Why human writing can be flagged as AI
False positives are rare but real. These are the kinds of writing most likely to trip a detector, and what to do when it happens.
Updated Oct 8, 20262 min read
AI detectors look for statistical patterns: predictable word choices, a steady sentence rhythm and familiar structure. Some human writing has those patterns too. On our benchmark, 0.6% of human-written documents are wrongly flagged as AI. That's low, but across thousands of essays it isn't zero.
Writing that's more likely to be flagged
- Formulaic genres: lab reports, abstracts, legal text, cover letters and product descriptions follow fixed templates.
- Very short pieces: with only a few sentences, there's little signal to go on.
- Text polished by tools: heavy use of grammar checkers or rewording apps smooths out the quirks that make writing look human.
- Writing in a second language: careful, textbook-correct prose can read as more uniform.
- Well-known passages: famous speeches, definitions and boilerplate are everywhere in the data models learn from, so they read as model-like.
If your own writing was flagged
- Step 1: Look at which sentences are highlighted. If it's a handful of generic lines, the result is weak evidence.
- Step 2: Gather your drafts. Version history in Google Docs or Word, notes and outlines show how the work developed.
- Step 3: Ask for a conversation. Explaining your argument and sources in your own words is the most convincing evidence there is.
If you're reviewing someone else's work
Treat a low human score as a reason to ask questions, not as proof. Compare the work with the writer's earlier pieces, look at drafts, and talk with them. Talking with a writer about a flagged result has a simple structure for that conversation.