Capstone Projects and AI Detection: Staged Submissions, Shared Documents

Three things make a capstone different from a coursework essay, and each one has its own consequence. It is submitted more than once, so your own interim report can end up in a repository and become something your final report is compared against. It is written by several people, so the AI writing percentage covers a document nobody wrote alone and carries no author column. And the person who merges everyone's sections into one voice at the end is doing the one editing pass that Turnitin's own documentation names as false-positive-prone.

None of the three is solved by rewriting the file at the end of the year. Two of them are decided by a setting somebody else chose, months earlier, and the useful action is a question rather than an edit.

Nothing here predicts a score. Turnitin publishes the shape of its pipeline and a set of file requirements. It does not publish thresholds, and neither do we.

Turnitin has a feature built for exactly the shape a capstone has. It is called a Multipart assignment, and the description on the guide reads like a project timetable:

"Multipart assignments let instructors connect two or more assignment parts into one larger assignment workflow. Use Multipart assignments when students need to complete a larger project in stages, such as an outline, annotated bibliography, first draft, and final draft."

The sentence that matters is the next one:

"A Multipart assignment is a connected set of assignment parts. Each part has its own title, instructions, dates, settings, and submissions, but the parts are grouped together as one assignment experience."

Its own settings. Its own submissions. And from your side of the screen, Turnitin says you will not necessarily notice the grouping at all: "Students see each part of a Multipart assignment as a separate assignment in their LMS or Turnitin, as they normally would." The student guide repeats it. Your parts "appear in your assignment list as separate assignments", and "each part of a Multipart assignment has its own submission".

So when you upload the interim report in week seven and the final report in week twenty-two, that is two submissions with two sets of settings, not one file being revised. The protection people rely on, where a resubmission overwrites the previous one and the two do not match against each other, is a rule about resubmitting to the same assignment. We went through exactly when it applies and when it does not in will my own earlier draft match my resubmission, and the short version is that the case that hurts is the second assignment, not the second upload.

Whether your week-seven file is sitting in a repository at all is a setting, and it is set two levels above you. Turnitin's account settings guide lists what an administrator can hand to instructors:

"Enable instructor standard repository options: Chosen instructors will be able to set the assignment option to either store student papers within the standard paper repository or not store the papers in any repository."

With expanded options on, the instructor chooses between the standard repository and the institutional one instead. And an administrator can take the choice away entirely by selecting "Submit all papers to the standard repository", which the same page describes as: "All student papers submitted to the account will be stored in the standard paper repository."

So the submission points on one module do not have to follow the same rule, and not one of these settings is visible from the upload screen.

That gives you one useful thing to do, and its timing is the whole point: ask before the interim submission, not after the final one. A single email to your supervisor asking whether the interim part stores submissions to the repository takes a minute in week six and is unanswerable in week twenty-three.

One more thing about the Multipart structure, since it is easy to assume your institution has it. Turnitin lists integrations where it is not available: Google Classroom, Sakai, Manaba, In Campus. If your capstone runs through one of those, the parts are just separate assignments with nothing connecting them, which changes nothing about the repository question and everything about whether your instructor can flip between your submissions while marking.

Here is the failure mode we would rather you skip. A group sees a high similarity figure on the final report, works out that their own literature review from the interim submission is the source, and then spends a week rewriting that literature review so it stops matching.

Two problems with that.

First, the similarity score and the AI percentage are not the same measurement and do not affect each other. Turnitin states it plainly: "The Similarity score and the AI writing detection percentage are completely independent and do not influence each other." A self-match is a similarity event. It tells you nothing about an AI number, and fixing it does not move one. AI Writing Report vs Similarity Report works through why all four combinations of high and low are possible.

Second, and this is the part that bites, rewriting your own literature review to make it stop matching itself pushes the text toward the profile Turnitin names when it describes 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."

The sentence immediately after that one is the sentence to keep:

"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 literature review rewritten purely so it stops resembling its earlier self, with no new reading behind the rewrite, sits close to one specific item in that sentence: text paraphrased without new ideas developing behind it. You would be trading a similarity number that has an instructor-side fix for a text profile that does not.

The instructor-side fix is the actual answer here. Excluding your previous submissions from the report is a control your instructor holds and you do not.

The mechanism explains why a group document behaves the way it does. Turnitin's account of it runs roughly like this. The submission is broken into overlapping sections. Every section gets a probability between zero and one. A qualifying sentence takes its score from whichever sections cover it, several scores get pooled into one, and those are aggregated upward until a single document figure comes out.

Read that as a pipeline and one thing is obvious. Nowhere in it is there a field for who typed the segment. The document is the unit. Five authors go in and one percentage comes out, and the percentage cannot be divided back among them. The general version of this problem, for people who are already in the middle of it, is group project flagged for AI when you didn't use AI.

But there is one thing in the report that does have a dimension the score lacks. The highlights have positions. A highlighted passage sits on a page, in a section, under a heading, and your team knows who wrote that heading.

That only helps if two conditions hold. You need the report, which you cannot get yourself, because Turnitin says "The AI writing detection indicator and report are not visible to students" and then says that "with the PDF download feature, instructors can download and share the AI report with students". So the ask is for the PDF, not a screenshot of a number. How to read a Turnitin AI writing report covers what you are looking at once you have it.

And you need a record of who wrote what, made at the time. A section-ownership table with dates, written in week two and updated as sections move around, is dull and takes ten minutes a term. Assembled in week twenty-three under pressure, the same table is a document created after the allegation, and everyone in the room knows it.

Turnitin only analyses what it calls qualifying text, which it defines as prose sentences written in standard grammatical sentences, inside a long-form piece of writing. It also states that the percentage "is not necessarily the percentage of the entire submission". The full version of that definition, with the exclusions, is in what is qualifying text in Turnitin.

Now think about what a capstone report is actually made of. Requirements tables. Numbered test procedures. Code listings. Figure captions. A Gantt chart. Interview schedules in an appendix. Very little of that is paragraphs of sentences.

Which leaves the background, the literature review, the discussion and the conclusion carrying nearly the whole measurement.

Here is why that is a group problem and not a formatting problem. In most teams those four sections are not distributed evenly. They land on whoever writes fluently in English, or whoever volunteered for the "writing bits" while others built the artefact. So the document-level percentage is computed almost entirely over text written by one or two people, and then reported as a property of a document with five names on it.

Worth knowing before the meeting, because the first question in that room is usually "who wrote this", and the honest answer for the parts that produced the number may be shorter than the author list.

Every group document has one. The person who takes five files the week before the deadline and makes them read like one document.

Some of what they do is necessary and carries no risk at all. Making the defined terms consistent so a "participant" in section three is not a "respondent" in section five. Fixing the tense in the methods. Renumbering figures after someone inserted one. Putting every citation into one style. Aligning heading levels. None of that touches the shape of anyone's sentences.

The part to be careful about is the pass where somebody smooths the whole document into a single voice because the joins are visible.

Go back to the list Turnitin gives for false positives. Content without a lot of structural variation is the first item. A document written by five people has structural variation in it by construction, and it has that variation because five people wrote it. A voice-smoothing pass is a pass that removes it.

So the practical version: harmonise the terminology, the formatting and the references, and leave the sentence rhythms alone. A capstone report is allowed to sound like it was written by a team, because it was. The joins are not a defect that a marker will punish, and the effort to hide them is spent on the one dimension you would rather keep.

If the document lives in a shared editor, there is a second reason to hold back. Whatever record that editor keeps of who wrote which section stops being useful the moment one person edits every paragraph. Our tool-by-tool walkthrough of what those records actually retain is in how to keep version history in Word, Google Docs and Overleaf.

Capstones with an external partner, a company, a hospital, a council, add a reader who sits outside every system described above. That produces three specific things worth sorting out early.

Nobody in that conversation has the report. Students cannot see the AI writing indicator. Neither can anyone outside the institution. If a partner asks whether the report was AI-checked and what it said, the only route to an answer is your instructor downloading the PDF. Ask through the university, not around it.

The file you hand the partner may not be the file you submit. Deliverable versions usually drop the reflective sections and the marking-scheme appendices, or add material the university never sees. Two files, two word counts. That matters at one specific boundary: Turnitin does not generate an AI writing report for a submission over 30,000 words, and a compiled capstone report with appendices lands near that line more often than people expect. A submission outside the file requirements does not come back clean, it comes back unprocessed, and we went through the difference in can Turnitin check a whole thesis.

If the project material is confidential, the repository question stops being administrative. Whether your submission is stored is the setting quoted earlier, and it can be set so that instructors choose per assignment, or so that every paper in the account is stored regardless. Removal afterwards is not something you do. Turnitin describes it as a request that goes upward: "instructors have the ability to request the deletion of any submissions in their assignments. Administrators can approve or reject requests." So if the partner's data is in your appendices, that is a conversation with your supervisor before the first upload, not a form after the last one.

We cannot tell you what your partner organisation will ask for or accept. Nobody can, and anyone who publishes a confident answer to that is guessing on your behalf. What is knowable is the university side, and the university side is a set of switches with names.

Every quotation above comes from Turnitin's own help pages, so it holds for institutions that use Turnitin and for nobody else. If yours runs something different, the part that transfers is not the vendor's vocabulary. It is three questions, and they fit in one email to a supervisor or a programme administrator:

• Is each submission point on this module a separate deposit, or are they versions of one thing?
• Who decides whether a submission is stored, and at what level is that decision made?
• If something has to be withdrawn afterwards, what is the process and who signs it off?

Every system has an answer to those three under some other set of names. What goes wrong is almost never that you did not know what a feature was called. It is that you assumed several submissions were one event.

All of this is cheap at the start of a capstone and impossible at the end.

• Ask which parts store submissions. One email to the supervisor covering every submission point in the module, asked before the first one.
• Keep a section-ownership table. Section, owner, date started, date last substantially changed. In the shared drive, updated as you go, not reconstructed later.
• Agree who owns the shared document. In most collaborative editors, the person who created the file has powers over its history that the others do not.
• Name a version at each milestone. Whatever your editor calls it, a labelled version with a date and a two-word note is worth more than a hundred automatic ones.
• Decide the harmonising rules before the last week. Terminology list, citation style, figure numbering, heading levels. Write them down in week two and nobody has to smooth anything in week twenty-two.
• Find out whether your final submission is over 30,000 words early enough to split it, rather than discovering an unprocessed report the night before.

If you are already past the point where this is preventative advice, flagged, but you wrote it yourself covers assembling a response, and if your module uses Turnitin's browser writing space rather than file uploads, what that space records is a separate subject entirely, covered in what Turnitin Clarity records.

A rewrite does not answer any of the three problems above. It cannot change which assignment stored your interim report, it cannot add an author column to a percentage, and it is not a substitute for the record of how the document was built.

Where it does have a place is narrow and specific: you have the PDF report, some passages are highlighted, and you want to revise those passages without touching the rest of a document that four other people wrote and checked. That last part is the constraint that makes group documents different. Anything a tool rewrites has to be checked by a human before it goes back in, and on a capstone that human is often not the person who ran the tool.

That constraint is what HumanPen is shaped around. You upload the document together with the report, the highlights in the report set the scope, and anything they did not cover comes back exactly as it was written. A paragraph is the smallest unit it will touch, so a partial match is expanded to the whole paragraph and shown to you before anything runs. Billing follows the words actually rewritten rather than the size of the file. Eligible results can continue lowering AI for free.

That is a statement about scope and cost. It is not a statement about what your next report will say, and we do not make those. On a document with five authors, the useful property is the short list of paragraphs your team has to re-read before it goes back in.

Does the AI percentage tell my supervisor which of us used AI? No. Turnitin's own description of the mechanism has no author field in it: the submission is segmented, each segment is scored, and the results are pooled and aggregated into a document-level figure. The highlights have positions in the document, which is the only thing in the report that can be connected to a person, and only if your team already knows who wrote what.

My interim report and my final report are both in the system. Is that a problem? It is a similarity question, not an AI question, and the answer depends on whether the interim submission was stored. That is an assignment setting your instructor controls, chosen from options your administrator enabled. Ask before the interim submission goes in, and if a self-match has already happened, take it to your instructor, since excluding previous submissions is a control on their side of the report.

Our final report is 45,000 words with appendices. What happens? Turnitin's file requirements state that a submission must not exceed 30,000 words for an AI writing report to be generated. Over the ceiling there is no percentage, and the indicator shows that it could not process the submission rather than showing a clean result.

One person edited the whole document at the end to make it read consistently. Was that a mistake? Not automatically, and not for the parts that were about terminology, references and formatting. The part worth being careful about is smoothing everyone's sentences into one voice, because "content without a lot of structural variation" is the first item on Turnitin's own list of what can produce false positives, and a document written by five people starts out with that variation for free.

Can I show a version history instead of arguing about the score? You can show it, and it is a reasonable thing to bring, but it settles nothing on its own. Turnitin says its own score "should not be used as the sole basis for adverse actions against a student" and describes it elsewhere as "a single data point rather than a definitive response". Your version history is another data point in the same sense. What weight either carries is your institution's decision, set out in its academic integrity procedure, and that document is the one to read.