How to Prove You Didn't Use AI to Write Something

A solid paper trail serves as the best safeguard. When your version history displays each draft, each edit, and each rewrite in real time, it leaves no room for doubt. No one can claim the work is unoriginal when the evolution is documented step by step.

That distinction matters because of how these cases actually get resolved. The University of South Carolina's own Office of Student Conduct and Academic Integrity describes exactly what it looks for when a professor flags a paper: "We review the version history with the student and look for patterns in the creation of the document that may give insight into the student's process." Not a rerun through a different detector. Not an argument about writing style. The document's own edit history.

Why version history beats a detector rebuttal

A detector score is a probability estimate about statistical patterns in finished text. It can't see how the text was written, only what it looks like once it's done. Version history is the opposite: it's a direct record of the writing happening, timestamp by timestamp.

That's exactly how two UC Davis students cleared themselves in a case Rolling Stone covered in 2023. William Quarterman was flagged by GPTZero on an exam answer. Louise Stivers was flagged by Turnitin's newly launched AI-detection tool on a case brief. Both were referred to their university's academic-integrity office. Both were cleared, and in Stivers' case specifically because she handed the reviewing office, in her words, "step-by-step instructions on how to open Google Docs and review history." The timestamps showed the paper being written incrementally, not pasted in all at once.

That's the pattern a detector score can never show you: whether a document was built or dropped in. We've covered why detectors get this kind of judgment wrong on their own — non-native English writers and heavily-edited drafts both trigger false positives at real rates. Version history sidesteps the guesswork entirely by showing the actual process instead of a probability about the output.

The institutional pattern behind this shift

This isn't just individual professors changing their minds. Universities have started pulling back from treating a detector score as sufficient proof on its own, and they're doing it in writing.

Vanderbilt disabled Turnitin's AI-detection feature entirely in 2023, laying out the math in its own guidance memo: at a 1% false-positive rate on the roughly 75,000 papers the university submitted that year, that meant potentially 750 students wrongly flagged in a single academic year if the tool had been active. Washington State University went further in February 2026, canceling its Turnitin AI-detection contract outright. The Office of the Provost's memo to instructors was direct: "Therefore, suspicion of the use of AI is not sufficient for a finding of student responsibility for inappropriate use of AI." The student paper The Daily Evergreen confirmed the same quote from WSU's academic integrity office in its own reporting.

And in January 2026, a New York state court put a legal weight behind that same principle. Adelphi University freshman Orion Newby, who has autism spectrum disorder and worked with a support-program tutor on a history essay, was found responsible for academic misconduct after Turnitin returned a 100% AI-generated score — despite two other detectors returning 0% on the same paper. Justice Randy Sue Marber ruled the university's finding "without valid basis and devoid of reason" and ordered it expunged, according to court records and CBS News coverage of the case. We go deeper into what that ruling means for how schools should weigh a score in Why Does My Essay Get Flagged as AI? — this piece is about preventing that fight from starting in the first place.

The through-line across all three: a percentage by itself is treated as weak evidence now, in policy and in court. Process evidence is treated as strong evidence. That's the gap this piece is about closing, before you're the one explaining yourself to a review board.

Build the paper trail before you need it

Attempting to overhaul the writing record after the paper is questioned is futile. What you need is a procedural adjustment. This adjustment costs nothing beyond the normal automatic logging of your writing.

• Write in a tool with built-in version history. Google Docs and Microsoft Word (with Track Changes and version history enabled) both log edits automatically as you type — no extra software required, no extra step to remember. Writing in a plain text editor or pasting a finished draft into a submission box from somewhere else throws that record away.
• Draft in stages, not one sitting. A history showing outline, then a rough pass, then revisions over hours or days looks like writing. A history showing one enormous block of text appearing in a single edit looks exactly like a paste, whether or not it was one. This is the specific pattern Draftback — a Chrome extension with more than 500,000 users that plays a Google Doc's edit history back like a recording — is built to surface for reviewers, and it's the same pattern USC's own conduct office says it looks for.
• Keep your outlines, notes, and earlier drafts, not just the final file. Newby's case leaned partly on evidence that he'd worked with a named tutor across many hours. Dated notes and a visible outline make that kind of claim concrete instead of just asserted after the fact.
• If you're ever accused, ask for specifics in writing before you respond. Which tool flagged the paper, its exact score, and whether there's any evidence beyond the detector output. Most conduct offices, public or private, owe you that under their own published process.
• Check your own draft before you submit it, if you're worried about how it reads. TheChecker.AI's free demo gives you a sentence-level breakdown, not just one aggregate number, so you can see specifically which passages read as high-perplexity or formulaic before a professor's tool ever does. Catching an oddly uniform paragraph yourself, and then pointing to your own draft history if it's ever questioned, beats finding out about both after the fact.

No presumption of ill intent on the part of the instructor or the school is needed. What matters is considering the formulation of the work as something that can be demonstrated. Think of it as retaining a receipt for a purchase you may later wish to return.

FAQ

Does Google Docs version history expire or get deleted? Google Docs keeps version history indefinitely for a document unless you or a collaborator manually deletes named versions, and even unnamed automatic versions are typically retained. It's tied to the document, not to a time window, so a history from months earlier is generally still there if you need to pull it up later.

What if I wrote my draft on paper first, or used a tool without version history? Dated handwritten notes, an outline with visible revisions, or screenshots taken at different stages all serve the same function as an automatic version history — they show a process unfolding over time rather than a single moment. The USC conduct office's own process explicitly asks students to walk through how and when they worked, not just to produce a single artifact.

Is Draftback itself an AI detector? No. Draftback plays back a document's existing Google Docs edit history for a reviewer to interpret directly; it doesn't generate an AI-probability score of its own. That's a meaningful difference from a detector score, and it's part of why version history and a detection tool answer different questions rather than competing to answer the same one.

Does having a clean version history guarantee I won't be accused? No single piece of evidence guarantees anything, and a version history that shows minimal editing over a very short window can still raise questions on its own. But paired with your outlines and drafts, it's the evidence every documented case above turned on, and it's something you control before any accusation exists rather than something you have to reconstruct after one.