Turnitin and GPTZero give different answers. Here is why
Two detectors on the same essay routinely disagree. That is structural, not a sign that one of them is broken, and the disagreement itself is the useful output.
The common situation: an institution's report shows one number, you run the same text through a standalone tool and get another, and the two are far apart. It happens constantly, and it is worth understanding before treating either figure as the real one.
They are not measuring the same thing
Both are classifiers that estimate whether text has the statistical properties of generated writing. Underneath, four things differ, and each one moves the number.
Training data. A detector learns from the examples it saw, so a tool trained on a narrow slice of generators inherits a narrow profile, and almost none of them will tell you what theirs was. Ours is trained across 14,892 generated documents spanning seven model families and five years of GPT versions, which is why it holds 89.9% on OpenAI GPT output.
The threshold. Every detector produces a continuous score and then cuts it somewhere. Two tools can compute nearly the same internal number and report different verdicts because one drew its line at a stricter false positive budget than the other. This alone accounts for a large share of disagreements.
Segmentation. Whether a tool scores sentences, paragraphs or whole documents changes the answer substantially. Turnitin publishes both of its own figures, under 1% of documents against around 4% of sentences, which is the same tool four times worse at the smaller unit.
What counts as text. References, equations, tables, code blocks and markup get handled differently by different tools. On a technical document this shifts the denominator enough to move the percentage on its own. See why LaTeX breaks AI detectors.
The practical differences
Turnitin runs inside the submission system. You generally do not choose it, you do not run it, and the report goes to the institution rather than to you. The AI indicator sits next to the similarity score on the same page, which invites people to read them as the same kind of evidence. They are not: similarity points at a source document you can click through to, and the AI indicator points at nothing, because nothing was found. See how Turnitin AI detection works.
Standalone tools are ones you run yourself. You get the number directly, on text you choose, as often as you like, subject to whatever cap the tool has. That makes them useful as a second measurement and useless as a prediction of what an institution's tool will say.
Which number is right
Whichever one is wrong about human writing less often.
Accuracy blends missing generated text with flagging text a person wrote, and only one of those lands on somebody. Turnitin publishes 4%. Ours is 0.4%, which is 99.6% specificity, measured on 15,900 documents of real academic writing. Many tools publish neither number, and that absence is itself information.
Using the disagreement
Wide disagreement between two detectors is genuinely informative. It usually means the text sits near a boundary, where small differences in threshold or segmentation flip the verdict. Text that is unambiguous to one classifier is generally unambiguous to the others.
So running a second tool is worth doing. Agreement raises confidence and disagreement tells you the text sits near a boundary, which is worth knowing before somebody else runs their own check on it.
Ours is free and unlimited with no account, takes .pdf, .docx and .tex, and scores each paragraph separately. The method is in how the unslop AI detector works.