Humanize AI Text for GPTZero: Perplexity and Burstiness

Detector guide

Humanize AI Text for GPTZero

GPTZero is the original AI detector. Edward Tian launched it in January 2023 and the company has shipped multiple model upgrades since. To humanize text for GPTZero specifically, you need to move two numbers: perplexity and burstiness.

Jan 2023

GPTZero launched

first widely-used AI detector

Core metrics

perplexity and burstiness

1-3%

False-positive rate

in independent tests

100M+

Documents scanned

per the company's public stats

The two metrics that drive GPTZero

Perplexity estimates how surprising your word choices are to a language model. Low perplexity means the model would have predicted your words. High perplexity means you wrote something the model did not expect. AI text scores low. Human text scores higher.

Burstiness measures variance: how much sentence length and complexity change from one sentence to the next. Humans write in waves. AI writes in a metronome. A passage that scores AI on GPTZero has low mean perplexity and low burstiness. A passage that scores human has higher mean perplexity and visible variance.

Where AI and human text live on the GPTZero plane: AI clusters in the bottom-left (low perplexity, low burstiness).

What raises perplexity

Specific verbs, idioms, unusual modifiers, and first-person observation. A model rarely picks these on its own.

Low perplexity

"The committee will utilize various methods to comprehensively evaluate the proposal."

Higher perplexity

"The committee plans to poke at the proposal from a few different angles before saying yes or no."

Same intent. The second sentence uses 'poke at' (idiom), 'plans to' (specific) and 'saying yes or no' (concrete), all words a model would not have predicted.

What raises burstiness

Aggressive length variation, fragments, coordinating-conjunction starts, asides, and varied paragraph shapes.

Move	Effect	Risk
Mix 4-word and 30-word sentences	Sharp increase in burstiness	Reads jumpy if overdone
Sentence fragments	Spikes the variance	Avoid in formal writing
Start sentences with And or But	Breaks AI's clean flow	Some style guides forbid
Parenthetical asides	Adds rhythm variation	Use sparingly
Vary paragraph length	Different shape on the page	Always safe

What humanizers actually do

A simple paraphraser swaps synonyms one for one. That moves perplexity slightly but does almost nothing to burstiness. Output stays uniform-rhythm and GPTZero still flags it. A real humanizer restructures.

Re-tokenize

Substitute high-probability AI words for human-pattern equivalents.

Re-shape sentences

Split long sentences. Combine short ones. Add at least one fragment.

Re-mix structure

Collapse parallel lists into prose with varied connectives.

Re-check meaning

Verify the original ideas, facts, and tone are preserved.

One detail that matters

GPTZero looks at per-sentence perplexity, not just document-level. A document with a few high-perplexity sentences and many low-perplexity ones can still score human if the variance is high enough. This is why varied sentence length helps even when individual sentences read AI-like.

Related guides

Run your draft through the humanizer

Output is rewritten to score in GPTZero's human range while keeping your meaning intact.

Open the humanizer