard structures and the kind of neutral prose that schools and businesses often ask for are predictable by design.

The same trick works in reverse

When the Stanford researchers asked ChatGPT to rewrite those human essays with more sophisticated vocabulary, the detectors started calling them human (ScienceDaily). A wider test of 14 tools found the same weakness: accuracy dropped further once AI text was edited, paraphrased or machine-translated (Weber-Wulff et al., 2023). So a detector score says more about how a text sounds than about who wrote it.

It costs people real things

This isn’t just an academic problem. Students have faced integrity procedures over detector scores, and in June 2024 Gizmodo reported freelance writers and journalists losing work after detection software wrongly classified their writing as AI-generated (summary on Wikipedia).

If a detector flagged your writing

Nothing about your writing needs fixing. What helps is evidence of how you wrote it: the version history of your document, your drafts, your notes. A percentage can’t compete with a timeline of your work. I’ve put the full method in how to prove you wrote it yourself.

And please don’t start writing “fancier” to please a detector. As the research shows, that’s exactly what AI rewriting does to slip past them. Your plain voice is fine.

Tired of being judged by a score? Let someone look at how you actually write. See how the Human Origin review works →

Back to the overview: do AI detectors actually work?.

Full disclosure: Human Origin, the label mentioned on this page, is something I created. I’m flagging it so you can weigh what I say about it. The facts above rest on the sources below.

Sources

• Liang et al. — GPT detectors are biased against non-native English writers, Patterns (2023)
• Liang et al. — same study, full text on arXiv
• ScienceDaily — GPT detectors can be biased against non-native English writers (July 2023)
• OpenAI — New AI classifier for indicating AI-written text (with the July 2023 discontinuation note)
• Weber-Wulff et al. — Testing of detection tools for AI-generated text (2023), via ERIC
• Wikipedia — Artificial intelligence content detection (reporting on freelance job losses)

Update log

September 29, 2026 — First publication.