Method
A style classifier, not a key
Turnitin scores prose segments for how likely they are to be AI-written, as its help center explains. It looks for patterns common in AI writing, with no key involved.
TURNITIN, GPTZERO AND CLAUDE
Can Turnitin detect Claude? Watermark facts and 2026 data
Turnitin and GPTZero guess from writing style. Anthropic's watermark sits in word choice and only Anthropic's key can read it. Here is what each tool sees, with the published numbers and what to do if you get flagged.
Scan my text free Updated
claude.ai
Today
Write the opening paragraph of my essay on urban heat islands.
Message Claude
MacBook Pro
9:41
ClaudeCleanLocal scanPASTED FROM CLAUDE
SCANNING
…hidden characters
U+200Bzero-width space1
U+00A0no-break space1
Turnitin scores writing patterns, not these characters. Rewrite in your own words.
9:41
ClaudeCleanLocal scanPASTED FROM CLAUDE
SCANNING
…hidden characters
U+200Bzero-width space1
U+00A0no-break space1
Turnitin scores writing patterns, not these characters. Rewrite in your own words.
Short answer
Turnitin can flag Claude text with its AI writing classifier, but it can't read Anthropic's watermark. That check needs Anthropic's key, and its detection API is in private preview for eligible organizations. GPTZero, Pangram, Copyleaks and Originality.ai are in the same spot: they predict from style, and every one of them makes mistakes.
WHO READS WHAT
Which AI detectors can read the Claude watermark?
Only one tool on this list holds the key. The rest are classifiers trained to spot AI-style writing, and they were around long before Claude had a watermark.
TWO WAYS TO READ ONE ESSAY
Classifiers read style. Only Anthropic's key reads the watermark.
The same Claude text goes down two very different paths. Tap a classifier to see its row from the table below.
Claude essayStyle and word choice both in it
Style classifiers
- Reads
- AI writing classifier that scores segments of a submission
- Who can use it
- Instructors, through a school license
- False positive notes
- Under 1% per document with 20%+ AI writing, about 4% per sentence (Turnitin, 2023)
Output: a prediction from style. Watermark read? No.
Anthropic detection API
- Reads
- Statistical check of word choices against Anthropic's secret key
- Who can use it
- Private preview: regulators, media, researchers, educational organizations and others
- Reads Claude watermark?
- Yes, the only tool on this list that holds the key
Output: Claude was likely involved at some point. Unreliable on short text.
| Detector | Method | Reads Claude watermark? | Who can use it | False positive notes |
|---|---|---|---|---|
| Turnitin | AI writing classifier that scores segments of a submission | No | Instructors, through a school license | Under 1% per document with 20%+ AI writing, about 4% per sentence (Turnitin, 2023) |
| GPTZero | Deep learning model plus sentence-by-sentence classification | No | Public web app | Says it keeps false positives at no more than 1%; 0 of 495 human passages flagged by Epoch AI |
| Pangram | Neural classifier trained on paired human and AI writing | No | Public web app | Pangram 4 claims about one error per 24,000 documents; 0 of 495 in Epoch AI's test |
| Copyleaks | AI content classifier | No | Public web app and LMS plugins | Markets a 0.2% false positive rate from its own testing |
| Originality.ai | Classifier built on a modified BERT model | No | Paid, plus 3 free scans a day up to 2,000 words | Flagged 19 of 495 human passages (3.84%) in Epoch AI's 2026 test |
| Anthropic detection APIKeyed | Statistical check of word choices against Anthropic's secret key | Yes | Private preview: regulators, media, researchers, educational organizations and others | Unreliable on short text; a hit means Claude was likely involved at some point |
Swipe to compare →
Vendor figures are self-reported. Epoch AI's numbers come from one July 2026 test with its own sample, so treat them as a snapshot, not a ranking.
THE PUBLISHED NUMBERS
How accurate is Turnitin on Claude text?
Turnitin hasn't published a rate for Claude specifically. These are the figures it and independent researchers have put on the record.
61%
Average rate seven detectors misread non-native TOEFL essays as AI (Turnitin not tested)
TRY THE DISPLAY RULES
When does the report show a number?
Move the sliders. Length and AI share are your own made-up inputs; the rules applied to them are Turnitin's 2023 changes, as reported on this page.
- Turnitin raised its minimum from 150 to 300 words of prose in 2023.
- Scores under 20% get an asterisk instead of a number, because Turnitin saw more false positives there.
- The under 1% document false positive figure covers files with 20% or more AI writing.
Similarity and AI writing reportMOCK UI
AI WRITING35%Shown as a number
Turnitin's chief product officer Annie Chechitelli wrote that the document-level rate stays under 1% only for documents with 20% or more AI writing. Below that line the report shows an asterisk instead of a number, according to K12 Dive.
The same 2023 change raised the minimum length from 150 to 300 words. As of May 14, 2023, Turnitin had scored 38.5 million submissions, and 9.6% showed more than 20% AI writing.
BEFORE YOU HIT SUBMIT
Check what's in the file, not what a score might say.
ClaudeClean finds zero-width characters, soft hyphens and lookalike letters in your browser, and can run a best-effort rewrite with a before and after diff. It can't read the watermark or predict a Turnitin score.
- 300 rewrite credits a month
- 20,000 characters per request
- .txt, .docx and .pdf files
INSIDE THE REPORT
What does Turnitin look for in Claude writing?
Turnitin's detector is a prediction about style. These are the parts of it that matter for Claude text.
WHAT THE INDICATOR READS
Inside an AI writing report, step by step
A generic course page, not a real product screen. It follows the four things this section explains, in order, and ends without a score on purpose.
lms.example.edu/env-sci-101/essay-2/report
Why cities run hot at night
Essay 2 · submitted on time
Cities run hotter than the land around them because asphalt and concrete store heat long after sunset.
On my street in July, the pavement still felt warm at ten at night, and our window unit never shut off.
Researchers call this the urban heat island effect, and it shapes energy use, health and comfort in dense neighborhoods.
Our city planted street trees on two blocks last spring, so I measured shade temperatures there myself.
Green roofs, lighter pavement and street trees each lower surface temperatures in different ways.
AI WRITING INDICATOR
Counts the prose300+ words, so it can score
Splits it into segmentsEach segment scored on its own
Reads writing patternsPatterns common in AI writing
Checks Anthropic's watermarkNo key, so this step does not exist
Shown to the instructor
Highlighted segments and a prediction about style. Not proof of how the essay was written.
MacBook Pro
lms.example.edu/essay-2
Why cities run hot at night
Essay 2 · submitted on time
Cities run hotter than the land around them because asphalt and concrete store heat long after sunset.
On my street in July, the pavement still felt warm at ten at night, and our window unit never shut off.
Researchers call this the urban heat island effect, and it shapes energy use, health and comfort in dense neighborhoods.
Our city planted street trees on two blocks last spring, so I measured shade temperatures there myself.
Green roofs, lighter pavement and street trees each lower surface temperatures in different ways.
AI WRITING INDICATOR
Counts the prose300+ words, so it can score
Splits it into segmentsEach segment scored on its own
Reads writing patternsPatterns common in AI writing
Checks Anthropic's watermarkNo key, so this step does not exist
Shown to the instructor
Highlighted segments and a prediction about style. Not proof of how the essay was written.
Coverage
Claude models are on its list
Turnitin's AI writing FAQ lists Claude models, Sonnet 4.5, Sonnet 4.6 and Opus 4.5 among them, for its English detector. It publishes no per-model catch rate.
Length
300 words before it scores
Turnitin raised its minimum from 150 to 300 words of prose in 2023, per K12 Dive. Short answers, lists and code fall outside what it scores well.
Low scores
An asterisk under 20%
Since 2023, documents scoring under 20% AI writing get an asterisk instead of a number. Turnitin said it saw a higher incidence of false positives in that range.
Humanizers
It flags humanizer output too
In August 2025 Turnitin added detection for text run through AI humanizer tools, English only, as Plagiarism Today and Campus Technology reported.
Watermark
No public sign it holds the key
Anthropic says third-party detectors look for stylistic tells and cannot verify its watermark. Turnitin hasn't announced access to the API.
INDEPENDENT TEST, JULY 2026
Can GPTZero detect Claude?
Epoch AI ran Pangram, GPTZero and Originality.ai on human writing and on text from Claude Opus 4.8, GPT-5.5 and Gemini 3.1 Pro. The Claude slices are below.
- Pangram0/495 (0.0%)
- GPTZero0/495 (0.0%)
- Originality.ai0/495 (0.0%)
False positives on 495 human-written passages. Pangram and GPTZero flagged none.
- Pangram0/99 (0.0%)
- GPTZero0/99 (0.0%)
- Originality.ai0/99 (0.0%)
Claude passages written from a basic prompt. All three caught almost every one.
- Pangram0/99 (0.0%)
- GPTZero0/99 (0.0%)
- Originality.ai0/99 (0.0%)
Claude asked to write in a particular style. Misses rise for every tool.
Source: Epoch AI, 'AI detectors false negatives', July 15, 2026. Versions tested: Pangram v3.3.2, GPTZero 2026-05-11, Originality.ai Turbo 3.0.2. Bars are scaled to 15%.
So yes, GPTZero catches most Claude writing by style, and so do Pangram and Originality.ai. Across all three models in Epoch AI's test, style-imitation misses ran 10.10% for Pangram, 10.77% for GPTZero and 17.85% for Originality.ai.
None of this involves the watermark. GPTZero and Pangram describe trained classifiers, and neither page mentions reading Anthropic's mark.
We keep GPTZero's false positive rate at no more than 1% when evaluating AI versus human text.
HOW WE GOT HERE
When did Turnitin and Claude detection change?
Detection by style came first. The watermark arrived in 2026 and still isn't something schools can check on their own.
April 2023
Stanford study questions detectors
Liang et al. found seven detectors misread 61% of non-native TOEFL essays as AI on average, while scoring US eighth-grade essays almost perfectly.June 2023
Turnitin publishes false positive data
Turnitin reports under 1% at the document level and about 4% per sentence, adds the asterisk for scores under 20% and raises the floor to 300 words.August 16, 2023
Vanderbilt turns Turnitin's AI detector off
Vanderbilt disables the feature, citing the false positive rate and the lack of detail on how it works.August 27, 2025
Turnitin targets humanizer tools
Turnitin adds detection for text passed through AI humanizers, English only at launch.September 4, 2025
Curtin University schedules a switch-off
Curtin announces that Turnitin's AI writing detection will be disabled from January 1, 2026.July 15, 2026
Epoch AI tests detectors on Claude Opus 4.8
Pangram, GPTZero and Originality.ai catch nearly all plain Claude text, and miss more when Claude imitates a style.August 2, 2026
EU transparency rules apply
Article 50 of the EU AI Act starts to apply, asking AI providers to mark generated content, per the Commission FAQ.August 14, 2026
Anthropic watermarks Claude text
Anthropic announces its text watermark and a detection API in private preview, open to educational organizations among others.
SCHOOLS AND SCORES
Do universities trust Turnitin's AI score?
Some do, some switched it off, and most leave it to the instructor. Your course policy matters more than any percentage.
Academic integrity comes first
If your course bans AI help, no tool changes that, and disclosing how you used Claude is usually the safest move. If AI is allowed, follow the citation or disclosure rule your institution sets. This page explains detectors; it isn't advice to hide AI use.
WHAT 1% LOOKS LIKE
75,000 papers, about 750 wrongly flagged
Vanderbilt's own math for 2022. Each dot below stands for 75 papers.
SUBMITTED IN 2022
0
papers, 75 per dot
AT A 1% RATE
~0
could be wrongly flagged, 10 dots
Source: Vanderbilt University, 2023, explaining why it turned off Turnitin's AI detector. The dot positions are random; the count is not.
Vanderbilt, 2023
750 papers at a 1% rate
Vanderbilt submitted 75,000 papers to Turnitin in 2022. At a 1% false positive rate, it estimated about 750 could have been wrongly flagged.
Curtin, 2026
Detection off from January 1
Curtin University said the AI writing detection feature in Turnitin would be disabled from January 1, 2026.
Non-native writers
Simpler prose gets flagged more
In the Stanford study, 89 of 91 TOEFL essays were flagged by at least one detector. Turnitin wasn't part of that test.
FLAGGED?
What should you do if Turnitin flags your Claude text?
A flag starts a conversation. These are the questions that tend to matter in it.
A FLAG, HANDLED
What a calm reply to a flag looks like
A made-up exchange that walks through the four questions below in order.
9:41
PRProf. Rivera
Today 4:12 PM
Hi Sam, the AI writing report flagged parts of your Essay 2. Can we talk this week?
Sure. Could you show me which sentences were flagged?
I'll bring my outline, drafts and the version history.
Outline.docxEssay 2, draft 1Version history
And the course AI policy plus my Claude chat, since we were allowed to brainstorm with it.
Thanks, that helps. Bring all of it Thursday.
iMessage
- 01What did the report flag?Ask for the highlighted sentences and the percentage. Under 20% shows only an asterisk.
- 02What can show how you wrote it?Drafts, outlines, notes and version history all show a process.
- 03Did your course allow Claude?Allowed use with disclosure is a different talk from banned use.
- 04Could a watermark check settle it?Only eligible organizations can run one, and a hit is not fully conclusive.
What did the report flag?
Ask to see the highlighted sentences and the overall percentage. A score under 20% shows only an asterisk, and Turnitin has said false positives are more common in that range.
What can show how you wrote it?
Drafts, outlines, notes and the version history in Google Docs or Word all show a process. If you used Claude within the rules, the chat history shows exactly what it did.
Did your course allow Claude?
Read the syllabus wording. Many policies allow brainstorming or proofreading with disclosure, and a flag on allowed use is a different talk from a flag on banned use.
Could a watermark check settle it?
Only an eligible organization can run one through Anthropic's preview. Even then, Anthropic says a hit means Claude was likely involved and can't separate "Claude wrote this" from "Claude heavily edited this". Its help center calls a detected mark a signal that is not fully conclusive.
WHAT CLAUDECLEAN DOES
A character check you can run before any detector sees your file.
ClaudeClean doesn't read watermarks or predict scores. It cleans the character layer and can restyle text with a non-Claude model. No detector result is promised.
01
> source text
Paste or upload
Drop in text or a .txt, .docx or .pdf file. Scanning runs in your browser and ClaudeClean doesn't store your text.
02
scan report
3 zero-width1 soft hyphencleanable signals2Find hidden characters
The scan highlights zero-width spaces, soft hyphens, odd whitespace and lookalike letters, then removes them in one click.
03
before → after
clearusefulanswerfocusedhelpfulresponse
Copied ✓Optional rewrite with a diff
Rewrite sends text to OpenAI in Light, Medium or Strong passes. You see every change before you keep it.
PRICING
Clean the text free. Rewrite more with Pro.
The scan and hidden-character cleanup are free. Pro adds room for rewrites and files.
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.
RECEIPTS
Where these numbers come from
Every figure on this page links to the page it came from. Vendor claims are marked as vendor claims.
Anthropic
Claude text watermark explainer
How the mark works, its limits, and who can apply for the detection API preview.
View sourceTurnitin
False positive rate for sentences
Annie Chechitelli's 2023 post with the under 1% document and about 4% sentence figures.
View sourceEpoch AI
AI detectors and false negatives
July 2026 test of Pangram, GPTZero and Originality.ai on human, Claude, GPT and Gemini text.
View sourceGPTZero
How GPTZero works
Its method, the 1% false positive target and the 1.1% rate on TOEFL essays.
View sourceVanderbilt University
Why we're disabling Turnitin's AI detector
The 2023 decision and the 75,000-paper math behind it.
View sourcearXiv / Patterns
GPT detectors are biased against non-native English writers
Liang et al., 2023: the 61% average misread rate on TOEFL essays.
View sourceMore context: The Decoder on Pangram 4's claimed error rate, Curtin University on switching detection off, and the Anthropic access form for organizations that qualify.
WATCH
Turnitin explains its own false positives
A three-and-a-half-minute clip from Turnitin on what counts as a false positive and how its model handles them.
FAQ
Turnitin and Claude FAQ.
Turnitin can flag Claude text with its own AI writing classifier, which guesses from style. It can't read Anthropic's watermark, because checking that needs Anthropic's key and a detection API that is in private preview.
Often, yes, by style. In Epoch AI's July 2026 test GPTZero missed 2 of 99 plain Claude passages and 8 of 99 written to imitate a style. It does not read the Claude watermark.
Nothing public says so. Anthropic says third-party detectors look for stylistic tells and cannot verify the watermark. Watermark checks run through Anthropic's own API, open only to eligible organizations for now.
Turnitin hasn't published Claude-specific rates. Its own 2023 figures are under 1% false positives per document with 20% or more AI writing, and about 4% at the sentence level.
Ask which sentences were flagged, share drafts and version history, and check your course's AI policy. A Turnitin score is a prediction about style, not proof of how you wrote.
It might, since classifiers read style, but no detector result is promised. Anthropic says light editing probably won't remove its watermark, and only a full rewrite that replaces every word will.