Turnitin does not keep every AI detection change on one page. The dated record is split across a model log, a general release log, a product updates page and a current FAQ. This table puts the changes in order and keeps four different things separate: the classifier, supported languages, the report interface and old bugs.
What changed in Turnitin AI detection in 2026?
Using a narrow rule, Turnitin's official pages document five relevant changes through 28 August 2026. The model was updated on 12 February, the Spanish detector was updated on 5 May, the FAQ records an architecture change in July, the English AI Writing Report merged its blue and purple categories on 4 August, and Modern Standard Arabic support arrived on 18 August. Those are five different kinds of change. The February and May entries update existing detection behavior, July changes model architecture, 4 August changes report display, and 18 August adds a separate Arabic-language model.
The rule for this changelog is simple: an entry has to change model behavior or architecture, supported languages, or how the AI Writing Report must be interpreted. New download buttons, login changes and unrelated Turnitin products are outside the table. Nothing below predicts whether a particular paper will receive a higher or lower score.
Four official sources, not one changelog
We read the rendered text of all four pages on 28 August 2026. Their jobs overlap, but they do not mirror one another.
| Official page | What it contains | What it misses |
|---|---|---|
| AI writing detection model | Dated model changes from 2023 onward. Its 2026 section has two dated entries: 12 February and 5 May. | It has no July architecture entry, 4 August category merge or 18 August Arabic launch. |
| Turnitin release notes | Shipped features with a date and a product license label. This is where the August category merge and Arabic launch appear. | It mixes AI changes with Feedback Studio, Clarity and other product releases. |
| Turnitin product updates | Rollout scope and the vendor's reason for a change. It adds details absent from the shorter release entries. | It is not a chronological model archive. |
| AI writing detection capabilities FAQ | The current mechanism and a July 2026 architecture statement. | The July statement has no day, and it does not appear on the model log. |
Calling any one of these pages "the release notes" hides the source boundary. The page name matters because the same date can carry the model change on one page, its product scope on another, and no entry at all on a third.
The 2026 Turnitin AI detection changelog
This table uses the inclusion rule above. The final column is there to stop a dated fact from being stretched past what the source says.
| Date | Type and official source | Published change | Scope and reading boundary |
|---|---|---|---|
| 12 February 2026 | Model update, AI writing detection model | Turnitin says it updated the AI writing detection model to improve recall while keeping its false-positive rate low. Settings and report highlighting did not change. | The entry does not name a language. It says old reports are not recalculated; an existing submission has to be resubmitted to use the updated model. It gives no basis for predicting the direction of one paper's score. |
| 5 May 2026 | Spanish model update, model log and product updates | The Spanish-language model was updated to detect content from newer GPT and Gemini model families. There was no visual change to the report. | Spanish submissions only. Previously generated reports stay as they were unless the file is resubmitted. |
| July 2026, no day published | Architecture, capabilities FAQ | The FAQ says a multi-model ensemble was consolidated into a single model. | This appears on the FAQ, not the model log. Turnitin says it retained its false-positive target, but publishes no before-and-after score table here. Do not infer how one document moved. |
| 4 August 2026 | Report interface, release notes and product updates | The AI Writing Report combined the former blue and purple categories into one blue category. Text suspected of being AI-generated and then paraphrased or processed by a bypasser remains inside that single category. | English submissions only. The release changes what the report displays, not the documented detection capability. Older reports keep purple until the file is resubmitted. This is an AI Writing Report change, not a Turnitin Clarity change. |
| 18 August 2026 | Language launch, release notes and product updates | Modern Standard Arabic joined English, Spanish and Japanese as a supported language. Turnitin describes the Arabic detector as a separate model. | The institution must license and enable AI writing detection. The launch also covers enabled iThenticate 2.0 customers. Separate language models should not be assumed to have identical capabilities. |
The five rows do not form a simple "old model versus new model" sequence. February and May concern model behavior, July concerns architecture, 4 August concerns report interpretation, and 18 August concerns which submissions can be processed.
Earlier changes that explain the 2026 report
The model page starts in May 2023. These are the earlier entries that changed classifier behavior, language coverage or the meaning of the report under the same rule used above.
| Date | Published change | How to read it now |
|---|---|---|
| 24 May 2023 | Detection logic was changed after Turnitin observed more false positives in the first and last few sentences, and segment boundaries were made more precise. A text-formatting bug was also fixed. | This is a record of changes already made. It is not evidence that introductions and conclusions remain a current weakness. |
| 9 August 2023 | Long-form prose inside tables became processable, the maximum sentence length was increased, and bibliographies were excluded after a highlighting bug. | Existing submissions had to be resubmitted to use the changes. The table entry applies to prose inside tables, not every table cell. |
| 6 December 2023 | Detection of likely AI text revised with what Turnitin calls an "AI word spinner" was added without a visual report change. | Earlier reports needed resubmission. This is the first dated paraphrasing-detection entry on the model page. |
| 15 December 2023 | Turnitin fixed a bug introduced on 6 December that had processed submissions below the 300-word file requirement. | The affected window was 6 to 15 December 2023. The record is historical and says the issue was resolved. |
| 16 July 2024 | The report added two interactive categories, raised the supported qualifying-text maximum to 30,000 words and stopped surfacing numerical scores from 1% to 19%. | The two-category interface was later replaced by the single blue category on 4 August 2026. |
| 12 September 2024 | A separate Spanish AI writing model launched. | A Spanish submission is not processed by the English detector. |
| 2 April 2025 | A separate Japanese AI writing model launched. | A Japanese submission is not processed by the English or Spanish detector. |
| 27 August 2025 | Likely AI text modified by bypasser tools was added to the report's "AI-generated only" category. | This capability later remained in place when the visible categories were merged in August 2026. |
| 14 October 2025 | The AI writing detection model was updated to improve recall, with no settings or visual changes. | Old reports were not recalculated. Existing submissions had to be resubmitted to use the update. |
The sequence explains two pieces of today's report. Separate language detectors arrived one at a time, and the blue-purple distinction had a two-year life before Turnitin removed it from the AI Writing Report.
The December 2023 short-document bug is historical
The 15 December 2023 entry is easy to misquote because its last paragraph says scores on submissions under 300 words may be less accurate. The heading and the paragraphs before it supply the missing frame: a bug had let files below the minimum through, the affected reports were generated from 6 to 15 December, and Turnitin had fixed it.
"This issue has now been resolved."
The current AI Writing Report guide requires at least 300 words of long-form prose before a report is generated. The old sentence therefore does not describe a present-day score below that floor. It also says nothing about the accuracy of a current 300-word or 500-word paper.
The AI Writing Report is not Turnitin Clarity
Turnitin's general release page covers several products on one screen. The product label attached to an entry is part of the fact, not small print.
| Change | Product labels on the official pages | Boundary |
|---|---|---|
| 4 August category merge | The release notes name Feedback Studio with the Originality add-on. The product updates page names Originality-enabled configurations and enabled iThenticate 2.0 customers. | The entry repeatedly names the AI Writing Report and says it applies to English submissions. It does not say that Clarity's Writing Report changed. |
| 18 August Arabic launch | The release notes name the Originality add-on. The product updates page names Originality-enabled configurations and enabled iThenticate 2.0 customers. | This is a language detector launch. It is separate from Clarity features published elsewhere on the same page. |
| February and May model updates | The model-specific page names the AI writing detection model; the May product entry narrows its update to Spanish submissions. | Neither entry is a suite-wide change to every Turnitin product or every language detector. |
This distinction matters most when search results drop a reader into the middle of the release page. A nearby Clarity heading does not extend the scope of an Originality entry, and the word "purple" in another interface does not turn it into the retired AI-paraphrased category.
How to compare two reports without guessing
Before comparing two AI percentages, record enough information to tell a changed document from a changed detector.
- The generation date printed on each report.
- The exact product and report name, including whether it is the AI Writing Report or a Clarity Writing Report.
- The submission language and whether the institution had AI writing detection enabled.
- Whether the same file was resubmitted after a dated model or report update.
- The category names and color legend printed inside each report.
- Whether the document bytes or exported file changed between submissions.
A score difference by itself cannot identify the cause. It may follow an edited file, a resubmission under a newer model, a different language detector or more than one of those at once. Why Turnitin detection changes over time explains the resubmission part; this changelog supplies the dates to check against it.
If revision is permitted
During an active academic-integrity review, preserve both artifacts exactly as submitted: the document and its original report. They record what the instructor reviewed. A replacement file and a newly generated report do not describe that same review.
When revision is allowed and the aim is to limit collateral changes, HumanPen can import the document with its AI report and keep the rewrite scope to the highlighted paragraphs you confirm.
That workflow controls which text is touched. It does not predict a future detector result, and this changelog does not turn a later score into proof that one particular update caused it.
Frequently asked questions
Does an old Turnitin AI report update automatically? No. The dated model entries repeatedly say previously generated reports are not recalculated. A file has to be resubmitted to generate a report under the current model.
Did Turnitin stop detecting paraphrased or bypassed AI text on 4 August 2026? No. The release entry says the capability remains; the report now draws one blue category instead of separate blue and purple categories. The full blue-purple change is covered separately.
Is the Arabic detector the same as the English detector? No. Turnitin calls Arabic, English, Spanish and Japanese separate models. The 18 August launch expanded language coverage rather than replacing all four with one multilingual model. How the language detectors differ has the current scope.
Does a model update mean my score will rise? The official pages do not support that prediction. They say reports may differ when a file is resubmitted under an updated model, but the direction for one document is unknown.
Does the 2023 note prove that current scores below 300 words are less accurate? No. That line describes reports created during a resolved bug window when files below the current minimum were processed by mistake. Under today's published requirement, a file needs at least 300 words of qualifying prose to generate the report.
Originally published at HumanPen: https://humanpen.net/blog/turnitin-detection-model-changelog?utm_source=devto&utm_medium=article&utm_campaign=syndication-2026-09
Top comments (0)