Why an AI Detector Says Your Essay Is AI (When It Isn't)

You spent hours on it. You balanced the argument, worked the transitions, chose the right words. Then, out of curiosity or caution, you ran it through an AI detector.

85% AI-generated.

The result is alarming and, in a large number of cases, simply wrong. Here is what that number is measuring, and why your careful essay is exactly the kind of writing most likely to trip it.

Detectors do not detect AI

This is the single most important thing to understand, and almost nobody explains it plainly.

An AI detector has no way of knowing who or what wrote a document. It cannot see your drafts, your keystrokes, or your search history. It sees finished text. From that text it computes two statistical properties:

• Perplexity — how predictable each word is given the preceding words. Language models are optimised to select likely words, so their output sits in a narrow, low-surprise band.
• Burstiness — how much sentence length and rhythm vary across a passage. Human writing tends to swing between clipped sentences and sprawling ones. Model output tends to level out.

Low perplexity plus low burstiness produces a high AI score. That is the entire mechanism.

Notice what it implies. The detector is not asking "did a machine write this?" It is asking "is this text predictable?" Those are different questions, and conflating them is where the harm comes from.

Why good essays are the most at risk

Academic writing is trained predictability. Every convention you were taught pushes your prose toward the exact profile detectors flag.

Structured repetition

Essay form demands signposting. Topic sentences, transitional phrases, a thesis restated in the conclusion. That deliberate scaffolding looks a great deal like a language model following a template.

Consistent register

You were told to hold a formal, even tone throughout. Consistency of tone reads as low variance, and low variance reads as machine-generated.

Sophisticated but conventional vocabulary

Strong essayists reach for precise, slightly formal words. So do language models, because they were trained on exactly this kind of text. The detector cannot separate "well-read human" from "model trained on well-read humans."

Clean, edited prose

Every round of editing removes an oddity. Polished writing has fewer surprises by construction, which is another way of saying it has lower perplexity.

The perverse conclusion is that the better you write within academic convention, the more you look like a machine.

The evidence that detectors get this wrong

This is not speculation. It is documented in peer-reviewed research.

Liang et al., "GPT detectors are biased against non-native English writers", published in Patterns (2023), evaluated seven widely used detectors against 91 TOEFL essays by non-native English speakers and 88 essays by US eighth-graders.

• On the US student essays, the detectors were near-perfect.
• On the TOEFL essays, the average false positive rate was 61.22%.
• All seven detectors unanimously misclassified 18 of the 91 TOEFL essays.
• 97.8% of TOEFL essays were flagged by at least one detector.

Then the decisive finding. When the researchers rewrote those same human essays using more elaborate, literary vocabulary, the false positive rate fell from 61.3% to 11.6%.

The essays did not change authorship. Only the register changed. The detectors were measuring style, not origin.

If English is your second language, this is not a hypothetical risk — it is the documented default behaviour of these tools. Detector bias against non-native writers covers what that means practically.

What to do when you are flagged

Do not rewrite in a panic

The instinct is to mangle your prose until the score drops. Resist it. You will usually damage the writing, and a lower score on one detector does not transfer to another. If the work is yours, your energy belongs on evidence, not cosmetics.

Gather your version history

This is the strongest card you hold. Google Docs keeps a full revision timeline under File → Version history. Microsoft Word does the same through OneDrive or SharePoint. Both show the document accumulating over hours or days.

AI-generated text has no ugly middle. It arrives complete. A revision history showing a thesis you rewrote four times is very hard to argue with.

Keep everything around the essay

Handwritten notes, outlines, annotated PDFs, library records, browser history on your sources. Individually weak, collectively persuasive.

Run a second and third detector

Detectors disagree with each other constantly. If one says 85% and another says 4%, that spread is itself an argument: the measurement is not reliable enough to support an accusation. We compare how far apart the major tools land in Turnitin vs GPTZero vs Grammarly.

Bring the research

If it escalates to a formal process, cite the Liang study. A documented 61% false positive rate across a whole class of writers is a serious methodological problem with the evidence being used against you, and most institutions have no policy answer to it.

Writing that is less likely to be flagged

If you want to reduce your exposure without gutting your style, the target is variation, not obfuscation.

1. Vary sentence length deliberately. Follow a long, subordinate-clause-heavy sentence with a short one. That swing is burstiness.
2. Keep your own oddities. The slightly unusual word choice, the aside, the specific example only you would use. Editing tools strip these; they are also the strongest human signal you have.
3. Use concrete specifics. Named sources, real numbers, particular examples. Models generate plausible generalities; people cite the thing they actually read.
4. Write a bad first draft on purpose, then edit. The version history that produces is worth more than any score.

The honest summary

Getting flagged is not a judgement on your writing ability. It is a judgement on the state of detection technology, which currently cannot separate careful human prose from machine output because both are, in the statistical sense, predictable.

Until something better than perplexity exists, your best protection is not a cleaner score. It is a documented trail proving the work grew under your hands.

For the full picture — the research on detector accuracy, who gets flagged most, and how to escalate a wrongful accusation — read our complete guide to AI detector false positives. If you want the underlying figures rather than the summary, we collected every published AI detector accuracy number in one table, including the study that found a 61% false positive rate on essays by non-native English speakers.

If you are working from an AI-assisted draft and want it to carry your voice rather than a model's cadence, RewriteAI's humanizer restores the sentence-level variation that both AI generation and heavy editing tend to flatten out.