How AI detectors work

What GPTZero, Turnitin and other detectors actually measure, why they produce false positives, and what that means for your writing.

Updated October 17, 2026

The short version

AI detectors estimate how likely it is that a language model produced a piece of text. They don't search a database of ChatGPT answers. Instead, they measure statistical properties of the writing and compare them with patterns learned from large collections of human and AI text. The result is a probability, not proof.

Signal 1: perplexity

Language models generate text by repeatedly choosing a likely next word. Perplexity measures how “surprised” a model is by a piece of text. AI-generated text tends to have low perplexity — it is made of the words a model would have predicted. Human writing usually has higher perplexity: we pick odd words, make unexpected jumps and use phrasing a model wouldn't choose.

Low perplexity: “Exercise has many benefits for both physical and mental health.”
Higher perplexity: “I started running to get out of the house, and somehow ended up liking hills.”

Signal 2: burstiness

Burstiness is how much perplexity and sentence structure vary across a text. People write in bursts: a long, winding sentence, then a short one. A fragment. Then something else entirely. AI output is often uniform — similar sentence lengths and similar predictability throughout. GPTZero popularised these two measures when it launched in early 2023.

Signal 3: trained classifiers

Modern detectors go further than two numbers. They train classifier models on millions of labelled examples of human and AI writing, across many models, topics and styles. The classifier learns subtler patterns — favourite words, punctuation habits, how paragraphs open and close. Commercial tools such as GPTZero, Turnitin and Originality.ai combine several signals and update their models as new AI systems appear.

Why detectors make mistakes

False positives (human text flagged as AI)

• Non-native English writing. Writers using a smaller vocabulary and simpler structures produce lower-perplexity text. A 2023 Stanford study (Liang et al.) found several popular detectors misclassified a large share of TOEFL essays by non-native speakers as AI-generated.
• Formulaic writing — abstracts, legal clauses, technical instructions — is predictable by design.
• Grammar tools that smooth out your natural irregularities.
• Famous text. Well-known passages (like constitutional text) can look “predictable” because models have seen them many times.

False negatives (AI text passing as human)

• AI text that has been substantially edited or rewritten.
• Prompts that ask for unusual styles, errors or personal tone.
• Newer models the detector hasn't been trained on yet.

What the score means in practice

Detector companies themselves advise against using a score as the only evidence of misconduct. A sensible reading is:

• High AI probability: worth a closer look and a conversation — not an automatic verdict.
• Mixed: common for drafts edited with AI tools, or human writing with formulaic sections.
• Low AI probability: reads as human, though heavily edited AI text can land here too.

How this relates to humanizers

Humanizers rewrite text to increase variation — the same burstiness and unpredictability detectors look for in human writing. That is also what good editing does: varied sentences, specific details, a real voice. You can see the effect directly by running text through the AI detector before and after using the humanizer.

Frequently asked questions

Keep exploring

• Free AI detectorCheck any text for AI writing with a free GPTZero scan.
• GPTZero vs Turnitin vs Originality.aiWho uses each detector and how they report scores.
• Why AI text sounds robotic — 12 tells and how to fix themThe 12 most common AI tells, with a fix for each.