Direct answer
An AI detection change is any shift in what a detector reports on the same or similar text, and it almost always comes from one of three sources: the detection model itself was updated, the submission was a new run rather than an edit to the old report, or the document's content changed in ways that move segment-level scoring. Turnitin's AI writing indicator is a percentage with highlighted segments, not a verdict, and Turnitin positions it as a signal that requires human judgement rather than a definitive finding [1]. That framing matters because it explains why "the score changed" is usually not a malfunction — it is the detector doing exactly what its design allows.
Why Did My Turnitin AI Score Change Between Checks?
The most common reason a score moves between two checks is that the second check is a genuinely new submission, not a revision of the first. Turnitin computes the AI writing percentage per submission against the model in use at that moment, so identical text run twice is two separate scoring events rather than one score being edited [2]. Nothing about the first report is amended; a second report is generated alongside it, and the two can legitimately disagree.
The second reason is granularity. The headline percentage is an aggregate of segment-level predictions, and individual highlighted passages can move in or out of the flagged set even when the overall number shifts by only a few points [2]. A reader who compares two reports side by side often notices the highlighted spans differ before they notice the percentage does, and that is expected behaviour rather than inconsistency.
The third reason is the text itself. Editing, reordering, or replacing passages — including light paraphrasing that keeps the same sentence architecture — changes the input the detector evaluates, so a different input producing a different output is the correct outcome [2]. Students who "fix" a flagged paragraph and resubmit are not observing a detector change at all; they are observing their own edit being scored.
Does Turnitin's AI Detector Update, and Do Old Reports Change Too?
Yes, the detector updates — and no, old reports do not change with it. Turnitin has shipped successive detection model versions over time, and the model version that produced a report is recorded on that report, which is precisely why two reports on the same text can disagree without either being wrong [3]. The version stamp is the audit trail: it tells you which detector produced the number you are looking at.
What does not happen is retroactive recalculation. When the model is updated, reports already generated are not silently rewritten to match the new model, so a report you downloaded months ago still reflects the model of that period [3]. This is a deliberate design choice with real consequences: an archived report is a historical record, not a live measurement, and treating it as current is the single most common source of confusion about an "AI detection change."
The decision to re-run anything sits with the institution, not with the detector. Turnitin provides the indicator and the version information; whether a department re-checks existing submissions, and under which model, is an academic policy question [3]. That is why the same essay can carry one score in a module that re-ran everything after an update and a different score in a module that did not.
Can You See Your Current Turnitin AI Score Before It Counts?
You can, and doing so is the cleanest way to separate a real detection change from a stale number. The key property to look for is a non-repository check: the file is evaluated without being added to Turnitin's student paper database, so previewing your draft does not create the very match that would later appear in your instructor's similarity report [4]. A preview that pollutes the database is worse than no preview.
A useful preview also delivers both reports together. The AI writing percentage and the similarity score answer different questions — one about how the text was produced, the other about where it came from — and a change in one does not imply a change in the other [4]. Seeing them side by side prevents the classic misreading in which a rising similarity figure is mistaken for an AI detection change.
The final property is currency. A preview is only comparable to what your instructor will see if it is produced by the current model version, because a report from an older model is a historical artefact rather than a prediction [4]. Checking under the current detector, shortly before you submit, is what makes the number you see meaningful.
If the uncertainty is really about which number you are looking at, the fastest way to settle it is to look at a real report generated under the current detector rather than reasoning from an old one. turnitin0 produces the same two PDFs your instructor sees in the LMS — the Turnitin AI detection report and the similarity report — from a non-repository check, so the file is not added to the student paper database and the result reflects the model in use today.
※ Turnitin0.com - Actual Turnitin AI Report Cover, Score, Flag And Similarity Summary
FAQ
Why would the same essay get two different AI scores?
Because each check is a separate scoring event against the model in use at that time, and the highlighted segments can shift between runs [2]. Neither report is corrupt; they are measurements taken at different moments.
Does Turnitin recalculate my old report after a model update?
No. Reports already generated are not retroactively rewritten, and each report records the model version that produced it, so an archived report remains a historical record [3].
Is a changed score evidence that I used AI?
Not on its own. Turnitin describes the AI writing indicator as something that requires human interpretation rather than a definitive finding [1], and a moved percentage is explained equally well by a model update, a resubmission, or an edit to the text.
Can I see the current score without adding my file to Turnitin's database?
Yes — a non-repository check evaluates the document without adding it to the student paper database, and it returns the AI report and similarity report together [4].
Does the AI score and the similarity score change for the same reasons?
No. They measure different things — how the text was produced versus where it came from — so a shift in one does not imply a shift in the other [4].