DEV Community: Rayna Rabon

Rechecked After Humanizing and Still Flagged? What Turnitin Says

Rayna Rabon — Sat, 05 Sep 2026 13:01:13 +0000

You ran your text through a humanizer, rechecked it, and the AI score is still there. The FAQ says the detector can identify text that has been modified by AI paraphraser or bypasser tools. The tool list is not disclosed. Bypasser detection has been expanded over time. The mechanism is segment-based word probability. The score is not the sole basis for action. Here is what the documentation says.

What the FAQ Says About Paraphrasers and Bypassers

Turnitin's FAQ states that its detector can identify text that has been modified to evade detection:

"Furthermore, it can also identify instances where AI-generated text may have been modified by AI paraphraser or bypasser (also called humanizers) tools to evade detection."

This means the detector is not limited to identifying raw AI output. It also targets text that has been processed by paraphrasing or bypasser tools. If your text was flagged after being run through such a tool, this is consistent with what the FAQ documents.

The FAQ does not name which tools it can detect:

"Our AI writing detector has been trained and tested to detect leading paraphraser and bypasser tools. However, to safeguard the integrity of our solution and its effectiveness in maintaining academic honesty, we're unable to disclose the names of these tools."

You cannot look up whether a specific tool is on the list. The FAQ says the detector covers "leading paraphraser and bypasser tools" but does not disclose which ones. Does Turnitin detect AI humanizers collects everything the vendor has said on that.

Bypasser Detection Has Expanded Over Time

The FAQ's documentation shows that bypasser detection has been expanded in stages. A 2023 update stated:

"AI word spinners are sometimes used to avoid identification of AI-generated text. We now detect likely AI-generated text even if it may have been paraphrased using an AI word spinner."

A 2025 update added:

"With this release, the 'AI-generated only' category in the AI writing report will now include the percentage of AI-generated text that may have been modified by an AI bypasser tool."

This means the detector's coverage of paraphrased and bypassed text has grown over time. A recheck that produced a different score than an earlier check may reflect a model update that expanded bypasser detection. I paraphrased AI text with another AI and Turnitin still flagged it is the same wall from the other side.

How the Detector Scores Your Text

The detection mechanism works at the segment level:

"When a paper is submitted to Turnitin, sentences from the submission are extracted and segmented into overlapping sections for prediction analysis. Each segment is classified by the AI detection model and given a value between 0 and 1, denoting the probability of the text being likely human or AI-generated."

Each segment is classified independently. The overall percentage reflects the proportion of segments the model classifies as AI-generated. When you modify text and resubmit it, the model re-evaluates each segment. A segment that was previously scored as human could be reclassified based on the new text patterns.

The segment-based approach means that partial modifications may not shift the overall percentage as much as expected. If some segments still exhibit patterns the model associates with AI-generated text, those segments will continue to be flagged regardless of changes made elsewhere in the document. That is the case for rewriting only the paragraphs a Turnitin report flagged rather than the whole file.

The Score Is Not the Sole Basis

The FAQ explicitly cautions against treating the percentage as a definitive measure:

"Hence, we must emphasize that the percentage on the AI writing indicator should not be used as the sole basis for action or a definitive grading measure by instructors."

The score is a data point, not a verdict. A separate guidance page, "How should I review the AI Writing report?", frames the score as one input among many:

"It is not meant to provide definitive answers in isolation. More important than any tool is the educator who sees the score and makes decisions balancing this information with their personal knowledge of their students, their work, and institutional policy."

If your text is still flagged after a recheck, the documentation says the score should be considered alongside context, not treated as the final word.

What This Means for You

To summarize what the FAQ documents:

The detector can identify text modified by AI paraphraser or bypasser tools.
The FAQ does not disclose which specific tools it can detect.
Bypasser detection has been expanded over time, with updates in 2023 and 2025.
The detector works segment by segment, assigning each a probability score between 0 and 1.
The percentage should not be used as the sole basis for action or a definitive grading measure.
The score is one input meant to be considered alongside the educator's knowledge of the student and institutional policy.

If you received a Turnitin AI report and want to address the flagged passages, import the report and work on them. Where a passage qualifies, sending it through again costs nothing.

Originally published at HumanPen: https://humanpen.net/blog/recheck-after-humanizing-still-flagged?utm_source=devto&utm_medium=article&utm_campaign=syndication-2026-09

How to Lower a Turnitin AI Score Without Breaking Your Citations

Rayna Rabon — Sat, 05 Sep 2026 12:56:28 +0000

To lower a Turnitin AI score without breaking citations, revise the flagged prose around each citation, not the citation itself. Freeze quotations, locators, numbers, claim limits, and reference-manager fields before editing. Then compare every changed paragraph with the original, reopen each source, and refresh the citation links in a copy. No workflow can promise the next score.

Before you edit

This is a revision method for work you wrote and are allowed to revise. It is not a way to make borrowed writing defensible, and it cannot override a course, journal, or disclosure rule. If the rule does not permit AI-assisted rewriting at this stage, stop at manual revision.

Do not treat the score, the sentence, and the citation as one object

A Turnitin highlight points to prose that the model classified. A citation marker points to a source. The sentence between them makes a claim. Those three things can occupy the same line while needing three different checks.

Turnitin says its AI percentage is based on qualifying text, which it describes as prose in standard grammatical sentences, and that the percentage is not necessarily the share of the whole submission. Its model release notes also say bibliographies are excluded when the AI writing report is processed. The same note says an older affected submission must be resubmitted to be reprocessed.

So the edit target is the highlighted prose. Do not rewrite a reference entry in the hope that it will lower the AI percentage. The reference list has a different job, and changing it can create a citation error without addressing the reported passage.

Object	What must remain true	How it can fail during revision
Report highlight	You can locate the same passage in the source document	A PDF highlight is copied as a fragment and the surrounding paragraph is missed
Claim	Population, relationship, strength, conditions, and polarity stay intact	"Associated with" becomes "caused" or a limited sample becomes everyone
Citation	The source still supports the sentence beside it	The marker survives after a sentence split but now appears to support a different claim
Locator	Page, section, timestamp, or paragraph still identifies the evidence	A quotation is shortened while its page number is carried over without checking
Citation object	The reference manager still recognizes and updates it	A live field becomes ordinary text that looks identical on screen
Reference entry	One correct entry still resolves from the in-text citation	An author or year changes in one place and creates an orphaned reference

Make a citation lock sheet before changing a word

Do this only for report-highlighted paragraphs that contain citations. It takes less time than reconstructing a broken evidence chain afterward.

Paragraph	Claim that the source supports	Source and locator	Text or data that is locked	Citation object before edit
P12	Late buses are associated with missed seminars among first-year commuters	Chen, 2023, p. 41	`associated`, population, page 41	Active manager citation
P27	The intervention did not change the primary outcome	Malik, 2022, Table 3	`did not`, outcome name, table number	Active manager citation
P34	Direct quotation	Ortiz, 2021, p. 88	Every quoted character, quotation marks, locator	Active manager citation

The rows above are invented examples, not research findings. Replace them with the obligations in your own paper. A locked item is not necessarily text that can never move. It is an item that cannot change without returning to the source, data, or reference manager first.

The sheet catches a failure that a bracket count misses. Suppose the original says:

Late buses were associated with missed seminars among first-year commuters (Chen, 2023, p. 41).

An unsafe rewrite might say that unreliable transport caused students to miss class. It sounds plausible, but it changes association to causation and drops the population boundary. A defensible revision could change the rhythm while keeping the obligations visible:

In Chen's sample of first-year commuters, late buses and missed seminars were associated (Chen, 2023, p. 41).

The second version still needs a source check. The example only shows what must be preserved; it does not establish that the fictional source exists.

1. Preserve the evidence state

Use the lock, revise, verify loop.

Keep the report, the exact submitted or checked document, and the last editable source file. Make a separately named working copy. If you are responding to an accusation rather than preparing an authorized revision, preserve the files and version history before editing anything.

2. Map highlights to full paragraphs

Open the report beside the source document and locate each highlighted passage. Put the full paragraph on the edit list, even if the visible highlight starts halfway through a sentence. Include the paragraph before and after in the review view, but do not add them to the rewrite scope automatically.

This is a scope decision, not a detector claim. Using a full paragraph avoids stitching two voices into one sentence. It also keeps unrelated citations in nearby paragraphs outside the edit.

3. Record the citation obligations

Fill one lock-sheet row for every citation-bearing paragraph in scope. Record the claim, the source passage or locator, numbers, quotations, named methods, technical terms, negatives, qualifiers, and the current citation-object state.

If you cannot say what a source supports, pause there. Rewriting that sentence first only makes the uncertainty harder to see.

4. Revise from the source and your notes

Read the source passage and your own notes, then restate the claim. Do not walk through the old sentence swapping words one by one. That often keeps the surface pattern while quietly changing the verb, scope, or emphasis that mattered.

Leave quotations unchanged unless you return to the source and choose a different exact excerpt. Keep numbers and units paired. Keep hedges such as may, in this sample, and was associated with unless the evidence supports a stronger statement.

5. Compare before accepting

Use a side-by-side diff or Word's Compare command, which can place differences in a third document without changing the two source files when configured that way. Review deletions before insertions. A missing not, page number, or condition is usually more consequential than a new transition.

For each paragraph, choose one action: accept, repair, or revert. Fluency alone is not a pass condition.

6. Reopen the cited source

Read the revised sentence and the source passage together. Ask:

Is the attributed author still responsible for this claim?
Is the relationship unchanged, especially association versus causation?
Is the population, time window, jurisdiction, or experimental condition still present?
Does the locator still point to the evidence or exact quotation?
If several sources share one marker, does each still support the part assigned to it?

If any answer is uncertain, revert the sentence or rewrite it again from the evidence. A lower score would not repair an unsupported citation.

7. Test the citation machinery in a copy

A marker that looks right can still be dead text. If the document uses Zotero, EndNote, Mendeley, or another manager, use that manager's refresh or update path on a copy and confirm that the edited citations and bibliography still respond.

For Zotero, its Word plugin documentation says Refresh updates citations and the bibliography from the library. It also warns that unlinking citations removes field codes and prevents further automatic updates. Do not flatten citations merely to make editing easier.

For Word fields on Windows, Microsoft documents field updates and notes that fields inside tables may need separate selection. Content controls, bookmarks, and fields are not interchangeable, so use the procedure for the reference manager and Word version that created the document.

8. Run the acceptance gate

Do not submit because the file opens or because every pair of parentheses survived. The changed file passes only when all of these are true:

Check	Pass condition
Scope	Every prose change is inside a paragraph you deliberately put in scope
Meaning	Claim strength, polarity, population, conditions, numbers, and terms still match
Evidence	Each revised claim is supported by the cited passage
Quotations	Wording, quotation marks, attribution, and locator match the source
Citation links	The reference manager recognizes the citation in a copy
Bibliography	Each in-text citation resolves to the intended entry and no required entry is orphaned
Document behavior	Fields, cross-references, notes, tables, and contents update without new errors
Rendered output	The final DOCX and exported PDF show no clipped text, detached markers, or broken pages

Only after this gate should an authorized recheck happen. Record the result as an observation about that file and that report, not as proof that the same edits will produce the same number later.

What to do if the score is still high

Do not respond by widening the rewrite to the whole paper. Read the new report and build a new edit list from what it marks. Some paragraphs may be unchanged, some may need a second evidence-led revision, and some may be better left alone because changing them would weaken the argument or distort a source.

If the report and the visible highlights seem out of proportion, remember the denominator. Turnitin's AI writing FAQ says the percentage is calculated from qualifying prose and is not necessarily the percentage of the whole document. Do not infer a paragraph count from the number alone.

Software can reduce the passage-matching work, but it cannot perform the source check for you. I work on HumanPen. For an English-language document, its public workflow can import an existing Turnitin or iThenticate report, show the full-paragraph scope for confirmation, and keep text outside that confirmed scope out of the rewrite. That controls the edit surface. It does not guarantee a detector result or prove that a revised claim is supported.

Protect the evidence chain first. Keep the edit scope small, then treat any score movement as a later measurement. If a sentence cannot be changed without making its citation less accurate, keep the defensible sentence.

Originally published at HumanPen: https://humanpen.net/blog/lower-turnitin-ai-score-without-breaking-citations?utm_source=devto&utm_medium=article&utm_campaign=syndication-2026-09

Originality.ai vs Turnitin AI Detection: What the Documentation Says

Rayna Rabon — Sat, 05 Sep 2026 12:45:43 +0000

Students sometimes run their paper through Originality.ai and get a different score than Turnitin. This is expected. Different detectors use different models trained on different data. Turnitin's FAQ describes its mechanism, false positive targets, and the asterisk convention for low scores. Here is what the documentation says.

The Short Answer

Originality.ai and Turnitin are different AI detectors built on different models with different training data and different classification thresholds. A score from one will not predict the score from the other. Turnitin's FAQ describes its mechanism: text is split into overlapping segments, each classified with a probability score between 0 and 1. The FAQ also states a false positive target of under 1% for documents with over 20% AI writing, and uses an asterisk convention for scores in the 1-19% range. Originality.ai publishes its own accuracy claims, but those are vendor-reported and cannot be independently verified against Turnitin's documentation. The score that matters is the one from the tool your institution uses. If your school uses Turnitin, no third-party score can tell you what your Turnitin score will be.

How Turnitin's Detector Works

The FAQ describes the detection process:

"When a paper is submitted to Turnitin, sentences from the submission are extracted and segmented into overlapping sections for prediction analysis. Each segment is classified by the AI detection model and given a value between 0 and 1, denoting the probability of the text being likely human or AI-generated."

Segments overlap, meaning sentences can receive multiple scores that are pooled together. This pooling smooths individual prediction errors and contributes to the overall document score. Originality.ai uses its own classification approach, which is not documented in Turnitin's materials. The two systems are built differently and evaluate text differently.

Turnitin's False Positive Target

The FAQ states a specific target:

"We strive to maximize the effectiveness of our detector while keeping our false positive rate - incorrectly identifying fully human-written text as AI-generated - under 1% for documents with over 20% of AI writing."

The next sentence: "In other words, we might flag a human-written document as AI-written for one out of every 100 fully-human written documents."

This target is specific to Turnitin and applies to documents with over 20% AI writing. Originality.ai may have a different false positive rate, but comparing the two numbers requires understanding how each vendor defines and measures false positives. We do not make accuracy claims about third-party detectors. What the documentation tells us is that Turnitin has explicitly stated this target for its own model. What that target costs at scale is worked out in what a 1% false positive rate means when a university submits 75,000 papers.

The Asterisk Convention

Turnitin has a specific display rule for low scores:

"To avoid potential incidence of false positives, no score or highlights are attributed for AI detection scores in the 1% to 19% range. When AI is detected below the 20% threshold in the report, it is now indicated with an asterisk (*%) and no percentage is attributed."

This means a Turnitin score showing "*%" indicates AI was detected below 20%, but the exact percentage is withheld to avoid overstating accuracy in a range prone to false positives. Originality.ai does not use this convention and will display a specific percentage in this range. A paper showing "*%" on Turnitin might show "15%" on Originality.ai. This does not mean one tool is more accurate than the other. It means they handle the low-confidence range differently. What the asterisk (*%) means on a Turnitin AI score covers the Turnitin side of it.

Why Scores Disagree Between Tools

Different AI detectors produce different scores on the same text because they use different models, different training data, and different thresholds. Turnitin's FAQ describes how its model pools overlapping segment scores into a document-level percentage. A different detector might use a different segmentation strategy, a different scoring scale, or a different aggregation method. The FAQ also notes that short documents produce all-or-nothing predictions: "In shorter documents where there are only a few hundred words, the prediction will be mostly 'all or nothing' because we're predicting on a single segment without the opportunity to overlap." Originality.ai may or may not have the same behavior on short documents.

The AI score and the similarity score are also independent on Turnitin: "The Similarity score and the AI writing detection percentage are completely independent and do not influence each other." A detector that combines AI and similarity into a single score will produce a number that is not comparable to Turnitin's separate AI percentage. Why the same text scores differently on every detector collects the reasons.

What This Means for You

To summarize what we have covered:

Originality.ai and Turnitin are different detectors with different models. Scores will not match.
Turnitin's detector splits text into overlapping segments and assigns probability scores between 0 and 1.
Turnitin's false positive target is under 1% for documents with over 20% AI writing.
Turnitin uses an asterisk for scores in the 1-19% range to avoid overstating accuracy.
Short documents get all-or-nothing predictions on Turnitin due to single-segment evaluation.
The score that matters is the one from the tool your institution uses.

If you receive a Turnitin AI report and want to address the flagged passages, import the report and work on them. Passages that qualify can go through again without a further charge.

Originally published at HumanPen: https://humanpen.net/blog/originality-ai-vs-turnitin-ai-detection?utm_source=devto&utm_medium=article&utm_campaign=syndication-2026-09

Why AI Detectors Flag Well-Written Essays

Rayna Rabon — Sat, 05 Sep 2026 12:35:24 +0000

The question assumes the detector read your essay and marked it down for reading too well. Nothing the vendor publishes about the tool has anywhere to put "well written", which changes what a flag can be telling you.

The short answer

It is not grading the writing. Across the three Turnitin help pages we searched, "quality", "style", "polished", "well written", "vocabulary" and "clarity" were absent. The sensitivity control "300 words" was present on all three pages, so the search path was reading page text rather than returning blanks. The FAQ does not frame the detector as a measure of how good the writing is. What Turnitin does publish is a non-quantified list of text that false positives can include. Careful academic prose can resemble two items on that list when it repeats fixed terms or uses parallel structures, but Turnitin does not say how often those properties appear among errors.

That does not prove polish is harmless, and we are not going to pretend it does. What a company documents is the part it chose to document, and a trained model contains a great deal more than its help pages do. It establishes something narrower and more useful: "well written" is not a quantity this system reports on, so a flag is not a verdict on it, and the advice people build on the assumption that it is, which always ends at writing below your own level, has nothing published behind it.

Counting the vendor's own vocabulary

We searched the rendered text of three pages retrieved on 18 August 2026 because guides.turnitin.com returned 403 to a plain command-line fetch: "Using the AI Writing Report", "Turnitin's AI writing detection capabilities FAQs", and "File requirements for an AI Writing Report". Those pages set out the mechanism and file rules; release notes and known-issues pages were outside this search. The table records case-insensitive presence or absence rather than exact counts, which can drift as live pages change.

Search term	Report guide	Detection FAQs	File requirements
`quality`	Absent	Absent	Absent
`style`	Absent	Absent	Absent
`well written` / `well-written`	Absent	Absent	Absent
`polished`	Absent	Absent	Absent
`genre`	Absent	Absent	Absent
`discipline`	Absent	Absent	Absent
`false positive` (secondary control)	Present	Present	Absent
`300 words` (sensitivity control)	Present	Present	Present

The bottom row is the reason the rest of the table means anything: an empty result and a broken retrieval leave the same trace, so a term known to sit on all three pages has to be present first. On the FAQ, "writing" and "segment" are present, while "vocabulary", "word choice" and "clarity" are absent. "Breakdown" is present on the report guide, but only as the name of an interface panel, not as a statistic.

Then there are the pictures, which a text search cannot see at all. The FAQ article itself contains no images; the site logo belongs to the page chrome, not the article. The report guide carries screenshots with no alt text, so we opened them separately. They include a sample report showing 56% AI beside 23% similarity, a page of highlighted prose, and status messages. None uses any of the words above. One does carry a sentence that appears nowhere in the page's own text:

"AI detection includes the possibility of false positives. Although some text in this submission is likely AI generated, scores below the 20% threshold are not surfaced because they have a higher likelihood of false positives."

That is the product declining to print a number in the band where it trusts itself least, and it is worth holding on to, because everywhere else in this material the error rate is described rather than acted on.

What it does have words for

Something is doing the work, and the FAQ names it under the heading "What parameters or flags does Turnitin's model take into account when detecting AI writing?":

"Our classifiers are trained to detect these differences in word probability and are adept at the particular word probability sequences of human writers."

The same answer is careful to add that the model "is not explicitly programmed to evaluate specific signals such as 'burstiness,' 'perplexity,' or other individual metrics sometimes referenced in public discussions", and that it "learns statistical patterns from our training data" instead. Both sentences are on the page and reading either without the other gets you somewhere wrong: the vendor is denying two named public metrics, not denying that word probability is the currency. What that means for the terms people argue about is laid out in what AI detectors measure.

Now follow the published units upward rather than downward. Words, then sentences, then this:

"When a paper is submitted to Turnitin, sentences from the submission are extracted and segmented into overlapping sections for prediction analysis. Each segment is classified by the AI detection model and given a value between 0 and 1 ... Each qualifying sentence within these segments inherits the segment's score. Since segments overlap, some sentences may have multiple scores, which are then pooled into a single score. These sentence scores are further aggregated and used to compute the overall document AI writing score."

The chain stops there. The largest object that is ever classified is a segment, a handful of sentences wide, and the document number is an aggregation of those, not a second judgement made after reading the whole thing. Almost everything that makes an essay good is bigger than a segment: an argument that holds for eight pages, evidence chosen well, an objection raised in section two and answered in section five. None of that is an object at the scale where classification happens. It is not that the system weighs your reasoning and dislikes it. There is no slot.

What "human" is calibrated against

When discussing false positives in academic prose, the vendor points to a pre-ChatGPT test corpus:

"To bolster our testing framework and diagnose statistical trends of false positives, before every update or new model release, we perform tests on over 700,000 additional academic papers that were written before the release of ChatGPT to further validate our less than 1% false positive rate."

The FAQ asks itself the same thing again under its own heading, "How does Turnitin ensure that the false positive rate for a document remains less than 1%?", and repeats the figure with the word "over" dropped: "we perform additional tests on 700,000 additional academic papers that were written before the release of ChatGPT". Same corpus, two slightly different sentences on one page.

Read how Turnitin describes that reference population: academic papers selected because they were written before ChatGPT's release. That is a date-based description, not a statement that every paper was independently verified as human-authored. The narrower claim is still useful: Turnitin says it re-runs this pre-ChatGPT academic corpus before model releases to watch the false positive rate. The page does not describe paper-by-paper authorship verification.

Which does not settle it, and we would be overclaiming if we said it did. The vendor publishes the target and not the distribution behind it, and the target itself carries a limit that its own next sentence quietly drops. The full version of that argument, including what happens when you multiply a small rate by a real university's annual volume, is in what a 1% false positive rate means.

The error distribution is not published

The page states a false-positive rate of less than 1%, but it does not publish how errors are distributed across the test corpus. It gives no subgroup rates and does not show whether any kind of text accounts for more errors than another.

What the page does publish is a non-quantified list of text that false positives can include, followed by advice for reading the percentage:

"Sometimes false positives (incorrectly flagging human-written text as AI-generated), can include content without a lot of structural variation, text that literally repeats itself, or text that has been paraphrased without developing new ideas. If our indicator shows a higher amount of AI writing in such text, we advise you to take that into consideration when looking at the percentage indicated."

Read what is on that list and what is not. Not sophisticated vocabulary, not long sentences, not a strong argument. It names three possible text properties, but gives no frequency for any of them. The third describes a process rather than anything you would have produced writing on your own.

The first two are worth comparing with ordinary academic practice, and what follows is our reading, not the vendor's. If your field asks you to call one thing by one name every time it appears, you may repeat the same term on purpose because a synonym would introduce ambiguity. If you set comparable items in parallel sentences so a reader can hold them side by side, you may reduce structural variation for the same reason. Neither choice is a lapse. Turnitin says false positives can include these properties. It does not say how often, that errors are concentrated there, or that the properties correlate with errors.

What we are not saying, because it is a much larger claim than any of this supports, is that polished writing gets flagged. That sentence circulates widely, and what it usually arrives attached to is an instruction to sound less capable than you are. The edits the published list actually supports, and what doing them by hand costs, are worked through in how to humanize AI text without a tool.

Notice too who that second sentence is written for. It asks whoever is looking at the percentage to weigh what kind of text produced it. The instruction is pointed at the person holding the report, not at the person who wrote the paper.

Whether your kind of writing is worse off is not published

The obvious next question is whether false positives are more common in your discipline, your section, or your genre. The FAQ has a heading that comes close, "Does the Turnitin model take into account that AI writing detection technology might be biased against particular subject-areas or second language writers?", and the answer is about the training sample:

"One of the guiding principles of our company and of our AI team has been to minimize the risk of harm to students, especially those disadvantaged or disenfranchised by the history and structure of our society. Hence, while creating our sample dataset, we took into account statistically under-represented groups like second-language learners, English users from non-English speaking countries, students at colleges and universities with diverse enrollments and less common subject areas such as anthropology, geology, sociology, and others."

The subject names "anthropology", "geology" and "sociology" appear only in paragraphs describing how the sample was built. "Discipline", "genre" and "subgroup" are absent from all three pages. So subject areas appear as an input to training, not as a published testing breakdown. There is no published false-positive rate for anthropology, and the vendor puts no numbers behind the bias answer, which is why it has to be read as a statement of intent.

That gap is why the figure everyone reaches for on this question comes from outside the company: a 2023 study on 91 TOEFL essays, whose scope, controlled edits and real limits we went through in why non-native English writers get flagged by AI detectors more often. It is also why two detectors can hand the same paragraph two different verdicts, which is a separate mess covered in why detectors disagree.

Some of your best work is not in the measurement at all

There is one more reason "it flagged my good essay" is the wrong frame, and it is about what the percentage is a percentage of. Only some of your document is eligible to be scored:

"This qualifying text includes only prose sentences, meaning that we only analyze blocks of text that are written in standard grammatical sentences and do not include other types of writing such as lists, bullet points (short non-sentence structures), or other non-sentence structures. This percentage is not necessarily the percentage of the entire submission."

Two changes announced in August 2023 moved the boundary further. One says long-form prose inside tables is now processed; the other says a bug that sometimes highlighted AI writing inside a bibliography was fixed and bibliographies are now excluded from processing. Both of those release notes end the same way, telling you to resubmit an existing paper before either change applies to it.

So the bibliography you spent two evenings aligning and inclusion criteria written as short, non-sentence bullets sit outside the scored set. Tables are conditional: long-form prose inside them can be processed, while numeric cells, fragments, and other non-sentence table content may remain outside qualifying prose. The percentage therefore need not describe the whole submission. That is also the mechanical reason it can disagree with the amount of coloured text in front of you, which we take apart line by line in how to read a Turnitin AI writing report.

What to do with all this

Very little of it is about how you write, and most of it is about how much weight the number can carry.

Sounding less capable is the one move to rule out. Everything above argues against it, and the trade is lopsided in an ordinary way: an examiner's marks land for certain, while the effect on the score is something nobody has measured and you are usually not allowed to watch.
Ask which passages, not what number. The highlights tell you where the classifier landed. The percentage tells you how much of the qualifying text it landed on, which is a different quantity from how much of your essay.
Expect a short document to be measured coarsely. The vendor's words: "In shorter documents where there are only a few hundred words, the prediction will be mostly 'all or nothing' because we're predicting on a single segment without the opportunity to overlap." The sentence after it is the one that matters: "This means that some text that is a mix of AI-generated and original content could be flagged as entirely AI-generated." Below 300 words of qualifying text, the file requirements say there is no report to read at all.
If your prose is conventional because the genre demands it, check who the vendor's advice is written for. That sentence is addressed to whoever reads the percentage, not to you. Worth knowing before you conclude that the paper is the thing that has to change.
Distrust any promise about the resulting number, ours included. Turnitin's own position is that individual predictions "may not always be explainable in simple feature-by-feature terms". Nobody outside the company can map an edit onto a movement in the score.

Where a rewrite fits, and where it does not

None of this is an argument for rewriting. Where the work is yours and you are prepared to say so, the paper is not the object that should be changing.

A rewrite earns its place at one specific point: when a report exists, particular passages are marked on it, and those passages are going to be revised anyway. That is the shape HumanPen is built for. Hand it the file plus the AI report from Turnitin or iThenticate and the marked passages become the scope, with everything outside them keeping the words you wrote. Our own page states the granularity rule: "A paragraph is the smallest unit the engine rewrites: if your selection covers only part of one, it is expanded to the full paragraph and shown that way for you to confirm." Billing counts the words actually rewritten, so a scoped job costs what the scope costs rather than what the file weighs.

If a new report still flags passages, eligible results can continue lowering AI for free. What nobody can hand you is the figure on the next report. After everything above, a tool that quotes you one is quoting a number it does not have.

Frequently asked questions

Is well-written prose more likely to be flagged? That is a bigger claim than the published material supports, and we are not making it. What is documented is a list of three kinds of text that Turnitin says false positives can include, none of which is a quality judgement, plus a measured effect on writers with a smaller lexical range in English from one 2023 sample. Treat the broad version as a claim in circulation rather than a finding.

Does using advanced vocabulary raise my AI score? Nothing the vendor publishes speaks to it: "vocabulary" and "word choice" were both absent from the three help pages we searched. One 2023 controlled experiment moved word choice directly and pushed against the folk theory rather than with it: the researchers enriched the wording of already-flagged essays, and the flags fell. That establishes that word choice was not inert in that experiment. It establishes nothing about your document.

My writing repeats itself because the terminology has to be consistent. That is the conventional case, and it answers to one item on the vendor's own false-positive list. Its advice in that situation is aimed at the person reading the report: take the nature of the text into consideration when looking at the percentage.

Should I add sentence-length variation on purpose? Not as a formatting exercise. Where variation was already carrying meaning, restoring it is ordinary revision. Manufacturing it in a section whose conventions require uniformity replaces writing your discipline asked for with writing it did not, in exchange for an effect nobody can measure from outside.

Can I check the score myself before I submit? Usually not. Turnitin's position is that the AI writing indicator and report are for instructors and administrators, and it reaches a student only if an instructor downloads the report and passes it on. Any advice that assumes you can watch the number move is assuming access most students do not have.

Originally published at HumanPen: https://humanpen.net/blog/why-polished-writing-gets-flagged?utm_source=devto&utm_medium=article&utm_campaign=syndication-2026-09

GPTZero vs Turnitin: Why Scores Won't Match and What Each Tool Actually Measures

Rayna Rabon — Sat, 05 Sep 2026 11:30:22 +0000

If you run the same text through GPTZero and Turnitin, the scores often will not match. This is not a sign that one tool is broken. Different detectors use different models, different training data, and different scoring conventions. Here is what Turnitin's documentation says about how its detector works and why cross-tool comparisons are unreliable.

Different Detectors Use Different Models

GPTZero and Turnitin are separate systems built by separate teams. Each has its own underlying model, its own training data, and its own approach to classifying text. When two detectors use different models, their outputs will not align on the same document. A text that GPTZero scores as 60% AI might receive a 30% score from Turnitin, or vice versa. Neither score is necessarily wrong. They are probability estimates from different models trained on different data.

There is no universal AI-detection score that all tools converge on, which is the shape of the problem in why the same text scores differently on every detector. Each tool produces its own estimate. Comparing scores across tools as if they measure the same thing in the same units leads to confusion.

We are not claiming one tool is more accurate than the other. The documentation describes how Turnitin's detector works, and that mechanism differs from what other tools do.

How Turnitin's Detector Scores a Document

Turnitin's documentation describes a specific process for arriving at a document-level AI writing score:

"When a paper is submitted to Turnitin, sentences from the submission are extracted and segmented into overlapping sections for prediction analysis. Each segment is classified by the AI detection model and given a value between 0 and 1, denoting the probability of the text being likely human or AI-generated. Each qualifying sentence within these segments inherits the segment's score. Since segments overlap, some sentences may have multiple scores, which are then pooled into a single score. These sentence scores are further aggregated and used to compute the overall document AI writing score."

The score is an aggregation of segment-level predictions. Segments overlap, and sentence scores are pooled before aggregation. Other detectors may segment text differently, use different prediction units, or aggregate scores in different ways.

Turnitin's documentation also clarifies that the AI writing detection percentage and the Similarity score are separate measurements. "The Similarity score and the AI writing detection percentage are completely independent and do not influence each other." The Similarity score indicates the percentage of matching text found in the submitted document when compared to Turnitin's comprehensive collection of content for similarity checking. A high similarity score does not mean a high AI score, and vice versa. Turnitin AI writing report vs Similarity report sets the two side by side.

Turnitin's False Positive Target

Turnitin states a specific false positive target in its documentation:

"We strive to maximize the effectiveness of our detector while keeping our false positive rate - incorrectly identifying fully human-written text as AI-generated - under 1% for documents with over 20% of AI writing."

The documentation then restates this as: "In other words, we might flag a human-written document as AI-written for one out of every 100 fully-human written documents." This restatement drops the "over 20% AI writing" qualifier. The version with the 20% threshold is the more precise statement.

To validate this rate, Turnitin describes a testing process: "To bolster our testing framework and diagnose statistical trends of false positives, before every update or new model release, we perform tests on over 700,000 additional academic papers that were written before the release of ChatGPT to further validate our less than 1% false positive rate."

This is Turnitin's stated target and validation process. What one percent actually costs at the scale a university submits is worked out in what a 1% false positive rate means. Other detectors may have different targets, different validation methods, or may not publish comparable figures. We cannot make a direct accuracy comparison between tools based on this information. We can only report what Turnitin's documentation says.

Short Documents Behave Differently

Document length affects how Turnitin's detector behaves. For short texts, the scoring mechanism can produce extreme results:

"In shorter documents where there are only a few hundred words, the prediction will be mostly 'all or nothing' because we're predicting on a single segment without the opportunity to overlap."

"This means that some text that is a mix of AI-generated and original content could be flagged as entirely AI-generated."

The overlap mechanism that smooths scores in longer documents does not function when there is only one segment to predict on. A few hundred words of mixed human and AI text could come back as 100% AI, one of the routes covered in why Turnitin says your work is 100% AI. This is a limitation on short inputs, not a definitive judgment about the text's composition.

If GPTZero scores a short document differently, the all-or-nothing behavior is one explanation. The two tools handle short texts differently because their scoring architectures differ.

The 1-19% Asterisk Range

Turnitin does not display a numeric score for AI detection results between 1% and 19%. The documentation states:

"To avoid potential incidence of false positives, no score or highlights are attributed for AI detection scores in the 1% to 19% range. When AI is detected below the 20% threshold in the report, it is now indicated with an asterisk (*%) and no percentage is attributed."

This means if Turnitin's model estimates the AI content at 12%, the report will show an asterisk rather than a number. No highlights are applied either. The reasoning is to avoid surfacing potential false positives in a low-confidence range, spelled out further in what the asterisk (*%) means on a Turnitin AI score.

GPTZero and other tools may display scores in this range differently. Some show a full percentage for any non-zero result. Turnitin suppresses the number below 20%. This is another reason scores will not match: one tool may show a numeric score where the other shows only an asterisk.

What This Means for You

If you are comparing GPTZero and Turnitin scores on the same document, here is what to keep in mind:

Different models, different scores. GPTZero and Turnitin use separate detection models trained on separate data. Their scores are not expected to match.
Turnitin scores by aggregating segments. Text is segmented, scored per segment, and aggregated. The AI score and Similarity score are fully independent.
Turnitin targets under 1% false positives for documents with over 20% AI writing, validated against over 700,000 additional academic papers written before ChatGPT's release.
Short documents get all-or-nothing scores. A few hundred words may produce a 0% or 100% result even when the text is mixed.
Scores below 20% show as an asterisk. No numeric percentage or highlights appear in the 1-19% range.

Eligible passages can be re-run at no charge.

Originally published at HumanPen: https://humanpen.net/blog/gptzero-vs-turnitin-accuracy?utm_source=devto&utm_medium=article&utm_campaign=syndication-2026-09

International Students and AI Scores: What to Know About Writing in a Second Language

Rayna Rabon — Sat, 05 Sep 2026 11:19:58 +0000

Writing academic English as a second language is hard enough without worrying about AI detection. Turnitin's FAQ says the training data included second-language learners to minimize bias. But the detector still reads word probability patterns, and certain writing characteristics are prone to false positives. Here is what that means for international students.

The Short Answer

Turnitin's FAQ says the model's training data "took into account statistically under-represented groups like second-language learners, English users from non-English speaking countries, students at colleges and universities with diverse enrollments, and less common subject areas such as anthropology, geology, sociology, and others to minimize bias when training our model." This is Turnitin's own claim, not an independently verified conclusion. The training set inclusion means the model was exposed to second-language writing during training. But the detector still operates by classifying word probability patterns in prose text. If your writing happens to share statistical patterns with AI-generated English, the detector can flag it regardless of whether English is your first language. Understanding this gap between the training claim and the detection mechanism is key to responding to a flag.

What Turnitin Claims About Bias

The FAQ addresses bias directly:

"While creating our sample dataset, we also took into account statistically under-represented groups like second-language learners, English users from non-English speaking countries, students at colleges and universities with diverse enrollments, and less common subject areas such as anthropology, geology, sociology, and others to minimize bias when training our model."

This is the company's own statement about its training methodology. It says the training data was designed to include diverse writing samples. But it does not provide any numbers, test results, or external validation of bias levels. The claim is that second-language writing was represented in training. What the claim does not say is that second-language writing is immune to false positives. The model was trained on diverse data, but it still classifies by word probability patterns. Why non-native English writers get flagged by AI detectors more often goes through the research on that gap.

How the Detector Reads Your Text

The detection pipeline works the same way regardless of who wrote the text:

"When a paper is submitted to Turnitin, sentences from the submission are extracted and segmented into overlapping sections for prediction analysis. Each segment is classified by the AI detection model and given a value between 0 and 1, denoting the probability of the text being likely human or AI-generated. Each qualifying sentence within these segments inherits the segment's score. Since segments overlap, some sentences may have multiple scores, which are then pooled into a single score. These sentence scores are further aggregated and used to compute the overall document AI writing score."

The detector does not know your language background. It does not adjust its classification based on whether you are a native or non-native English speaker. It reads the statistical features of your word sequences and classifies them. If your English writing patterns, perhaps influenced by formal instruction that emphasizes certain academic phrases and structures, happen to overlap with patterns the model associates with AI-generated text, the score can be higher than you expect.

False Positives and Second-Language Writing

The FAQ lists text characteristics prone to false positives:

"Sometimes false positives (incorrectly flagging human-written text as AI-generated), can include content without a lot of structural variation, text that literally repeats itself, or text that has been paraphrased without developing new ideas."

The next sentence: "If our indicator shows a higher amount of AI writing in such text, we advise you to take that into consideration when looking at the percentage indicated."

Some of these characteristics can be more common in second-language academic writing. Non-native speakers often learn English through structured templates and repeated academic phrases, which can produce writing with less structural variation. Transitions like "moreover," "furthermore," and "in addition" are taught as standard connectors, and their repeated use can produce the kind of uniform structure the FAQ describes. This does not mean second-language writing is always flagged. It means the characteristics associated with false positives can overlap with patterns common in instructed second-language writing. Whether the answer is to write differently is its own question: should you change how you write to avoid being flagged.

You Cannot See Your Own Score

One more thing international students should know. The FAQ states: "only instructors and administrators are able to see the indicator." Students cannot check their own AI score. The FAQ continues: "However, with the PDF download feature, instructors can download and share the AI report with students." If you are concerned about your score, the only way to see it is to ask your instructor to share the PDF report with you. You cannot run the same check yourself through a student account. Once you have the PDF, how to read a Turnitin AI writing report is the next step.

What to Do If You Are Flagged

To summarize what we have covered:

Turnitin claims its training data included second-language learners to minimize bias, but provides no numbers to verify this claim.
The detector reads word probability patterns regardless of who wrote the text or what language they speak natively.
False-positive-prone characteristics, such as low structural variation, can overlap with patterns common in second-language academic writing.
Students cannot see their own AI score. Only instructors can, but they can share the PDF.
The AI score and the similarity score are independent measurements.

If you have a Turnitin report showing which passages were flagged, import the report and work on those specific passages. Eligible passages can be re-run at no charge.

Originally published at HumanPen: https://humanpen.net/blog/international-students-reduce-ai-rate?utm_source=devto&utm_medium=article&utm_campaign=syndication-2026-09

Systematic Reviews and AI Detection: Why Methodology Sections Get Flagged

Rayna Rabon — Sat, 05 Sep 2026 11:09:21 +0000

Systematic reviews follow rigid methodological templates. The methodology section repeats the same verbs and sentence structures across studies, which matches the false-positive patterns Turnitin describes. The detector is not measuring burstiness or perplexity. It is a word probability classifier. Here is what the documentation says about how the score is produced and how to interpret it.

Why Methodology Sections Match False-Positive Patterns

Systematic reviews demand a standardized methodology section. You describe the databases searched, the search strings, the inclusion and exclusion criteria, and the screening process. The language is necessarily repetitive because the PRISMA checklist requires specific reporting elements. Turnitin's documentation describes text types that produce false positives:

"Sometimes false positives (incorrectly flagging human-written text as AI-generated), can include content without a lot of structural variation, text that literally repeats itself, or text that has been paraphrased without developing new ideas."

The next sentence: "If our indicator shows a higher amount of AI writing in such text, we advise you to take that into consideration when looking at the percentage indicated."

A methodology section hits all three criteria, which is why Turnitin flagged my whole methodology section is such a common complaint. The structural variation is low because every systematic review reports the same steps. The phrasing repeats ("we searched," "we identified," "we screened"). The text is paraphrased from protocol descriptions without developing new ideas. The documentation says to take the percentage into consideration when the text matches these patterns, not to accept it as definitive.

How the Detector Computes the Score

The detection pipeline breaks qualifying text into segments and scores each one:

"When a paper is submitted to Turnitin, sentences from the submission are extracted and segmented into overlapping sections for prediction analysis. Each segment is classified by the AI detection model and given a value between 0 and 1, denoting the probability of the text being likely human or AI-generated. Each qualifying sentence within these segments inherits the segment's score. Since segments overlap, some sentences may have multiple scores, which are then pooled into a single score. These sentence scores are further aggregated and used to compute the overall document AI writing score."

In a systematic review, the methodology section's repetitive patterns could produce high segment scores across multiple overlapping segments. Those scores aggregate into a high document-level percentage. An 800-word methodology section could pull up the score for an entire 5000-word review. If you end up revising it, what you can rephrase and what must stay exact marks the lines you cannot move.

What the Detector Actually Measures

The detector does not evaluate burstiness, perplexity, or other named metrics, a claim examined in does Turnitin use perplexity and burstiness to detect AI:

"Our model is not explicitly programmed to evaluate specific signals such as 'burstiness,' 'perplexity,' or other individual metrics sometimes referenced in public discussions."

The next sentence: "Instead, it learns statistical patterns from our training data."

The same source explains what the model actually does:

"Our classifiers are trained to detect these differences in word probability and are adept at the particular word probability sequences of human writers."

The detector evaluates word probability sequences. It checks whether the words in a segment match patterns learned from human writing or from AI-generated text. In a methodology section, the word sequence is driven by the reporting template. Phrases like "we searched the following databases" and "studies were included if" appear across thousands of papers. These sequences may align more closely with statistical patterns from training data than with an individual writer's word choices. That alignment can produce a high probability score without AI involvement.

What Qualifies as Analyzed Text

The detector processes only qualifying text:

"This qualifying text includes only prose sentences, meaning that we only analyze blocks of text that are written in standard grammatical sentences and do not include other types of writing such as lists, bullet points (short non-sentence structures), or other non-sentence structures."

The next sentence: "This percentage is not necessarily the percentage of the entire submission."

Systematic reviews contain tables (PRISMA flow diagrams, extraction tables) and structured lists (inclusion and exclusion criteria). These are not qualifying prose. The percentage covers only prose sentences. If your results section is mostly tables, the percentage is weighted toward prose-heavy sections like methodology. The documentation also notes:

"The model does not reliably detect AI-generated text in the form of non-prose, or code, nor does it detect short-form/unconventional writing such as bullet points (short non-sentence structures)."

The next sentence: "This means that a document containing several different writing types would result in a disparity between the percentage and the highlights."

A systematic review is a multi-format document. The disparity between the percentage and the highlights is expected behavior.

Short Segments

Methodology sections are sometimes broken into short subsections. A search strategy subsection might be only 200 words. The documentation warns:

"In shorter documents where there are only a few hundred words, the prediction will be mostly 'all or nothing' because we're predicting on a single segment without the opportunity to overlap."

The next sentence: "This means that some text that is a mix of AI-generated and original content could be flagged as entirely AI-generated."

For a short subsection, the detector may predict on a single segment with no overlap to moderate the score. A human-written but template-following subsection could receive an all-or-nothing high score.

Turnitin's Boundary Improvement

In 2023, Turnitin also addressed false positives at document boundaries:

"Since launch, we have observed a higher incidence of false positive detection in the first few or last few sentences of a document. Many times these sentences consist of introduction or conclusion content written in a generic way. As a result, we have changed our detection logic to help reduce these false positives."

The next sentence: "We also worked on making our segment boundaries detection more precise which could lead in some rare cases to change of boundaries compared with a previous version."

This was a 2023 improvement. Systematic review introductions and conclusions follow templates (background, gap, objective, findings summary). The logic change addressed this pattern.

Do Not Treat the Score as Sole Evidence

Turnitin's documentation is explicit about the limitations:

"Our AI writing detection model may not always be accurate (it may misidentify human-written, AI-generated, and AI-paraphrased text), so it should not be used as the sole basis for adverse actions against a student."

The next sentence: "It takes further scrutiny and human judgment in conjunction with an organization's application of its specific academic policies to determine whether academic misconduct has occurred."

A systematic review with a high AI score on its methodology section is not proof of misconduct. The methodology section matches false-positive patterns. The detector is a word probability classifier. The score requires human judgment and institutional policy. What your instructor is told to do when your AI percentage is high is the guidance the marker is working from.

What This Means for You

To summarize:

Methodology sections match false-positive patterns: low structural variation, repetitive phrasing, and paraphrased text without new ideas.
The detector is a word probability classifier, not a burstiness or perplexity evaluator.
Only qualifying prose is analyzed. Tables, lists, and bullet points are excluded.
Short methodology subsections face an all-or-nothing scoring problem.
In 2023, Turnitin improved detection logic to reduce false positives at document boundaries.
The score should not be used as the sole basis for adverse actions. It requires human judgment.

If you receive a Turnitin AI report on a systematic review and want to address the flagged passages, import the report and work on them. Eligible passages can be re-run at no charge.

Originally published at HumanPen: https://humanpen.net/blog/ai-detection-systematic-review?utm_source=devto&utm_medium=article&utm_campaign=syndication-2026-09

How Does Turnitin Detect Collusion? The Due-Date Check Explained (With the Eric & Jane Example)

Rayna Rabon — Sat, 05 Sep 2026 10:05:12 +0000

Your similarity score was 18% before the deadline and 72% two days after — and the report now shows a match against a classmate's paper. Before you panic about being flagged, know what actually happened: Turnitin runs a final peer-to-peer comparison after the due date, so papers that never matched during the term suddenly surface against each other. Turnitin's official documentation describes this mechanism, the Eric & Jane scenario that shows why the copier's score can be the lower one, and the settings that decide whether the check even runs at your institution. This article explains the mechanism, the confusion it creates, and how to respond when your score changes overnight.

Short answer

Turnitin identifies collusion by comparing submissions to the same assignment against one another — a peer-to-peer check that, for most assignments, runs automatically after the due date. Turnitin's official student guide defines collusion as "identified when a student's work matches with another student's submission on the same assignment or to previously submitted papers," and its help center explains that this comparison "typically happens automatically only after the assignment due date has passed." A score that jumps after the deadline is often this check working as designed, not a new accusation against you.

What collusion detection actually is

The official student guide (Understanding the similarity score for students) opens with the definition:

"Collusion is typically identified when a student's work matches with another student's submission on the same assignment or to previously submitted papers."

Two things to notice. First, the match is against another student's submission — not against a website. Second, it also covers previously submitted papers (including your own earlier work from a different class, which shades into self-plagiarism; more below). The collusion check is not a separate AI-style detector; it is the similarity database comparing papers to papers.

The Eric & Jane scenario

Turnitin's official guide uses a pair of students to show how the mechanism works:

"Consider the following scenario: Eric acquired a copy of his classmate Jane's paper. Eric submits Jane's paper as his own and receives a similarity score of 25%. Jane, who originally wrote the paper, submits her work a few days later and receives a 100% similarity score."

The guide then explains the timing:

"Turnitin can identify that collusion has taken place in this scenario by running a final similarity check against all submitted assignments after the due date."

Read that carefully, because it inverts the usual assumption: Eric (the copier) scored 25%, and Jane (the author) scored 100%. If the check ran before the due date, Eric's paper only matched the internet — no one else had submitted yet. After the due date, the final check matched Jane's paper against everything, including Eric's copy, which is why her score jumps to 100%. The copier's low score is not evidence of innocence, and the author's high score is not evidence of guilt — the numbers reflect submission order, not authorship.

Why the check runs after the due date

Turnitin's help center explains the delay in a dedicated article, "Why are papers submitted to the same Turnitin assignment not matching to each other?":

"Papers submitted to the same Turnitin assignment do not match against each other immediately upon submission."

"To prevent students from matching against their own drafts, peer-to-peer comparison (known as collusion checking) typically happens automatically only after the assignment due date has passed."

"If your Standard (Classic) Assignment is set to allow resubmissions, all Similarity Reports are automatically regenerated within one hour after the assignment due date and time."

The due-date timing serves two purposes: every student gets the same level of scrutiny regardless of when they submitted, and students' own earlier drafts are excluded so that resubmission does not count against them. Before the deadline, reports only check external sources (internet, periodicals, repository) — classmates simply are not part of the comparison yet.

Why your score changed after the deadline

The same help-center page describes the visible consequence:

"Similarity Reports generated before the due date only check submissions against external sources... meaning student papers will not match against one another until the deadline passes."

So a report that was 18% on Monday and 72% on Thursday is most likely the collusion check regenerating the report — the match against a classmate's paper was always there, it just was not visible until the final run. This is the mechanism behind our guide on scores changing after the due date; this article is the mechanism, that one is the phenomenon.

Turnitin's blog on collusion adds the settings that make the check possible:

"The standard and institutional paper repository must be selected to compare student work against previous submissions. If 'no repository' is selected, a collusion check cannot take place."

If your institution did not enable a repository, the check cannot even run — another reason a "classmate match" may have other explanations.

Collusion vs self-plagiarism

Two different patterns both show up as "match with another student's paper," and Turnitin's documentation distinguishes them. The official blog (Effective solutions) uses Celine:

"Celine has submitted the same paper to multiple classes... She's checked her Similarity Report, which shows a match against another student in her institution... This is an example of self-plagiarism."

The instructor-side interpretation guide adds the reverse case:

"A student may have used Turnitin to submit drafts of the same paper, meaning their final draft has resulted in a score of 100%... the instructor can rectify this issue by excluding the student's previous submissions from the Similarity Report."

And the collusion blog notes the boundary:

"Turnitin automatically excludes papers submitted by the same student to the same assignment, preventing previous drafts from skewing similarity scores."

So: same student, same assignment → drafts excluded automatically; same student, different classes → self-plagiarism shows up as a match; different students, same work → collusion scenario. If your match is against a paper you yourself submitted elsewhere, that is not the collusion check — it is self-plagiarism, which requires a different conversation (and a different fix: the instructor can exclude your previous submission).

What a classmate match does and doesn't mean

Turnitin's blog is careful about what a match proves:

"A high similarity score does not always suggest that a piece of writing has been plagiarized, just as a low similarity score does not always indicate that no plagiarism or academic dishonesty, in the form of contract cheating, has occurred."

And its collusion-checking post, quoting Nicholls and Lewis (2017), warns against jumping to conclusions on circumstantial signals:

"Sometimes roommates using the same Wi-Fi router take the class together so this isn't necessarily a sign of collusion. However, if two or more students normally log in with different IP addresses, but suddenly are using the same IP address and starting exams at nearly the same time, it suggests [unauthorized] collaboration."

A shared Wi-Fi or similar phrasing is a hint for the investigation, not a verdict. The report flags or excludes matching text; a human (your instructor) decides what it means, and your response should be a conversation backed by your drafts and revision history — the same evidence playbook as our appeal and version-history guides. What you should not do is treat "matched a classmate" as a conviction you cannot respond to, or try to make your submission evade the check — altering your work to dodge the final comparison is the exact behavior the check exists to catch.

When the check runs in your assignment

The help center ties the timing to assignment configuration:

"If your Standard (Classic) Assignment is set to allow resubmissions, all Similarity Reports are automatically regenerated within one hour after the assignment due date and time... If your assignment is configured to generate reports immediately but treats the first submission as final, Turnitin will not automatically regenerate reports on the due date."

In the newer Standard Assignment, the check can also be forced early — the help center describes a Refresh Reports button that cross-compares everything on demand. The practical point for students: whether your report changes after the deadline is determined by how the assignment was configured, not by anything you did. If the score does move, the collusion check is the most likely reason.

Bottom line

Turnitin's collusion detection is a peer-to-peer comparison of submissions to the same assignment, and for most assignments it runs automatically after the due date. That is why the Eric & Jane scenario produces a 100% score for the original author and a 25% for the copier — the numbers follow submission order, not authorship. A score that jumps after the deadline is usually the final cross-check regenerating your report, and a match against a classmate is a starting point for a conversation, not a verdict. If the matched paper is your own from another class, that is self-plagiarism, which your instructor can exclude. And if the repository was disabled, the check cannot run at all. When a match appears, respond with evidence and a discussion — not with evasion.

Originally published at HumanPen: https://humanpen.net/blog/turnitin-collusion-check-explained?utm_source=devto&utm_medium=article&utm_campaign=syndication-2026-09