Scripli logoScripli
Back to Scripli

GPTZero Flagged Your Writing. Here's What the Score Means.

GPTZero uses statistical measures of text — perplexity and burstiness — that can produce false positives on formal human writing. The method has published limitations that work in your favour when appealing.

How GPTZero works — and where it fails

GPTZero measures two properties of text: perplexity (how predictable each word choice is) and burstiness (how varied sentence lengths are). The theory is that AI-generated text tends toward low perplexity (predictable word choices) and uniform burstiness (consistent sentence lengths).

The problem is that formal human writing — academic essays, professional reports, carefully edited prose — shares exactly these statistical properties. A student who writes formal, precise academic prose will produce text with low perplexity and uniform burstiness. GPTZero cannot distinguish this from AI output.

Non-native English speakers are especially vulnerable. Writing carefully in a second language tends toward conservative, predictable word choices — the exact signal GPTZero interprets as AI. Research has documented false positive rates of 20%+ for non-native academic writers.

What GPTZero says about its own accuracy

GPTZero's own published research and website acknowledge limitations. Their model is described as a tool to assist educator judgment — not as a standalone verdict. From GPTZero's documentation:

“GPTZero is designed to be a helpful tool that assists human judgment — it should not be the sole determiner of any academic decision.”

GPTZero also reports accuracy figures that, by definition, include false positives. When they report “98% accuracy,” that still means 2% of clean human text is falsely flagged. At scale — a university reviewing hundreds of submissions — that is a significant number of innocent writers.

This matters for your appeal. If an instructor is treating GPTZero's result as conclusive, they are using the tool against the developer's stated intent.

What to do right now

  1. 01

    Don't panic, and don't delete anything

    Every file, draft, and message is potential evidence. Don't delete or modify anything until your appeal is resolved.

  2. 02

    Gather your evidence

    Google Docs version history is the strongest: File → Version history → See version history. Screenshot the earliest versions and the full timeline. Also collect any draft files, research notes, and emails about the work.

  3. 03

    Write an appeal that addresses the statistical argument

    Your appeal should: (1) note that GPTZero is a probabilistic tool, not proof, (2) cite GPTZero's own guidance that it assists judgment rather than replaces it, (3) present your evidence of process.

  4. 04

    Request a conversation, not just a written response

    A live conversation where you can discuss your writing process — the sources you consulted, decisions you made, phrases you changed — is often more persuasive than a letter.

Prevent this from happening again

Scripli records your writing session and issues a certificate before you submit. GPTZero can't dispute a verified writing record.