Why AI detectors flag human writing

A detector is a model trained on examples of human writing and of model writing, and taught to tell them apart. What it learns is the texture of each: how predictable the next word is, how even the sentences are, which phrasings turn up. Given a new piece, it reports how far that texture resembles a model’s, as a percentage or a label.

That number is an estimate about the words on the page. It cannot see the drafts, the deleted paragraphs, the afternoon it took, or the person at the keyboard, because none of that is in the finished text.

Writing that is predictable in the way a model’s is: plain, even, carefully edited, or written to a tight format. Some writers are flagged far more than others.

• –People writing in a second language. A Stanford study of several widely used detectors found they “consistently misclassify non-native English writing samples as AI-generated”, while identifying native writing correctly: across seven detectors, 61% of essays by non-native speakers were wrongly flagged, against about 5% of essays by native speakers.
• –Formal and formulaic writing: lab reports, cover letters, summaries, anything that follows a template, where there are only so many ways to say a thing.
• –Short pieces, where there is little text to judge from.
• –Writing that has been edited hard, by its writer or by a grammar tool, until it reads smoothly.

Liang and others, GPT detectors are biased against non-native English writers, 2023.

Their makers publish low rates, and a low rate applied to a lot of writing is still a lot of people. When Turnitin launched its AI detector it gave a false positive rate of 1%. Vanderbilt University, turning the detector off in 2023, worked out what that meant: of the 75,000 papers its students submitted the year before, around 750 could have been wrongly labelled (Vanderbilt’s guidance).

OpenAI withdrew its own classifier for AI-written text in July 2023, six months after launching it, because of its low rate of accuracy (OpenAI). Substack, announcing the Pangram scan it gave every reader in 2026, said of it that it “is not perfect” (Substack).

At most, a reason to look closer: to ask the writer about the piece, and to look at how it was written. It is not, on its own, evidence of what somebody did, and a score that is treated as a verdict will sooner or later be a verdict on somebody who wrote every word.

Evidence of how the piece was written: version history, drafts, notes, sources, and a writer who can talk about their own choices. How to prove you wrote it lists them, and accused of using AI says what to do with them.

What cannot stand against one is changing the writing. Adding mistakes, dropping the em dashes or running a piece through a tool that promises to make it read as human moves the score, proves nothing, and costs the writer their voice.

No. iwrote does not guess from your sentences. It records how the words got onto the page as you write them, and certifies the piece with a link that hands that record to your reader, who can check it. The model that reads the session for signs it was manufactured is sent the counts and timings, never the words, and what that reading cannot establish is written down.