A PROBABILITY
The score is not a share of cheating
Turnitin says there is no “right” or “target” score. A 30% result means the model thinks about 30% of the text looks AI-written, not that 30% was copied from a chatbot.
FALSELY FLAGGED, OCTOBER 2026
How to prove you didn't use AI, step by step
An AI flag on work you wrote yourself is stressful, but it is not a verdict. Here is what the score means, which evidence carries weight, how to raise it with your instructor, and how a formal appeal works.
Scan my text free Updated
essay-final.txt
$ wc -w essay-final.txt
1840 essay-final.txt
$ head -c 240 essay-final.txt
MacBook Pro
9:41
ClaudeCleanLocal scanPASTED FROM MY ESSAY
SCANNING
…hidden characters
U+00A0no-break space1
U+00ADsoft hyphen1
U+200Bzero-width space1
These came in with pasted quotes. A clean scan says nothing about who wrote the essay.
9:41
ClaudeCleanLocal scanPASTED FROM MY ESSAY
SCANNING
…hidden characters
U+00A0no-break space1
U+00ADsoft hyphen1
U+200Bzero-width space1
These came in with pasted quotes. A clean scan says nothing about who wrote the essay.
Short answer
To prove you didn't use AI, show how the work was made: your Google Docs or Word version history, dated drafts and notes, your sources, and a calm meeting where you explain the argument. A detector score is a prediction, not proof. Turnitin itself says its AI score should not be the only basis for action against a student.
WHAT A FLAG MEANS
What does an AI detection score prove?
Less than it looks like. A score is a model's guess about how predictable the wording is. It does not record how you wrote anything.
KNOWN ERROR
Vendors publish false positive rates
Turnitin puts its rate under 1% per document and about 4% per sentence. It adds that 54% of wrongly flagged sentences sit right next to real AI writing, so in a fully human essay the picture is different.
A STARTING POINT
Built to start a talk, not end one
Turnitin tells instructors the model can misidentify human writing and should not be the sole basis for adverse action. It even publishes a guide for students handling false positives.
Use the information to initiate a conversation, not to draw a conclusion.
PUBLISHED FALSE POSITIVE RATES
How often human writing gets flagged
What the companies that sell detectors say about their own tools.
Turnitin, document levelunder 1%
documents with 20% or more AI writing
Turnitin, sentence levelabout 4%
single sentences marked as AI
GPTZero, its stated target1% or less
"no more than 1%"
Turnitin, in GPTZero's 2023 test2%
a rival's benchmark, not neutral
WHY 1% STILL MATTERS
Vanderbilt ran this math in 2023: about 75,000 papers a year at a 1% rate means roughly 750 honest papers flagged.
750human papers flagged at 1%
If you used Claude and want to know what Turnitin can see in its text, that is covered in can Turnitin detect Claude. This page is for people who wrote the work themselves and need to show it.
THE EVIDENCE
What evidence proves you wrote it?
Anything that shows the work being made beats anything that only judges the finished text. Collect it before you reply to anyone.
EVIDENCE STRENGTH
What carries weight, strongest first
Tap a rung to see why it helps and what to watch for. The order is our editorial ranking, based on how hard each item is to fake later.
RUNG 1 OF 7
Edit history across many sessions
Strength
Why it helps
Dozens of timestamped versions over days show the essay growing, getting cut and getting rewritten. That pattern is very hard to fake after the fact.
Watch out
Export or screenshot it now. Google Docs needs edit access to show versions, and Word only keeps them for files on OneDrive or SharePoint.
| Evidence | What it shows | Where to get it | Limits |
|---|---|---|---|
| Version historyStart here | Each saved version, who edited it and what changed | Google Docs: Last edit or File, then Version history. Word: File, Info, Version History | Docs needs edit access. Word needs OneDrive, SharePoint or AutoSave |
| Writing replay | The document being typed and edited over time, like a video | Draftback or Revision History in Chrome, for Google Docs | Google Docs only. Ask before installing anything on a school account |
| Authorship report | Which text was typed, pasted or AI-assisted, with a replay | Grammarly Authorship, switched on before you write | Cannot cover past writing. Reporters showed it can be fooled |
| Drafts and notes | The thinking before the final text: outlines, messy drafts, margin notes | Your notebook, notes app, email drafts, files sent to a friend or tutor | File dates can change when copied. Keep originals untouched |
| Research trail | That you found and read the sources you cite | Browser history, library loans, highlighted PDFs, database searches | Shows research, not who wrote the sentences |
| Another detector | That detectors disagree | Any free AI checker | Proves nothing about authorship. Do not lead with it |
Swipe to compare →
Our ranking is editorial. Product behavior is from Google, Microsoft, Draftback, Revision History and Grammarly's own pages, October 2026.
Freeze the original before anything else
Do not tidy up, rename or move the document you submitted. Make a copy for any edits. If the file changes after the accusation, the timeline gets harder to read, and that helps nobody.
GOOGLE DOCS AND WORD
How to find and save your version history
Most students already have the best evidence and do not know where it lives. Both editors keep it, with some limits.
WHAT VERSION HISTORY SHOWS
A week of drafts, one timestamp at a time
A sample essay we wrote, shown the way an editor's version history lays it out: dates on the right, changes highlighted in the text.
docs.google.com/document/d/heat-islands-essay
Urban heat islands: outline
1. What makes a city hotter than the land around it
2. Who pays for it (Austin, Phoenix)
3. Fixes: trees, cool roofs, less asphalt?
64 words
Oct 2, 9:14 PM
MacBook Pro
- 01
Open version history in Google Docs
Click Last edit at the top of the document, or go to File, Version history, See version history. Google says you need edit permission to browse earlier versions, so use the account that owns the doc. - 02
Click through each version
The panel lists when each version was saved and who made it, and selecting one shows what changed. Note the dates of your first outline, your drafts and your last edit. - 03
Name key versions
Use More, Name this version for milestones like Outline, First draft and Submitted. Google allows up to 40 named versions per document, and you can filter to show only those. - 04
Open version history in Word
Go to File, Info, Version History. Microsoft says this works for files saved to OneDrive or SharePoint, and personal accounts can get back the last 25 versions. A file that only ever lived on your desktop may have none. - 05
Capture it in a form you can share
Take dated screenshots of the history panel and of a few versions side by side, or share view access to a copy. A PDF of screenshots attaches cleanly to an email. - 06
Add a replay if you wrote in Google Docs
Draftback replays the history of any Google Doc you can edit, even one written before you installed it. A two-minute replay of your essay growing is easy for an instructor to follow.
Word's history is thinner than Google's
Microsoft notes that Word's version history does not show a detailed record of every change between versions. Pair it with drafts and notes, and if you are still writing, save to OneDrive with AutoSave on so future work keeps a fuller record.
QUOTED FROM A PDF OR WEBSITE?
See what invisible characters came in with your pasted quotes.
ClaudeClean shows whether pasted text carries no-break spaces, zero-width spaces or soft hyphens, and removes them in your browser. It cannot prove who wrote an essay, so use it on a copy, never on the original you may need as evidence.
- 300 rewrite credits a month
- 20,000 characters per request
- .txt, .docx and .pdf files
REPLAY TOOLS
Draftback, Revision History and Grammarly Authorship compared
These tools turn a document's history into something an instructor can watch. Two work on past writing; the rest only record from the day you switch them on.
| Tool | Works with | Covers past work? | What it shows | Worth knowing |
|---|---|---|---|---|
| DraftbackWorks on old docs | Google Docs, in Chrome | Yes, replays any Google Doc you can edit, including old ones | Keystroke replay with play, pause and rewind, plus writing time insights | 30-day free trial with no card on its store listing. Says document data stays on your computer |
| Revision History | Google Docs and Slides, in Chrome | Yes, reads the existing revision history | Video replay, active writing time, pastes with time and size | Built for teachers. Says it flags patterns as a starting point for conversations, not proof |
| Grammarly Authorship | Google Docs, Microsoft Word and the Grammarly editor | No, it records from the moment you switch it on | Typed, pasted and AI-assisted text, sources of pastes, a replay | Free and Pro plans. Plagiarism Today found reworded AI text labeled as mostly typed |
| GPTZero writing report | GPTZero's Writing Replay and Writing Report | GPTZero pitches it as part of your everyday writing, so plan ahead | How a piece was written, edited and pasted, plus activity insights | Made by a detector company, with built-in AI detection that flags large pastes |
| Turnitin Clarity | Inside your school's learning platform, if the school buys it | No, you write inside it from the start | The writing process to the instructor, alongside the final submission | Launched March 4, 2025. You cannot add it yourself |
Swipe to compare →
From each product's own page or store listing, checked October 2026. Features and prices change; check before you rely on one.
ONE CAVEAT
No tracker is tamper-proof
Plagiarism Today showed AI text reworded with Grammarly's own tools coming out labeled as mostly typed by a human. Process records are strong evidence, which is why they work best next to your drafts, notes and a conversation.
SCHOOL ACCOUNTS
Ask before you install
Some schools block Chrome extensions on managed accounts. If yours does, open the doc on a personal browser with the same account, or ask your instructor or IT office to run the replay with you.
NO HISTORY?
What if you don't have version history?
Plenty of honest writers have none: they wrote offline, in a notes app, or pasted from a draft at the last minute. You still have options.
I wrote it in a notes app and pasted it into the submission
Check the notes app for older versions or deleted notes, and look for the draft on other devices that sync. Phone screenshots, messages where you mentioned the essay, and photos of handwritten plans all help place the work in time.
I wrote it in Word on my laptop with AutoSave off
Look for earlier copies: email attachments, files sent to a tutor or friend, cloud backups, and Word's own recovered or unsaved files. Even one older draft with different wording shows the essay changed over time.
I wrote it in one long sitting
Say so plainly. Bring your notes and sources, describe the order you wrote it in, and offer to discuss the argument or write a short passage on the same topic under supervision.
I have nothing but the final file
Then the conversation is your evidence. Reread the essay, know why you chose each source, and be ready to explain one paragraph in detail. When Australian Catholic University investigated AI flags, students were asked for handwritten notes and search histories, so gather whatever research trail you have.
TALK TO YOUR INSTRUCTOR
How to talk to your instructor about an AI flag
Most cases start informally. A calm first message and a prepared meeting resolve more than an angry email ever will.
THE FIRST MESSAGE
Five short lines, no panic
A sample first email. It states the facts once, offers evidence and asks to meet. Names and details are made up, so swap in yours.
9:41
CancelNew Message
To: Professor Lee
Subject: Meeting about my essay
Hi Professor Lee,
I saw that my essay on urban heat islands was flagged for possible AI use. I wrote it myself, and I would like the chance to show how I wrote it.
I can share the Google Docs version history, my outline from October 2 and my source notes.
Could we meet this week? I am free Tuesday or Thursday afternoon, or I can come to office hours.
Thank you, Maya Ortiz, ENG 102
Version history (PDF)412 KB
- 01State the facts once, calmly
- 02Offer evidence up front
- 03Ask for a conversation
- 04Keep it short and polite
Leave out what you think of AI detectors and any guesses about who reported you. Bring that context to the meeting if it is relevant, with sources.
In the meeting, do
- Lead with the process. Open the version history and walk through how the essay grew, session by session.
- Explain your choices. Why this thesis, why these sources, what you cut and why.
- Ask what the concern is. Is it only the score, or something in the text? You can only answer what you know about.
- Offer a quick check. A short supervised writing sample or a few questions on the content can settle things fast.
- Follow up in writing. Send a short email summing up what was agreed, so there is a record.
In the meeting, avoid
- Arguing about AI in general. Keep it about your essay. Bring research on detector errors only if the score is the whole case.
- Showing other detector scores first. A low score elsewhere shows detectors disagree, not that you wrote it.
- Guessing or exaggerating. If you used a grammar checker or a tutor, say so. Small honest details build trust.
- Signing anything on the spot. If you are offered an admission or penalty, ask for time to read it and get advice.
- Going silent. Missing a meeting or deadline can count against you, even with good evidence.
Not sure what made the text look AI-like? Our signs of AI writing guide lists the phrasing readers react to. It is useful context for the meeting, but do not rewrite the submitted work to remove them.
FORMAL APPEAL
How the formal appeal process works
If the informal route does not settle it, most schools have an academic integrity process with written steps and deadlines. Names vary; the shape is similar.
- 01
Get the allegation in writing
Ask for the exact concern, the policy you are said to have broken, and the full detector report if a score is involved. - 02
Read the policy and the deadlines
Find the academic integrity policy and the appeal timeline. Missing a response window can end your options early. - 03
Get support
Student advocates, ombuds offices, student unions and academic advisors help with these cases. Many schools let you bring a support person to hearings. - 04
Write a clear statement
Say what you did, when, and what evidence shows it. Attach the version history, drafts, notes and sources, labeled by date. - 05
Prepare for the hearing
Expect questions on the content and your process. Reread the essay and your notes, and bring printed copies of key evidence. - 06
Appeal the decision if it is wrong
Most policies allow an appeal on grounds like new evidence or a process error. Point to what was missed, calmly and specifically.
YOUR CHECKLIST
Tick off each step as you go
Works in your browser only. Nothing is saved or sent. Copy it into your notes if you want to keep it.
0 of 15 done0%
01 FIRST 24 HOURS
0/502 BEFORE THE MEETING
0/503 IF IT GOES FORMAL
0/5A statement template you can adapt
Written statement: [course code], [assignment], [date submitted]
I wrote this assignment myself. I did not use AI tools to generate its text.
How I wrote it:
- [date]: outline (attached, page 1)
- [date to date]: first draft, written over [number] sessions
- [date]: revised the second section after feedback from [person or none]
Evidence attached:
1. Version history export from [Google Docs / Word] showing [number] versions
2. Notes and outline, with dates
3. List of sources with the dates I accessed them
I am happy to explain any part of the essay in person.Replace every bracket. Keep it factual and short; your attachments do the heavy lifting.
WHO GETS FLAGGED
Who gets falsely flagged by AI detectors?
Anyone can be, but research points to a few patterns worth knowing if you need to explain why your work was flagged.
NON-NATIVE WRITERS
Simpler vocabulary looks predictable
Liang et al. found the same detectors were near-perfect on US eighth-grade essays but flagged most TOEFL essays. When the TOEFL essays were rewritten with richer vocabulary, the average false positive rate fell to 11.77%.
LOW SCORES
Small scores are the least reliable
Turnitin hides scores from 1% to 19% because false positives are more common there. If your report shows an asterisk, that alone is weak ground for an accusation.
THE OTHER SIDE
Newer research is more mixed
A 2026 study of Czech-language writing found no systematic bias against non-native writers in current detectors. Bias may depend on language and detector, so cite the research that fits your case.
Bloomberg Businessweek also told the story of Moira Olmsted, a student who received a zero on an assignment after an AI detector flag. Note that Liang's study did not test Turnitin, so quote it as evidence about detectors in general, not about Turnitin specifically.
SCHOOLS THAT SWITCHED IT OFF
Which universities disabled Turnitin's AI detector?
If your school still uses it, these decisions are useful context. Each one is from the university's own page or a news report.
2023
Turnitin adds AI writing detection
Schools that already used Turnitin for plagiarism checks get an AI score in the same report.Jun 14, 2023
Turnitin publishes its false positive rates
Under 1% at the document level for documents with 20% or more AI writing, about 4% at the sentence level, and the advice to start a conversation rather than draw a conclusion.Aug 16, 2023
Vanderbilt turns the detector off
Vanderbilt noted it submitted about 75,000 papers to Turnitin in 2022, so even 1% would have meant around 750 wrongly flagged papers. It also cited bias against non-native English writers.Jul 2024
Low scores get an asterisk
For submissions after July 8, 2024, Turnitin shows scores from 1% to 19% as an asterisk with no highlights, because false positives are more common in that range.Mar 2025
Clarity launches, and ACU drops the detector
Turnitin launched Clarity, a writing space that shows the process. The same month, Australian Catholic University stopped using the AI detector after logging nearly 6,000 misconduct cases in 2024, about 90% of them about AI.Aug 27, 2025
Turnitin starts flagging humanizers
Reports from Plagiarism Today and Campus Technology describe a new check for text altered by AI bypasser tools, in English at launch.Jan 1, 2026
Curtin switches it off
Curtin University announced in September 2025 that Turnitin's AI writing detection would be disabled from the start of 2026.
UT Arlington never turned it on
UT Arlington's Turnitin page says: “At this time, UTA leadership has elected not to enable AI-Writing detection due to the potential for false positives.” Many other schools are reported to have made similar calls; we list only those we could confirm from a primary source.
GPTZERO VS TURNITIN
GPTZero vs Turnitin: which AI detector is more accurate?
Both publish low false positive rates for themselves. The only head-to-head numbers come from one of the two companies.
| Figure | False positives | Measured by | Caveat |
|---|---|---|---|
| Turnitin, own numbers | Under 1% per document (20%+ AI writing), about 4% per sentence | Turnitin, June 2023 | Scores of 1% to 19% shown only as an asterisk |
| GPTZero, own numbers | Kept at 1% or less, per GPTZero | GPTZero, January 2025 | Not stated on the benchmark page |
| GPTZero's test of both | 2% for each. Accuracy 99% for GPTZero, 95% for Turnitin | GPTZero, November 2023. GPTZero says newer Turnitin tests are not possible | A rival grading a rival |
| Liang et al., independent | 61.22% average on TOEFL essays across seven detectors, GPTZero included | Stanford researchers, Patterns, 2023 | Turnitin was not tested |
Swipe to compare →
GPTZero says its last Turnitin benchmark is from 2023 because Turnitin is closed. Treat vendor figures as claims, not audits.
For your case the question is narrower than which tool is better. Whatever flagged you, a percentage is a probability with a known error rate, and your process evidence answers a different question: how the work was made.
WHAT NOT TO DO
Mistakes that make a false flag worse
Panic moves feel productive and usually backfire. Avoid these.
HUMANIZERS
Do not run your own essay through a humanizer
It replaces your words with a tool's, and since August 2025 Turnitin flags text it thinks was altered by AI bypassers. You would be swapping a weak accusation for a stronger one. Our Claude humanizer guide explains what these tools change.
THE ORIGINAL
Do not edit or delete the source document
Even well-meant cleanup adds new versions after the accusation and muddies the timeline. Copy it, and leave the original alone.
SCORE SHOPPING
Do not lead with ten other detector results
A stack of low scores from other checkers mostly proves that detectors disagree. Bring one comparison at most, and only after your process evidence.
Cleaning characters is not proof of anything
Hidden characters such as no-break spaces often come from PDFs and web pages you quoted. Removing them changes no words, and a clean scan tells nobody who wrote the text. If you want to know what they are, see watermarks vs hidden characters.
WHERE CLAUDECLEAN FITS
A small, honest role: checking pasted text for hidden characters
ClaudeClean is a text cleaner, not an authorship tool and not an AI detector. Here is all it does with a copy of your essay.
01
> source text
Paste a copy
Paste a copy of your text or drop a .txt, .docx or .pdf. Up to 2,000 characters works without an account. Never work on the original you may need as evidence.
02
scan report
3 zero-width1 soft hyphencleanable signals2See hidden characters
The scan runs in your browser and marks zero-width spaces, no-break spaces, soft hyphens and lookalike letters, which often ride in with quotes pasted from PDFs or websites.
03
before → after
clearusefulanswerfocusedhelpfulresponse
Copied ✓Clean, or leave it
Remove them locally if you want tidy text for a new draft. The optional rewrite sends text to OpenAI and is not something to use on work you need to defend.
PRICING
The scan is free. Pro is for heavier cleanup.
Free accounts get 10 credits a month and Pro is $5 a month. Neither proves authorship or promises any detector result.
PRO
CLAUDECLEAN PRO
ClaudeClean Pro
$5/ month
Cancel anytime.
- 300 credits every month (≈300,000 characters of rewriting)
- 20,000 characters per request
- Clean .txt, .docx and .pdf files
- Light / medium / strong intensity
- Dashboard with history
Checking account…
Free: 10 credits a month with an account, or try 2,000 characters right here without one. Strong rewrites cost 2× credits.
TERMS YOU WILL HEAR
Words used in AI misconduct cases
Schools use slightly different names. These are the common ones.
- False positive
- Human writing that a detector labels as AI-generated. Every detector has a rate for this, and vendors publish theirs.
- AI writing score
- The share of a document a detector predicts is AI-written. A prediction about the text, not a record of how it was made.
- Version history
- The saved timeline of a document's edits in Google Docs, Word on OneDrive and similar editors, with dates and authors.
- Process evidence
- Anything that shows the writing happening: versions, replays, drafts, outlines, notes and research history.
- Academic integrity referral
- A formal report of suspected misconduct to the office that handles it, which starts the written process.
- Viva or oral check
- A short conversation where you explain your work, used by some schools to confirm authorship.
WATCH
False positives on Turnitin, explained for accused students
A three-minute video for teachers and students, including how a revision replay shows the writing process.
SOURCES
Where these facts come from
Detector makers first, then research, universities and the tools you can use.
Turnitin
Understanding the false positive rate for sentences
June 2023: under 1% document rate, about 4% sentence rate, 54% near real AI text, conversation not conclusion.
View sourceTurnitin Guides
AI writing detection in the classic report view
Asterisk for scores from 1% to 19% since July 2024, and the warning against using the score as the sole basis for action.
View sourceGPTZero
AI accuracy benchmarking
January 2025: false positive rate kept at 1% or less, and GPTZero's November 2023 comparison with Turnitin.
View sourcePatterns (arXiv)
GPT detectors are biased against non-native English writers
Liang et al., 2023: seven detectors, 91 TOEFL essays, 61.22% average false positive rate, 11.77% after vocabulary edits.
View sourcearXiv
Revisiting the bias against non-native speakers in GPT detectors
2026 study on Czech-language writing that found no systematic bias in current detectors.
View sourceBloomberg Businessweek
AI detectors falsely accuse students of cheating
October 2024: 1% to 2% of 500 pre-ChatGPT essays flagged by GPTZero and Copyleaks, and students' stories.
View sourceVanderbilt University
Why we're disabling Turnitin's AI detector
August 2023: the 75,000-paper math and concerns about non-native English writers.
View sourceCurtin University
Update on Turnitin AI detection tool
AI writing detection disabled from January 1, 2026.
View sourceABC News (Australia)
Australian Catholic University and AI cheating cases
Nearly 6,000 misconduct cases in 2024, about 90% about AI; detector dropped in March 2025.
View sourceGoogle Docs Editors Help
Find what's changed in a file
Version history, naming versions, the 40 named version limit and the edit permission requirement.
View sourceMicrosoft Support
View previous versions of Office files
Version history needs OneDrive, SharePoint or AutoSave; personal accounts keep the last 25 versions.
View sourceGrammarly
Grammarly Authorship
Records typed, pasted and AI-assisted text with a replay, in Google Docs, Word and the Grammarly editor.
View sourceFAQ
Proving you didn't use AI: FAQ.
Show how the work was made. Export your Google Docs or Word version history, gather dated drafts, outlines, notes and sources, then ask your instructor to meet and walk them through your argument. A detector score is a prediction, not proof.
Yes. Turnitin puts its document-level false positive rate under 1% for documents with 20% or more AI writing and its sentence-level rate around 4%. It hides scores from 1% to 19% behind an asterisk and says the score should not be the only basis for action against a student.
Nobody neutral has settled it. GPTZero says it keeps its false positive rate at 1% or less, and its own 2023 test scored GPTZero ahead of Turnitin, but a vendor testing a rival is not independent. Independent studies found much higher error rates for non-native English writers.
It is the strongest evidence most students have. Many timestamped versions over several days, with text added, cut and reworded, show the writing happening. One big paste with few edits shows less, so explain where pasted parts came from.
Use what you do have: handwritten notes, outlines, earlier drafts, research history, library loans and your past writing. Offer to explain the essay in person or write a short sample under supervision, and ask how the school weighs a detector score.
Policies differ by school, so read yours. Turnitin's own guidance says the AI score should not be used as the sole basis for adverse action, and Vanderbilt, Curtin, UT Arlington and Australian Catholic University have turned the detector off.
No. It changes your own words, it can make the text read worse, and since August 2025 Turnitin flags text it thinks was altered by AI bypasser tools. Your original document and its history are worth more than any lower score.
Detectors lean on how predictable the wording is, and writers with a smaller English vocabulary tend to use more common words. In Liang et al. (2023), seven detectors flagged an average 61.22% of TOEFL essays written by people as AI.
No. ClaudeClean shows whether pasted text carries invisible characters such as no-break spaces or zero-width spaces, and can remove them in your browser. That says nothing about who wrote the text, so do not present it as evidence of authorship.