GPTZero estimates how likely a text is to be AI-generated by looking at how predictable the writing is and running it through a trained classification model, then reports a document label (human, mixed or AI), probabilities, and highlighted sentences. The scores are confidence estimates, not measurements of how much AI was used, so they need careful reading, especially in the middle range.
ZeroTell publishes this blog and makes a competing AI detector. This piece focuses on GPTZero's detection method and how to interpret its results, based on what GPTZero has publicly explained. For features, workflow and who it suits, see our GPTZero review.
How GPTZero's detection works
Where it started: perplexity and burstiness
When GPTZero launched in early 2023, it became known for two concepts:
- Perplexity: how predictable the text is to a language model. AI text tends to use highly likely word sequences, so it scores low.
- Burstiness: how much that predictability varies from sentence to sentence. Human writing tends to swing between plain and surprising sentences; AI writing tends to stay level.
These ideas were easy to explain and made GPTZero's approach understandable to teachers.
Where it is now: a layered model
GPTZero has since described its detector as a multi-part system rather than just two metrics. It combines signals like those above with a classifier trained on large collections of human and AI writing, and it works at both the sentence and document level. It has also said it updates the model as new AI systems appear.
The practical takeaway: you can't predict GPTZero's result just by measuring sentence length or word rarity yourself. The model weighs many features together.
Why sentence-level matters
GPTZero scores individual sentences and then combines them. That's why reports highlight specific lines. It's also why a document can be labeled "mixed": some sentences look human-written, others AI-like.
Reading the result: label, probability, highlights
A GPTZero report usually gives you three layers of information. Exact names and layout may change, but the logic stays similar.
| Part of the report | What it tells you | How to use it |
|---|---|---|
| Document label | The model's overall call: likely human, mixed, or likely AI | A starting point, not a verdict |
| Probability breakdown | How confident the model is in each category | High confidence in one class means more weight; split probabilities mean uncertainty |
| Highlighted sentences | Which lines most influenced the AI judgment | Read these closely to see whether they actually seem generic or out of character |
What the percentage is not
A common misreading: "80% AI" means 80% of the essay was written by AI. Usually it means the model is about 80% confident the document belongs in the AI category. A fully human essay written in a very formulaic style can get a high confidence score, and a partly AI essay can get a lower one.
Why scores vary
People are often surprised that the same essay can score differently. Common reasons:
- Length. More text gives the model more signal. Short passages swing more.
- Context. Adding or removing a paragraph changes the surrounding context for every sentence.
- Genre. Abstracts, lab reports and definitions are naturally predictable.
- Editing. Text revised by a person after AI drafting lands in uncertain territory.
- Model updates. GPTZero improves its detector over time, which can shift scores on the same text.
- Language background. Writers using simpler, careful English, often second-language writers, can produce text that looks more predictable.
Interpreting results fairly: a step-by-step approach
If you're a teacher or editor looking at a GPTZero report, this order helps:
- Check the length. If it's short, weigh the result lightly.
- Look at the probabilities, not just the label. A narrow lead for "AI" is much weaker than a confident one.
- Read every highlighted sentence. Are they generic statements anyone might write, or do they contrast sharply with the rest of the text?
- Compare with known writing from the same person.
- Check sources. AI-written text sometimes cites sources that don't exist. Real, accurate citations point toward genuine research.
- Ask about process. Drafts, outlines and version history are stronger evidence than any score.
- Talk before you act. A conversation resolves many cases.
GPTZero itself has said its results shouldn't be used as the sole basis for punishment. That guidance matters.
A worked example
Imagine a 900-word student reflection labeled "mixed," with modest AI probability, and four sentences highlighted. Three of them are in the introduction and say things like "Technology has transformed the way we learn." The fourth is the conclusion.
A careful reader would notice that the highlighted lines are the most generic in the essay, while the middle, with personal details from a class field trip, isn't highlighted. That pattern is consistent with a student who wrote a bland opening and closing, and also with one who asked AI for an intro and conclusion. The score alone can't tell which. The student's draft history and a short conversation can.
AI DetectorSentence-level AI scoreCommon misreadings to avoid
A few interpretation mistakes come up again and again.
"Mixed means partly cheated." Mixed means the model saw both patterns. Plenty of honest writers produce a few generic sentences, especially in introductions and conclusions, which are the most formulaic parts of any essay.
"Human means definitely human." A human label is the model's best guess too. Heavily edited AI text, or AI text in an unusual style, can land there. A label in either direction is an estimate.
"The highlighted sentence is the AI sentence." Highlights show which sentences most influenced the score. They don't prove those exact lines were generated, and unhighlighted lines aren't proven human.
"Two tools agreeing settles it." Detectors share similar ideas and can share similar blind spots. Agreement adds some weight, but it doesn't replace process evidence or a conversation.
"A score from last month still applies." If the model has been updated since, the same text may score differently today. Rescan before relying on an old report, and note the date on any report you keep.
If GPTZero flags your writing
If you wrote the work yourself:
- Stay calm. A flag is the start of a conversation.
- Gather your evidence: version history, notes, outlines, research tabs.
- Look at the highlighted sentences. Often they're generic lines. Revising them improves your writing anyway.
- Explain your process to your instructor and offer to discuss the content.
If you used AI in ways your course allows, say so clearly and show what you did. If you used it in ways that weren't allowed, honesty usually leads to a better outcome than trying to defend the score. Our guide on AI detection false positives has more detail.
What GPTZero's detection is good for
Used well, GPTZero's detection is valuable for:
- Triage: pointing reviewers to passages worth reading closely.
- Teaching: showing students what generic, predictable writing looks like.
- Self-review: helping writers spot flat sentences that need specifics.
It isn't designed to prove authorship, and no detector can do that alone.
Key takeaways
- GPTZero began with perplexity and burstiness and now uses a layered classifier working at sentence and document level.
- Reports combine a label, probabilities and highlights; the probability reflects confidence, not the share of AI used.
- Scores vary with length, genre, editing, language background and model updates.
- Interpret results by reading highlights, comparing with known writing, and asking about process.
- No detector score should be the only basis for an integrity decision.
Next time you see a GPTZero report, ignore the headline label for a moment and read the highlighted sentences first. You'll form a much fairer judgment.