Why Turnitin, GPTZero, and Other Detectors Flag AI Writing

Turnitin, GPTZero, Originality.ai, Copyleaks, and ZeroGPT do not read for meaning — they score statistical patterns in how text was likely generated. Two of the biggest signals are perplexity and burstiness.

Perplexity

Perplexity measures how predictable each word choice is. Language models tend to pick the statistically most likely next word, which keeps perplexity low. Human writing is comparatively less predictable.

Burstiness

Burstiness measures variation in sentence length and structure. AI-generated text tends toward a uniform rhythm, sentence after sentence of similar length and complexity. Human writing swings between short fragments and longer, winding sentences.

Why Synonym Swaps Do Not Work

Basic paraphrasing tools change words, not sentence structure. Since perplexity and burstiness are about structure and rhythm, not vocabulary, that kind of surface-level rewriting rarely moves the needle.

How We Check Our Own Output

We test humanized output against Turnitin, GPTZero, Originality.ai, Copyleaks, and ZeroGPT directly and tune the engine as those tools update. See our full methodology on the How We Test page.