Why Turnitin Says Your Work Is 100% AI

Three situations produce a 100 much more often than "every sentence you wrote reads like a machine wrote it", and you can rule them in or out in about ten minutes. Check which number you are holding: Turnitin says the AI writing indicator is not visible to students, so a percentage you found yourself, inside your own view, is the similarity score rather than the AI one. Check how much prose the file contains: Turnitin says that in documents of only a few hundred words the prediction is "mostly 'all or nothing'", because it is predicting on a single segment with no opportunity to overlap. Check what the 100 is 100% of: the percentage is computed over qualifying text, and Turnitin says that figure "is not necessarily the percentage of the entire submission". Only when all three are ruled out does it make sense to ask what the model saw in the writing itself.

The three call for completely different responses, which is why the order matters more than the explanation. This page is about how an extreme number gets produced. It is not a script for a misconduct meeting, and nothing here predicts what a future report will say.

Turnitin is explicit about who can see the AI number. Its guidance says "only instructors and administrators are able to see the indicator", and separately that "The AI writing detection indicator and report are not visible to students." The sentence immediately after the second one matters just as much: "However, with the PDF download feature, instructors can download and share the AI report with students." So both routes exist. A PDF someone sent you can genuinely be the AI report. A percentage you clicked into yourself, in your own submission view, is not the AI indicator.

That leaves the other number, and 100 is a much more ordinary value there. A similarity score of 100% has a documented and completely innocent route: your own earlier drafts, stored by a different assignment, matching your final version almost word for word. We traced how that happens, and when it does not, in will my earlier draft match my resubmission. The two reports are separate analyses that can disagree in all four directions, which is the subject of AI report vs similarity report.

A 100 on the similarity side and a 100 on the AI side are different findings with different causes and different responses. Spending an evening on the wrong one is the most common way this goes badly.

Turnitin's published description of the calculation has three moves in it: sentences are pulled out and grouped into overlapping segments, each segment gets a probability between 0 and 1, and every qualifying sentence inherits the score of each segment it sits in. Because the segments overlap, a sentence near a boundary carries more than one score, which are then pooled. Those pooled sentence scores are what get aggregated into the document figure. The averaging is the whole reason a document can come back at 43 rather than at one of the ends.

Now take that away. Turnitin's FAQ says what happens when there is nothing to average across:

"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."

Read that as a statement about resolution rather than about accuracy. In a long document, a 100 would mean a large number of separate predictions all landed on the same side. In a document of a few hundred words it can mean one prediction, which the interface then renders as a percentage. The two are displayed identically and are not the same kind of evidence, and the second sentence of that quote says outright that mixed content is what gets swept up.

There is a floor underneath this. A submission needs at least 300 words of prose in a long-form format before an AI report is generated at all. So the shortest file that can receive a number is also the one whose number rests on the fewest independent observations, and the band just above the floor is where those two facts overlap. A 900-word reflective piece, a short problem set write-up, a conference abstract padded to length: all of them sit in the zone Turnitin describes. What happens at the other end of the range, where a thesis exceeds what one report covers, we went through in can Turnitin check a whole thesis.

The percentage is not calculated over your document. It is calculated over what Turnitin calls qualifying text, and the definition is narrow: "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 sentence that follows is the one to hold on to: "This percentage is not necessarily the percentage of the entire submission."

Some exclusions are stated separately. Turnitin's release notes record that a bug highlighting AI writing inside bibliographies was fixed and that "Bibliographies are now excluded when processing the AI writing report", and elsewhere that "We are now able to process long-form prose text in tables." Both entries end the same way, with an instruction not to assume the change applies to what you already submitted: resubmit to reprocess. So a reference list is outside the calculation, and paragraphs sitting inside a table are inside it, which is not the intuition most people have about either.

Put a real file through that. A 5,000-word lab report with a methods table, three bulleted protocol lists, an appendix and 60 references might contain 2,000 words that qualify. A 100 on that report is a statement about those 2,000 words. It is not a statement about the other 3,000, which were never assessed and cannot be defended or attacked on the basis of the number.

This is also why a 100 can coexist with pages that carry no highlighting at all. Turnitin says a document containing several different writing types "would result in a disparity between the percentage and the highlights". At the top of the range that disparity looks like a contradiction, and it is expected behaviour.

Counting highlighted lines and comparing them against the percentage will not resolve it, because the two are computed over different things. How to read a Turnitin AI writing report works through the rest of the report states for the same reason.

If the file is 8,000 words of continuous prose, the single-segment explanation is gone. Many separate predictions did land on the same side, and the question becomes what would make an entire document look the same way to the model all the way through. Turnitin publishes a partial answer, in the form of the properties it says show up in its own 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. 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."

Those three properties are all properties of a whole document rather than of a sentence. That is our reading and not something Turnitin claims, but it is the part that fits the shape of the question: a paper written to a mandated section order, in a discipline with a fixed way of phrasing a methods paragraph, restating its own aims in the abstract and again in the introduction and again in the discussion, has roughly the same character in every segment. There is no passage that differs enough from its neighbours to pull the aggregate down. Uniformity is not a writing fault, and in several genres it is the requirement, which is exactly why the result is hard to argue with by pointing at any one paragraph.

The second sentence of that quote is a vendor instruction to whoever is reading the score, not a defence you construct. What has actually moved decisions is a separate subject, and we wrote it up from published casework in flagged, but you wrote it yourself. If editing is appropriate and permitted, the properties in that list are also the closest thing to a specification for doing it by hand, which is the approach in how to humanize AI text without a tool.

One thing worth knowing before you build any argument on the number being exaggerated: the published tuning runs the other way. Turnitin says that "In order to maintain this low rate of 1% for false positives, there is a chance that we might miss some AI written text in a document", and gives the direction of the error: "if we identify that 50% of a document is likely written by an AI tool, it could contain as much as 65% AI writing." Whatever else is going on, "the tool is built to over-report" is not the design the vendor describes.

That does not make a 100 a finding. Turnitin states in three separate help pages that its model may misidentify human-written, AI-generated and AI-paraphrased text and should not be the sole basis for adverse action against a student. And the accuracy claim people quote at each other carries a qualifier that gets dropped in transit, which we took apart in what a 1% false positive rate means. There is also no published threshold at which a number becomes misconduct, in either direction: is 20% AI too high covers why the 20 in the interface is a display rule rather than a rule about you.

If you wrote it, rewriting it is the wrong first move. It changes the object under discussion, and it answers a question nobody asked.

Worth saying plainly, because it cuts against us: at 100, the thing HumanPen is normally good for has nothing to bite on. Upload the document together with the Turnitin or iThenticate report and only passages matched from the report are rewritten, with the rest preserved verbatim, and billing counts only the words actually rewritten. That is a real saving at 30%, where two thirds of the file never gets touched or charged for. At 100% the report has marked everything, so the scope is the document and the cost is a full pass. If you were hoping the report would narrow the job, a 100 is the one number where it does not.

If a new report still flags passages, eligible results can continue lowering AI for free.

None of that is a promise about a detection result, and we will not make one. No tool can tell you what a future model version will output on a document it has not seen, and anything offering a guaranteed score is describing something it cannot control. There are also cases where the tool is the wrong instrument entirely: if your position is that the work is yours, the number is not the object to be fixed.

Does 100% mean every sentence I wrote was flagged? No. The percentage is computed over qualifying text, which excludes lists, bullet points and other non-sentence structures, and Turnitin says the figure is not necessarily the percentage of the entire submission. A file with substantial tables, lists and a reference section can show 100% while large parts of it were never assessed.

My essay is only 600 words. Is that the explanation? It is the first thing to check. Turnitin's own account is that in documents of a few hundred words the prediction is mostly "all or nothing" because it is working from a single segment with no opportunity to overlap, and that a mix of AI-generated and original content can therefore be flagged as entirely AI-generated.

I saw the number in my own Turnitin view. Is that the AI score? Almost certainly not. Turnitin says only instructors and administrators can see the AI indicator and that it is not visible to students, though instructors can download the AI report as a PDF and share it. A percentage you reached yourself is the similarity score, which has its own ordinary routes to 100.

Can any tool guarantee the next report will be lower? No, and treat the claim as a warning sign. Detector versions change, and nobody outside the vendor can commit to the output of a model version that has not run yet. What a revision service can define is scope and cost, not a score.