How to prove your work is original when an AI detector gives a false positive

No tool can prove you are the original author of a piece of text. What you can do is build an evidence chain: version history that shows the document being edited over time, source materials like notes and outlines and annotated sources, records of the writing process from supervisor emails to writing center visits, and an explanation for why your text looks the way it does. None of these, alone, disproves AI use. Together they give whoever is reviewing the score something to look at besides the score itself. And the company that made the tool has said, in three separate help documents, that the score should not be used as the sole basis for action against a student. That is not your argument. It is theirs.

The rest of this lays out the four layers of evidence, what each one does and does not show, and how to bring them to the conversation without overclaiming. There is also a section on what Turnitin itself has said about its own score, because those statements belong in the room and they are not yours to make.

The difference between this and a denial is the difference between "I didn't use AI" and "here is how this document came to exist." The first is a claim. The second is evidence. Only the second gives the reviewer something to work with.

AI writing detectors classify text. They take a document, segment it, assign each segment a probability, and aggregate those into a number. That number describes a statistical property of the prose. It does not describe who wrote it, how it was produced, or what happened before the file was submitted.

This is worth stating plainly because the question "how do I prove my work is original" often gets answered as if there were a symmetric tool to Turnitin, something that takes your draft and outputs "yes, you wrote this." There is no such tool. There is no classifier for originality. Originality is a claim about process and provenance, and process and provenance are things you establish with records, not with another score.

So the work is not to find a counter-detection. The work is to assemble the material that makes the writing process visible, and to pair it with the statements Turnitin itself has made about what its score does and does not settle. We went through the general version of what to do first in flagged but you wrote it yourself. This article is the wider framework.

Version history is the timestamped log of how a document changed over time. Word keeps one through OneDrive autosave and Track Changes. Google Docs keeps one natively and it is on by default. Overleaf keeps a Git-style page history. All three record when edits happened and roughly how large each revision was.

What version history can show:

• The document was edited across multiple sessions on different dates.
• Some revisions were small (a paragraph, a sentence) and some were larger.
• The file existed in earlier states, sometimes weeks before submission.
• Edits cluster around feedback events, supervisor comments, or your own revision passes.

What it cannot show:

• That every sentence was written by a person.
• That no AI tool was used at any stage.
• That the final text matches the intermediate drafts line for line.

The value of version history is not that it disproves AI use. It does not. The value is that it shows the document has a history consistent with being written over time. A pasted-in block has no history. A written paper has one. That distinction is something the AI score cannot see, because the score only looks at the final text.

If you wrote in Word or Google Docs, export the version history before the meeting. If you wrote in Overleaf, the page history serves the same purpose. Screenshot the list of revisions with dates visible, pick three or four versions along the timeline and note what changed in each, and if you used Track Changes keep a copy with the marks still visible. The marked-up version is the one that shows the editing process inline.

Before the draft existed, there were other things. Notes from the readings. An outline. Annotated PDFs. A reference list with your own markings on it. Search histories in the library database. These are the residue of research, not of generation. Their dates trace a path that does not look like a single text output.

What to bring:

• The earliest draft you still have, even if it was rough. A rough draft is evidence of a writing process. Its date is evidence the document existed before the final version.
• Notes, outlines, or annotated PDFs you produced while working. If you highlighted passages in sources and wrote marginal notes, those markings are evidence you engaged with the material before writing about it.
• Your reference list in its working state, before it was formatted. A reference list that grew over weeks as you found sources looks different from one that was generated in one pass.

None of this is a detector result. None of it produces a number. What it does is put the writing in a context that has a timeline, and that timeline runs in a direction that starts before the document and ends after it. A generated paper has no before.

The third layer is records of the writing process that exist outside the document itself.

• Email threads with your supervisor or anyone who gave feedback. If you sent a draft on a specific date and got comments back, that exchange anchors a point on the timeline. It also shows the document was being reviewed by another person during its production.
• Peer review records or writing center visits. If you took a draft to a writing tutor or exchanged drafts with a classmate, those sessions have dates and sometimes notes. They are evidence the writing went through a social process, which is how most academic writing actually gets done.
• Any earlier assignment that built toward this one. If this paper grew out of a proposal, an annotated bibliography, or a methods section you submitted separately, those earlier submissions are part of the same writing trajectory.

What this layer does is connect the document to other people and other dates. A document that sits alone with no connection to any prior interaction is easier to question. A document that sits inside a chain of emails, tutorials, and earlier submissions is harder to separate from the person who produced it.

The fourth layer is your ability to explain why your text looks the way it does. This is where Turnitin's own guidance on false positives becomes relevant.

The AI writing detection FAQs describe characteristics of text that tends to 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 is the one that matters in your context:

"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 that covers. Content without a lot of structural variation. Text that literally repeats itself. Text that has been paraphrased without developing new ideas. Those describe a large share of academic writing. Literature reviews summarize and synthesise, which is paraphrasing by nature. Methodology sections describe standard procedures in language that is supposed to be repetitive because the procedure is the same. Introductions and conclusions are often generic by design because they frame work that is specific in the middle.

If your flagged sections are in that territory, the company that made the tool is telling the person reading the score to discount it for exactly that reason. Not you. The company. Your job is to identify which of your flagged sections fit those characteristics and say so. "The methodology section describes a standard protocol, and the language is repetitive because the protocol is." That is not an excuse. It is a direct match to the characteristics Turnitin itself listed.

There is a second set of statements that belongs in the room, and none of them are yours.

The AI Writing Report guide says:

"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 sentence immediately after it is the one that gets dropped every time this quote is used:

"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."

Read both together. The first says do not use this number alone. The second says what should happen instead: further scrutiny, human judgment, the institution's own policies. That is the manufacturer telling the person holding the number to do more work before drawing a conclusion.

A separate page titled How should I review the AI Writing report? puts it differently:

"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."

The next sentence completes it:

"When educators look at the AI writing score and utilize it as a single data point rather than a definitive response, then it is being used as intended."

That is the manufacturer calling its own score a single data point. Not a verdict. Not a definitive response. The evidence chain you are building is what turns that single data point into multiple data points, which is the usage the manufacturer says is correct.

These statements appear in three separate Turnitin help documents. They are not a student's defense. They are the toolmaker's own boundary on what its number means.

Turnitin's FAQ states its accuracy 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 restates it:

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

Notice the qualifier. The under-1% target applies to documents with over 20% of AI writing. That is the band where the indicator shows a real number and highlights. Below 20%, the current display rule replaces the number with an asterisk. So the false positive guarantee covers the band where the tool commits to a figure, and does not extend to the asterisk band.

The same page explains the trade-off:

"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. We're comfortable with that since we do not want to incorrectly highlight human-written text as AI-written. For example, if we identify that 50% of a document is likely written by an AI tool, it could contain as much as 65% AI writing."

Read the direction of that. The model is tuned to avoid false positives, and the cost of that tuning is under-reporting. A high score is the model committing to a claim despite a tuning that biases it toward caution. A low score is consistent with both "nothing happened" and "something happened that the model did not catch." Neither direction settles the question of who wrote the document. We went through what 20% means and where the threshold sits in is 20% AI too high.

Before you can explain anything about the score, you need to see the report. Turnitin's FAQ states:

"Please note, only instructors and administrators are able to see the indicator."

The same page also says:

"The AI writing detection indicator and report are not visible to students."

That sounds absolute. The next sentence changes it:

"However, with the PDF download feature, instructors can download and share the AI report with students."

So if nobody has given you the report, ask for the PDF. Not a screenshot of the percentage. The full report. Because the full report shows you where the highlights sit, how many words were analysed, whether the score is a real number or an asterisk, and when the report was generated. All of that is information you need before you can explain anything.

A screenshot of a number tells you one thing. A PDF of the report tells you whether the number is a number at all. The distinction matters more than people assume, and we went through the four indicator states in how to read a Turnitin AI writing report.

You do not need a slide deck. You need one page with a timeline on it and a short list that pairs each flagged section with an explanation.

1. Draw a horizontal line. Mark the date of the earliest draft on the left, the submission date on the right.
2. Put the version history dates on the line. Not every version. Five or six. The ones where something changed.
3. Put the supervisor emails, peer review sessions, and writing center visits on the same line. These anchor the version history to events that happened outside the file.
4. Put the Turnitin report generation date on the line. This is the date the score was produced, not the date you submitted. They are sometimes different.
5. Under the line, list each highlighted section from the AI report and your explanation. "Methodology section, standard protocol, repetitive by design." "Literature review, synthesising three sources, structural variation is low because that is what synthesis looks like." These map directly to the characteristics Turnitin itself listed under false positive tendencies.
6. Bring the three Turnitin quotes. Printed, with URLs. The sole-basis statement from G1 and its next sentence. The single-data-point statement from G5 and its next sentence. The false-positive characteristics from G2 and its next sentence. These are not your words. They are the toolmaker's.

That page does two things. It gives the reviewer something to look at besides the score, and it shows you came to participate in a review, not to make a denial.

Do not claim the evidence chain disproves AI use. It does not. Version history shows the document was edited over time. Source material shows research happened. Process records show other people were involved. None of these is a counter-detection result. Claiming more than they show hands the other side an easy rebuttal and weakens everything else you brought.

Do not bring a tool's score as your defense. Running your paper through another AI detector and getting "0% AI" proves nothing, and the company that made Turnitin has explicitly said no score should be the sole basis. Bringing a different score to fight a score is using the same logic the manufacturer told the reviewer not to use.

Do not overclaim what the Turnitin quotes mean. The sole-basis statement says the score should not be the only basis. It does not say the score is wrong. The single-data-point statement says the score is one input. It does not say the score is meaningless. Present them as what they are: the manufacturer's own boundary on how its number should be used. Let the reviewer draw the conclusion.

HumanPen is a document rewriter. We are not going to tell you that running your paper through a rewriter will make an AI score disappear, because we do not make claims about future scores. What we can say is what the tool does: you upload the document with its AI writing report, and only the passages the report flagged are rewritten. The rest of the document stays verbatim. You review the output before it becomes anything final.

If you want to go into the meeting with the original document and let the evidence chain speak, that is one path. If you want to revise the flagged sections beforehand, that is another. Both are legitimate. The meeting comes first.

Can I prove I didn't use AI? No. No tool can prove a negative about your writing process. What you can do is assemble an evidence chain that makes "I wrote this" the simpler explanation. Version history, source material, process records, and an explanation for why your text looks the way it does. Together they give the reviewer something to look at besides the score.

What if I don't have version history? You lose the strongest single piece of timeline evidence, but you do not lose everything. File modification dates on earlier drafts, email attachments sent to your supervisor, and cloud-synced backup folders all carry timestamps. They are not as clean as a version history pane, but they tell the same story.

My professor only sent me a screenshot of the percentage. What do I ask for? The full AI writing report as a PDF. Turnitin's FAQ says students cannot see the indicator, but also says instructors can download and share the report through the PDF download feature. Ask for the PDF, not the screenshot. The report shows you where the highlights sit, how many words were analysed, and whether the score is a real number or an asterisk.

The flagged sections are all in my literature review. Does that matter? Yes. Turnitin's FAQ lists "content without a lot of structural variation" and "text that has been paraphrased without developing new ideas" as characteristics associated with false positives. Literature reviews synthesise and summarise, which fits both. The next sentence in the FAQ says: "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." That is the company telling the reviewer to discount the score for exactly that reason.

Can I bring quotes from Turnitin's own help pages to the meeting? Yes. The statements about the score not being the sole basis, about it being a single data point, and about false positive characteristics are published in Turnitin's help documents. They are the toolmaker's words, not yours. Bringing them printed with URLs is more effective than paraphrasing.