Why AI Detectors Flag Human Writing & How to Check Fairly
AI-writing detectors flag a lot of human text: formal essays, templated reports and, above all, writing by non-native English speakers. This guide explains why that happens, what a 62% AI score does and doesn't mean, and a fair way to check your own draft before a teacher, editor or client runs it through a detector.
Short version: detectors measure how predictable your text is, not who wrote it. Predictable human writing gets flagged. This page explains the mechanics, the research, and a practical way to check a draft before someone else does. Our partner ScanForAI is the detector we point to, because it shows the signals behind its score instead of a bare percentage.
Why detectors flag human writing
Almost every AI-writing detector works the same way. A language model reads your text and estimates how surprised it is by each word. Text that follows the most likely path word after word scores as "low perplexity", which is what AI output tends to look like. A second signal, often called burstiness, measures how much sentence length and structure vary. People mix short and long sentences; models keep an even rhythm. The detector is trained on samples labelled human or AI and learns where to draw the line.
Nothing in that process checks who typed the words. So anything that makes human writing predictable pushes it across the line:
- Formal or academic register. Standard phrasing, hedged claims and conventional transitions are exactly what a model predicts next.
- Templated documents. Cover letters, lab reports, abstracts, policy summaries and legal boilerplate follow a fixed pattern by design.
- A narrower vocabulary. Writers working in a second language, or in a field with fixed terminology, reuse the same words and structures.
- Heavy grammar-tool editing. Rewrite and "improve" features smooth sentences towards the statistical average. Plain spelling and punctuation fixes rarely matter; wholesale rewrites can.
- Short samples. Under roughly 300 words the detector has too little signal and the score swings between scans.
- Lists and definitions. Bullet points, step-by-step instructions and glossary-style text are predictable by nature.
The non-native English problem
The clearest evidence comes from a 2023 Stanford study (Liang et al., published in Patterns; preprint on arXiv). The researchers ran essays through seven widely used detectors.
| Sample | Result |
|---|---|
| 91 TOEFL essays by non-native English writers | More than half misclassified as AI-generated; 61.3% average false-positive rate across the seven detectors, and one detector flagged nearly all of them |
| Essays by US eighth-graders (native speakers) | Under 10% false positives for most detectors |
| The same TOEFL essays after a model was asked to enrich the vocabulary | False positives dropped sharply |
The last row matters twice. It shows the bias is about vocabulary range, not about being human. It also shows how easy the detectors are to steer, which is why no serious tool should be the only evidence against anyone. OpenAI retired its own AI text classifier in July 2023 citing low accuracy, and several universities, Vanderbilt among them, switched off Turnitin's AI indicator the same year.
What the percentage actually means
A "62% AI" result does not mean 62% of the words were written by a machine. Depending on the tool it means one of two things: the model is 62% confident the whole text is AI-generated, or about 62% of the sentences crossed the tool's threshold. Neither is a measurement of authorship, and the two readings can disagree for the same text. Read the tool's own explanation of its score before you repeat the number to anyone.
Three habits keep a score in proportion:
- Look at the sentence view, not the headline. A tool that highlights which sentences it doubts tells you something you can act on. A bare percentage does not.
- Scan the whole piece. Fragments and short paragraphs produce noise.
- Rescan after real edits. If the score barely moves when you rewrite the flagged sentences, the signal was weak to begin with.
How to check your own draft fairly
- Finish the draft first. Scan the final text, at least a few hundred words, not an outline or a single paragraph.
- Use a detector that shows its reasoning. ScanForAI marks the sentences it considers machine-like, lists the signals behind the score in plain English, and adjusts its scoring for non-native English writing and for the type of text being scanned. You can paste text, upload a file or give it a web page URL. No account is needed and anonymous scans are not stored.
- Read the heat-map. Flagged sentences are usually the ones that could appear in any essay on the topic. Replace them with your specific example, your data, your opinion. You are not disguising anything; you are making the text more yours.
- Vary rhythm, keep your voice. Put a short sentence after a run of long ones. Use the concrete word instead of the general one. Do not run the text through a paraphraser; that makes it less yours, not more.
- Keep evidence as you go. Version history in Google Docs or Word, dated notes and earlier drafts show how the text grew. Our text diff tool makes a before-and-after comparison you can keep or share.
- If you used an AI assistant legitimately, for brainstorming, an outline or a grammar pass, keep the conversation. The AI chat export cleaner turns a messy ChatGPT or Gemini log into a tidy PDF you can attach if asked.
If you are the one holding the detector
For teachers, editors and reviewers the same mechanics apply in reverse. A high score is a reason to look closer, never a finding. Compare the text with the writer's earlier work, ask about the process, look at drafts and version history, and scan with more than one tool before drawing any conclusion. A non-native English writer with a careful, formal style is the most likely person in the room to be flagged by mistake. ScanForAI's own documentation asks users not to present its scores as verdicts about people, and that is the right standard for any detector.
Why we point to ScanForAI
SimpleToolsHub tools run in your browser and never upload your files, so we are picky about the services we recommend. ScanForAI is a partner site, and we link to it because its design matches the advice above:
- Sentence-level heat-map plus a per-signal breakdown, with scoring adjusted for non-native English writers and for text type.
- Fast and Deep scan modes, and a mixed-text mode that scores human and AI sections separately.
- Scan pasted text, uploaded files or a web page by URL. A free Cleanup tool strips hidden characters, smart quotes and other paste artefacts.
- Browser extension for Chrome, Edge and Firefox, plus an API and an MCP server for developers.
- Works in English, Spanish and Turkish.
- Free Class links for teachers: students scan their own work and choose whether to send the result; a free account runs five open classes of 100 submissions each.
- A free image prompt-injection scan and a declared-provenance check for images, both with a free account.
- No account needed: 20,000 characters per scan and nothing stored on anonymous scans. A free account raises the limit to 50,000 characters and adds a year of history, batch scans and share links; Pro is $19 a month as of November 2026.
- Its own motto is "a signal, not a verdict", which is the attitude this whole page argues for.
Related guides: Essay flagged as AI? What to do · For teachers: free Class links and a fair policy · Image prompt injection: how to check an image.
Disclosure: ScanForAI is a partner site. Links on this page carry tracking tags so they can see visitors arriving from SimpleToolsHub; we don't earn a commission. We make no accuracy claims for any detector, and no detector score should be treated as proof of how a text was written.