GPTZero score: what the number means and what it does not
A GPTZero score is the probability that GPTZero, a widely used AI detection product, assigns to a piece of text having been written by AI. It is usually presented as a percentage for the whole document, a verdict such as human, AI or mixed, and per-sentence highlighting that shows where the tool is most suspicious. Humanize 360 is a different product with a different scale, and we do not know GPTZero's internals. What follows is drawn from what the company has said publicly and from how detectors of its kind generally work.
GPTZero became well known in early 2023 as one of the first tools built specifically to flag ChatGPT output, and it described its early approach in terms of two ideas, perplexity and burstiness, that we cover elsewhere in this glossary. It has since described adding trained components and multiple model versions. Which of those runs on your text today, and with what weights, is not published in detail, so anyone who tells you exactly why GPTZero gave a particular number is guessing.
How to read the report
- The document-level percentage is a probability, not a proportion. "Sixty percent AI" does not mean sixty percent of the words were machine-written.
- Sentence highlighting shows where the model is most confident, which is useful for finding flat passages whoever wrote them.
- The mixed verdict is the honest one for most edited drafts and deserves more attention than the extremes.
- Short texts get less reliable results, a limitation the vendor itself notes.
Why the number moves
Detection products update their models, and a text that scored one way last month can score differently now with no change to the words. The same text also scores differently on other vendors' tools, because each uses its own reference model and threshold. None of this is a flaw specific to GPTZero; it is the nature of estimating authorship from texture. It is why no humanizer, ours included, can promise a GPTZero result.
How it compares with the Human Score
The Human Score on /ai-detector runs the other way: 0 to 100 where higher means more human, with above 70 reading human, 45 to 70 mixed and below 45 likely machine. Instead of a probability, you get eleven named signals with their own readings, so you know whether the issue is sentence rhythm, stock phrases, transition words or something else. We designed it as an editing tool, not a verdict machine, and it runs on our own code with no third-party model.
Using both sensibly
If your institution or client uses GPTZero, the only honest advice is to write text that carries fewer of the tells all detectors respond to, then check with more than one tool and expect them to disagree. Keep your drafts. If you are the one assessing, read our page at /ai-detector-for-teachers before treating any percentage as a finding.
Common questions
What GPTZero score counts as AI?
GPTZero sets its own thresholds and may change them. We do not publish or guess at their cut-offs. Read the verdict label in their report and treat the percentage as a probability.
Does Humanize 360 predict my GPTZero score?
No. We measure the same family of statistical patterns, but we cannot know their current model or thresholds. A high Human Score means the common tells are gone; it is not a promise about any vendor.
Why did GPTZero flag my own writing?
Formal, careful or formulaic prose reads as predictable, and predictability is what these tools measure. That is a false positive. Our detector breakdown shows which signals your text carries so you can argue the point.
Is GPTZero free?
It has offered a free tier with limits and paid plans; check their site for current terms. Our detector is free for 300 words per check, three checks a day.