Academic · Educators & students
GPTZero: how it works & how to pass
The classroom-favorite detector built on perplexity and burstiness.
How GPTZero works
GPTZero scores text on two statistical properties: perplexity (how predictable each word is to a language model) and burstiness (how much sentence length and complexity vary). AI text tends to be low-perplexity and low-burstiness, so it reads as uniformly smooth and even. GPTZero highlights the specific sentences it judges most machine-like.
What it flags
- Uniformly smooth, predictable word choices (low perplexity)
- Sentences of similar length and rhythm (low burstiness)
- Repeated clause shapes and transition words
- An even, polished tone with no rough edges
A note on accuracy
GPTZero is among the more cautious detectors, but it still false-positives. Non-native English and tightly edited human writing get flagged because they are also low-perplexity. Treat any single score as a rough probability.
How NotepadLLM helps your text read as human
- Rebuilds sentence-length variation (burstiness) the way human drafts wander
- Raises perplexity with fresher, less predictable phrasing
- Breaks repeated clause templates and stock transitions
- Keeps meaning, facts, and citations intact while changing register
GPTZero FAQ
Does GPTZero detect ChatGPT and GPT-4 text?
Yes, it's built for exactly that. It works by measuring statistical smoothness, with no reliance on a hidden watermark, so it flags text from most large models, including GPT-4 and Claude.
Why did GPTZero flag my own writing?
Low-perplexity human writing, such as concise, heavily edited, or non-native English, shares the statistical fingerprint GPTZero associates with AI. False positives are a known limitation.
How do I make text read as human to GPTZero?
Increase variation: mix long and short sentences, use specific and occasionally unexpected wording, and avoid uniform paragraph shapes. That's exactly what NotepadLLM's humanizer does structurally.