1. Four things not to do
Most of the damage in these cases happens in the first day, before anyone has looked at the evidence. These are the four mistakes that make a recoverable situation much worse.
Do not admit to something you did not do
You may be offered a deal: admit it, take a zero on the assignment, and the matter ends quietly. When you did not do it, that is not a shortcut, it is a permanent record of academic dishonesty that follows you into graduate applications and some job screenings. If you did not use AI, say so plainly and keep saying so.
Do not edit, rewrite, or delete the file
The instinct is to go make the flagged sentences sound more like you. Resist it. Your edit history is the single most persuasive evidence you have, and modifying the document now both damages it and looks like tampering. Duplicate the file and work only on the duplicate.
Do not open with "AI detectors are inaccurate"
It is true, and there is real research behind it (section 5). But led with, it sounds like the argument of someone who has nothing else. Lead with your evidence. Bring the research as the second move, once you have shown your work.
Do not walk in without your timeline
An unprepared meeting turns into your word against a number on a screen, and the number sounds more objective than it is. Ten minutes of preparation changes the shape of the entire conversation.
2. Your strongest evidence is version history
Here is the thing most students do not realize: you probably already have a near complete recording of yourself writing the paper. Most writing software saves incremental versions automatically. That record shows text appearing gradually over hours or days, with false starts, deletions, and revisions.
Pasted AI output does not look like that. It appears in large blocks, at once. A reasonable reviewer looking at a genuine edit history can see the difference immediately, and it is far more convincing than any score from any tool.
Before you do anything else, preserve it. Make a copy of the document. Export or screenshot the version history with dates and times visible. Save those screenshots somewhere outside the document itself.
Google Docs
- Open the document, then File → Version history → See version history.
- The right panel lists every saved revision with timestamps. Expand the entries to show individual saves.
- Screenshot the panel with the timestamps readable. Capture several, including the earliest sessions.
- For a stronger record, the Draftback Chrome extension replays your document being typed, keystroke by keystroke, and produces a graph of writing sessions over time. It is the most compelling single artifact available to a student, and it only works on documents you own.
Microsoft Word (365, OneDrive, or SharePoint)
- File → Info → Version History, or click the filename in the title bar and choose Version History.
- This only exists for files saved to OneDrive or SharePoint. Locally saved files do not keep versions this way.
Word documents saved on your own computer
You still have options, and one of them is underused:
- File → Info → Properties → Advanced Properties → Statistics shows Total Editing Time and a revision count. A paper with several hours of accumulated editing time is hard to square with an accusation of pasting.
- File → Info → Manage Document may hold autorecover versions.
- Search your drive for
.asdand.wbkfiles, which are Word autorecover and backup copies. - File properties in Windows Explorer or Finder show created and modified dates.
Other apps
- Apple Pages: File → Revert To → Browse All Versions.
- Notion: page menu → Page history. How far back it goes depends on your plan, so check it early.
- Scrivener: Documents → Snapshots.
- Overleaf or LaTeX with git: the full commit history, which is excellent evidence.
- Obsidian, Ulysses, and similar: check for a version or snapshot feature, and check your system backup (Time Machine, File History) for earlier copies.
If you wrote it somewhere with no history
If you typed directly into a submission box, or wrote in an app that keeps no versions, you do not have this evidence and you should not pretend otherwise. Say so directly and move to the other categories in section 3. Being straightforward about a gap in your evidence builds more credibility than an overreach that falls apart.
Then start using a tool that keeps history for everything you write from now on. This is the cheapest insurance available to a student right now.
3. Every kind of evidence, ranked honestly
Not all evidence carries the same weight. Lead with the top of this list. Bring the bottom of it as support, and be honest with yourself about which is which.
Version and edit history
A timestamped record of the document growing over time. Draftback replays are the strongest single artifact. See section 2.
Your research trail
Browser history from your research sessions, library borrowing records, downloaded and annotated PDFs, database search logs, handwritten or digital notes, and outlines. These show the thinking that preceded the draft.
Timestamped communications
Drafts emailed to yourself or a classmate, messages asking someone to read it, writing center appointment records, tutoring session notes, and any earlier feedback from the same instructor on an outline or draft.
Your own command of the work
Be ready to discuss your argument without notes: why you chose those sources, what you cut and why, what you found hardest. Some instructors will simply ask you to explain a paragraph. Someone who wrote it can. Prepare for this, because it is common and it works in your favor.
Handwritten drafts, notes, and whiteboard photos
Physical artifacts with dates on them. Weaker on their own, useful as corroboration alongside a digital timeline.
A report from another AI detector, including ours
This is the bottom of the list and we want to be direct about why. A second detector is the same category of tool that produced the accusation, using the same kind of statistical guesswork. It is one data point, it is not proof, and presenting it as proof invites the reasonable reply that detectors disagree all the time. See section 10.
4. Build a one page timeline
This is the highest return on ten minutes of work in the whole guide. A single page, printed, that lays out how the paper actually came to exist. It reframes the meeting from an interrogation about a score into a review of a documented process.
Keep it factual and short. Attach your screenshots behind it.
Print two copies. Give one to the person you are meeting and keep one in front of you. Handing someone a document at the start of a meeting changes who is leading it.
5. What the research actually says
Use this as your second argument, after your evidence. Every claim below is sourced, and the sources are ones an instructor or a panel can check for themselves. That matters more than volume.
Turnitin marketed a false positive rate below 1% at the document level. In June 2023, its Chief Product Officer disclosed that in real world use the sentence level false positive rate was around 4%, attributing the gap to the difference between lab testing and live student writing.
Source: Inside Higher Ed, 1 June 2023
Vanderbilt University disabled Turnitin's AI detector in August 2023. Their stated reasoning: against roughly 75,000 papers submitted annually, even the claimed sub 1% error rate implied around 750 papers could be wrongly flagged. They also cited a lack of transparency about how the tool works and its documented bias against non native English writers.
Stanford researchers tested seven widely used AI detectors against 100 TOEFL essays written by non native English speakers. More than half were misclassified as AI generated, and one detector flagged up to 99% of them. Essays by native speaking US eighth graders were misclassified about 23% of the time. The proposed cause is that these tools treat simpler word choice as a sign of machine writing.
Source: Liang, Yuksekgonul, Mao, Wu and Zou, "GPT detectors are biased against non-native English writers," Patterns, 10 July 2023
In 2025 the US Federal Trade Commission acted against AI detection vendor Workado over an advertised 98% accuracy claim. The FTC found the tool's real accuracy on general purpose text was about 53%. The final order was approved in August 2025. Advertised detector accuracy figures are not, on their own, reliable.
Turnitin licenses to institutions, not to students. You cannot submit your own paper to it before turning it in, and whether you can even see the report generated about you depends on a setting your instructor controls. It is fair to point out that you are being judged by a tool you were never given access to.
How to use this well. Do not print all of it and slide it across the table. Pick the one or two items that fit your situation, cite them accurately, and be ready to say where they came from. One verifiable fact stated calmly is worth more than a stack of printouts.
6. If English is not your first language
You are in the group these tools fail most often, and it is worth understanding exactly why, because it is not a judgment about the quality of your work.
AI detectors largely measure how statistically predictable your word choices are. Writing that uses common words and steady sentence patterns scores as more predictable, and these tools read predictability as a sign of machine generation. Writing carefully in a second language tends to produce exactly that pattern. The Stanford study in section 5 found more than half of non native TOEFL essays wrongly flagged, with one detector flagging up to 99 percent, against about 23 percent for native speaker essays.
This is measurable, published, and specific. You can say so directly:
If your institution has an office for international students, an ombudsman, or a student advocate, contact them before the meeting. Many have handled this exact situation before and some have successfully argued it as an equity issue.
7. Questions to ask in the meeting
Asking these calmly does two things: it gets you information you need, and it establishes that you understand the process. Write them down and bring them.
- Which tool produced this result, and what exactly did it report? A percentage with no tool name attached is not evidence you can respond to.
- May I see the full report rather than just the score? Sentence level detail often shows flags on quotations, citations, or standard academic phrasing, which is worth pointing out.
- Is the detector result the sole basis for this concern, or is there other evidence? This is the most important question in the list. Many institutional policies state that a detector score alone is not sufficient grounds.
- What does our written policy say about using AI detection as evidence? Ask for the policy in writing. Read it before you respond to anything.
- Is this an informal conversation or a formal allegation, and is it being recorded anywhere? Your rights differ between the two. You are entitled to know which one you are in.
- May I respond in writing after I have gathered my drafts and version history? Almost always granted, and it converts a surprise conversation into one you can prepare for.
- May I bring an advisor or support person to any further meeting? Most institutions allow this in formal proceedings. Many students never ask.
- What are the possible outcomes, and what is the appeals process and deadline? Appeal windows are often short. Find out the date now.
8. What to actually say
Adapt these to sound like you. Reading them word for word will come across as rehearsed, which is the opposite of what you want. The point is the shape of the response: calm, factual, evidence first, no pleading.
Opening
When shown a score
When you are pressed to admit it
When you need more time
When the detector result is the only evidence
Stay calm even if the conversation is not fair. Anger, however justified, gets remembered instead of your evidence. If you feel yourself losing composure, use the "more time" script and leave.
9. If it becomes a formal case
If it moves past a conversation with your instructor, treat it as a process with rules rather than a conflict with a person.
- Get everything in writing. After any verbal meeting, send a short email summarizing what was said and asking for confirmation. That creates a record you control.
- Find your advocate. Most institutions have an ombudsman, a dean of students office, a student government advocacy service, or student legal services. These are free and they know the policies better than you will after a week of reading.
- Read the policy and the deadlines. Appeal windows are frequently ten business days or fewer. Diary the date the moment you learn it.
- If a disability or condition affects your writing style, and you are comfortable disclosing it, your disability services office may be relevant. Formulaic or highly structured writing is a documented feature of several conditions and has been raised in real cases.
- Bring your advisor to any hearing where it is permitted.
These cases are contestable, and students have won them. In February 2026 a New York court ordered a university to expunge a student's record after finding a "100% AI" determination was without valid basis. A separate federal suit was filed in California in May 2026 over a Turnitin flag on a high school essay. You are not the first person to be in this position and there is a growing record of these accusations being overturned.
10. Where a tool like ours fits, and where it does not
We make an AI detector, so read this section knowing that. We would rather be straight with you than sell you something that will not help.
What we cannot do
We cannot prove you wrote your paper. No detector can, ours included. We are the same class of tool that flagged you, working from the same kind of statistical guesswork, and we are wrong sometimes too. If you walk into your meeting and say "a different detector says I am fine," the fair response is that detectors disagree constantly, which is part of the problem.
What we are actually good for, in order:
- Before you submit. This is the real use. You cannot run Turnitin on your own work, so you have no way of knowing what your instructor will see. We show you, sentence by sentence, which passages read as machine-like, so you can look at them again before anyone else does. Whether you change them is up to you. Sometimes the honest answer is that a paragraph is just plainly written.
- As one supporting document, clearly labeled as one. Behind your version history, not in front of it. Presented as an additional data point rather than as proof.
- To understand what you are being accused of. Our linguistic breakdown shows which features drive a score, which makes the accusation something you can actually discuss rather than an unexplained number.
We show a confidence level with every result, including telling you when we are not sure, and there is a middle band for text that could genuinely go either way. We do not sell results to schools and we do not store your text.
Check your writing before someone else does.
Scan your text freeThree scans, no account, no card. Your text is not stored.
This is general information, not legal advice. Academic integrity policies differ substantially between institutions and between countries. Read your own institution's written policy, and contact your student advocacy, ombudsman, or legal services office for advice about your specific situation.
ClearTextAI produces probabilistic estimates, not determinations of fact. No AI detection tool, including ours, should be the sole basis for an academic disciplinary decision.
Last reviewed 8 September 2026.