How to reduce AI-detection false positives.

Why human essays sometimes look like AI.

The frustrating thing about modern AI detectors is that the writing patterns they flag aren't exclusive to language models. Three signals dominate most classifiers, and all three show up in legitimate human writing too:

• Low perplexity. Perplexity measures how 'surprised' a reference model is by your text. Human academic writing in the social sciences and humanities is often quite predictable, well-edited prose follows known structural patterns. If you've revised your essay carefully and it reads smoothly, you've probably also lowered its perplexity.
• Flat burstiness. Burstiness is variance in sentence length and complexity. The same revision pass that smooths out perplexity also tends to flatten burstiness, you remove the awkwardly long sentences and the abrupt short ones. The detector reads the polished result as machine-like.
• Generic discourse markers. Phrases like 'in conclusion', 'it is important to note', 'on the other hand' show up disproportionately in AI output because they're common in the training data. They're also common in essays written by students who learned a five-paragraph template in middle school.

This is why the false-positive rate is concentrated in two populations: ESL writers (who often rely more heavily on memorized academic phrases) and high-effort revisers (whose final draft is statistically smoother than a typical first draft). The five steps above target each of these signals directly without changing what your essay actually says.

What to do if you've already submitted.

If you submitted an essay you wrote yourself and the detector flagged it after the fact, the editing tactics in the steps above don't help anymore, the submission is locked. Different playbook:

1. Surface your draft history immediately, before the conversation. If you wrote in Google Docs, share the version history with timestamps. If you wrote in Word, export tracked changes. This is the single strongest exonerating signal you can produce.
2. Don't rewrite or 'fix' the essay retroactively. Editing a flagged essay after the fact looks worse than the flag itself. Lock the file. Print or PDF it. Make sure the version your professor evaluates is exactly what you submitted.
3. Run it through a second detector and document the result. If GPTZero says 8% AI and Originality.ai says 9%, save those screenshots with timestamps. Two independent classifiers agreeing is much harder to dismiss than one disagreeing detector.
4. Read the appeal procedure before the meeting. Most schools have a formal appeal path for academic-integrity findings. The procedural details matter, deadlines, who can attend, what evidence you can bring. Showing up prepared shifts the dynamic substantially.

If you need a structured appeal template, our how-to-appeal guide walks through the full process step-by-step.

The line between editing and cheating.

This guide gets a lot of pushback from instructors who worry it amounts to teaching students how to fool detectors. The distinction we hold is simple: if the words are yours, adjusting their rhythm to reduce a false alarm is editing. If the words aren't yours, adjusting their rhythm to hide that fact is cheating. The action looks similar; the underlying fact is different.

This is also why our humanizer is gated behind a free account, why it runs your text through our own AI detector before and after, and why it shows you both scores side-by-side. We want the legitimate use case (lowering false-positive risk on a paper you actually wrote) to be easy and obvious, and the illegitimate one (running AI-generated text through a laundromat) to be uncomfortable enough that most users self-select out of it. Read the full humanizer ethics policy for the longer version of this argument.