The same piece of writing can receive different AI-detector scores for documented reasons: you changed the text, the length or format differs, a different tool or version was used, or a cutoff changed what is displayed. This guide lists the sources of variation we can support, says who documents each one, and notes what we could not find. It will help you interpret a changed score without assuming the change means the text became more or less human.
Key takeaways
- A different score can come from edited input, different length or formatting, a different tool, a version update or a different display cutoff.
- Turnitin says an existing submission’s score changes only if the document is resubmitted for processing. This does not establish repeatability for other products or versions.
- Do not read a changed score as proof of authorship in either direction.
- When you compare scores, record the tool, version or date, settings and the exact text.
Documented reasons scores change
Scroll horizontally to see the full table. Keyboard: focus the table and use the left and right arrow keys.
| Source of variation | What is documented | Source and qualification |
|---|---|---|
| Edited or transformed input | Paraphrasing and adversarial changes reduced detector performance in controlled tests. | RAID (2024) and Weber-Wulff et al. (2023). Study conditions, not a prediction for your text. |
| Text length | Turnitin requires at least 300 words of prose for a report and says very short documents are mostly all or nothing. QuillBot’s detector page states an 80-word minimum and says 300 or more words improve reliability. | Vendor statements, checked October 11, 2026. Product-specific. |
| Format and what counts as text | Turnitin analyzes prose sentences and does not reliably detect lists, code or short unconventional writing. Its percentage is not necessarily the share of the whole submission. | Vendor FAQ. Other tools may treat formats differently. |
| Cutoffs and display rules | Turnitin shows no percentage or highlights for results of 1% to 19% and marks them with an asterisk. | Vendor FAQ. One example of a display rule. |
| Model or version updates | Turnitin says changes to its model are likely to affect the percentage, and that it consolidated a multi-model ensemble into a single model in July 2026. A submitted document’s percentage changes only when resubmitted. | Vendor FAQ. Update frequency for other tools was not documented in the pages we reviewed. |
| Different tools | In one study, 89 of 91 human-written essays were flagged by at least one of seven detectors, and 18 by all seven. | Liang et al. (2023). An example of disagreement, not a general rate. |
Same text, different result
Turnitin says a submitted document’s percentage changes only if the document is resubmitted for processing, which means a later model update does not silently change an old result. This statement describes Turnitin; it does not establish whether identical text always returns an identical score in other products. If you see two scores for what you believe is the same text, check the date, version, settings and whether the text was truly identical, including hidden formatting.
Training and threshold choices
Different vendors train on different data and choose different cutoffs. A lower cutoff catches more AI text and flags more human text. A higher cutoff does the reverse. Turnitin says it works to keep false positives under 1% for documents with over 20% AI writing and accepts that it may miss some AI text as a result. That is a design choice that makes its numbers incomparable with a tool that chooses differently.
Length, format and context
Short samples, mixed content and non-prose elements are documented sources of instability for at least some tools. Before comparing two scores, check that the same text was scored the same way: same length, same formatting, same included sections, with quotations and headings treated alike.
Why revision changes the score
Editing changes the input, and detectors respond to input. That is documented in RAID and Weber-Wulff for transformed text in their tests. It is not evidence about how any named rewriting tool performs against any named detector now. We do not make that claim, and a lower score after editing does not show the text was made “more human” in any sense the detector can verify.
How buyers should read score changes
- Do not buy a tool because of one before-and-after score pair.
- Treat vendor-supplied detector figures as marketing unless the data and method are published.
- Compare edits for accuracy and meaning, not only for score.
- Remember detector results can change after a model update, so an old result may not hold.
A calmer way to work with changing results
Keep a simple log so a change can be explained rather than guessed at. Illustrative example, with invented values:
Scroll horizontally to see the full table. Keyboard: focus the table and use the left and right arrow keys.
| Date | Tool and version | Text length | Change since last run | Result |
|---|---|---|---|---|
| Oct 1 | Tool X, version unknown | 650 words | None | Score A |
| Oct 8 | Tool X, version unknown | 650 words | Two paragraphs edited | Score B |
| Oct 8 | Tool Y | 650 words | Same as above | Score C |
With a log like this, a changed score can be tied to an edit, a different tool or a time gap. For how scores are produced and how to read the output, see how AI detectors work. If a result affects you personally, see what a flag means and how to inspect it.
Before you compare tools
To compare rewriting tools, start with the ranking hub and the AIHT Rating page. The AIHT Rating is an editorial rating, not a detector pass rate.
Sources
- Turnitin, AI writing detection capabilities FAQs (vendor; page showed “Updated October 07, 2026”)
- QuillBot, AI content detector page (vendor; checked October 11, 2026)
- Dugan et al., RAID (ACL 2024)
- Weber-Wulff et al., Testing of detection tools for AI-generated text (2023)
- Liang et al., GPT detectors are biased against non-native English writers (2023)
Frequently asked questions
Why does one detector say AI while another does not?
They are different systems with different training and cutoffs. In Liang et al., agreement was far from complete: 89 of 91 human-written essays were flagged by at least one of seven detectors, but only 18 by all.
Do longer texts always get more stable scores?
Vendors say very short texts are less stable, and some set minimum lengths. That does not mean longer text is always stable, so treat length as one factor.
Can rewriting lower the score without changing the meaning?
Editing can change a score, as RAID and Weber-Wulff report for transformed text in their tests. Whether meaning survived is a separate question you must check against the original.
What should I trust more than a single score?
Context: the tool and date, the vendor’s definition of the number, and records of how the text was written, such as drafts and sources.