HOW IT WORKS
A nudge on word choice
SynthID Text works as a logits processor: while Gemini picks each next word, a secret key slightly shifts the odds. Across a long answer that shift adds up to a pattern a detector with the same key can measure.
GEMINI TEXT, OCTOBER 2026
Gemini AI detector: who can check Gemini text in 2026
Gemini marks its text with an invisible watermark, and Google opened a public detector this month. The catch is that the public detector does not take text. Here is who can tell Gemini wrote something, how they do it and how often they get it wrong.
Scan my text free Updated
gemini.google.com
Today
Write a short intro paragraph for my lab report on plant growth.
Message Gemini
MacBook Pro
9:41
ClaudeCleanLocal scanPASTED FROM GEMINI
SCANNING
…hidden characters
U+200Bzero-width space1
U+00A0no-break space1
U+202Fnarrow no-break space1
3 characters removed. Gemini's SynthID text mark sits in word choice, and only Google can check it.
9:41
ClaudeCleanLocal scanPASTED FROM GEMINI
SCANNING
…hidden characters
U+200Bzero-width space1
U+00A0no-break space1
U+202Fnarrow no-break space1
3 characters removed. Gemini's SynthID text mark sits in word choice, and only Google can check it.
Short answer
Text from the Gemini app and web carries Google's SynthID watermark, but only Google can read it. The public SynthID Detector and the Gemini app's own check handle images, video and audio, not text. Classifiers like Pangram, GPTZero and Originality.ai caught every plain Gemini 3.1 Pro text in a July 2026 test, missed 14% to 20% when it imitated a human style, and can flag human writers.
THE MARK INSIDE GEMINI
Does Gemini watermark its text?
Yes, in the Gemini app and on the web. The mark is a pattern in which words get picked, so you cannot see it, copy it out or scan for it.
WHO HOLDS THE KEY
Google, and nobody else
The open-source SynthID Text lets developers mark and detect text with their own keys. It cannot check Gemini output, because Gemini's key and settings are not public.
WHERE IT IS WEAK
Short and factual answers
Google's launch post says the mark works best on longer, creative responses. A factual answer has few word choices to nudge, so it carries less signal.
WHAT IT IS NOT
Not hidden characters
Nothing is inserted into the text. Zero-width spaces and odd non-breaking spaces you find in a paste are a separate issue, often from editors and copy paths.
Detector confidence scores can be greatly reduced when an AI-generated text is thoroughly rewritten, or translated to another language.
ONE GEMINI ANSWER, FOUR CHECKERS
Who can check what Gemini made?
9:41
Gemini
Explain photosynthesis in two sentences for my lab report.
Photosynthesis turns light, water and carbon dioxide into sugar and oxygen. Plants store that sugar as energy, which is why seedlings in brighter light usually grow faster.
SynthID mark applied as it was generated
Ask Gemini
Who can check this reply?
- Google, with its private keyCan check
- SynthID Detector (public)Can't check
- Gemini app, “Is this AI-generated?”Can't check
- Turnitin, GPTZero, Pangram and co.Style guess
- Google, with its private keyCan checkThe only party that can test Gemini text for SynthID today.
- SynthID Detector (public)Can't checkOpened to everyone on Oct 7, 2026 for images, video and audio. Text is not on the list.
- Gemini app, “Is this AI-generated?”Can't checkChecks uploaded images, video and audio. It does not check pasted text.
- Turnitin, GPTZero, Pangram and co.Style guessThey judge writing style. None of them reads the SynthID mark.
Text: only Google can read the mark. Everyone else is left with style classifiers, which guess from how the writing reads.
May 14, 2024
SynthID comes to Gemini text
Google DeepMind extended SynthID to watermark AI-generated text in the Gemini app and web experience, with video for Veo.Oct 23, 2024
Nature paper and open source
The Nature paper described a live test on nearly 20 million Gemini responses with no significant quality difference, and Google open-sourced SynthID Text for other developers.May 20, 2025
SynthID Detector portal, waitlist only
Google announced a portal that could check images, audio, video or text made with Google AI, starting with early testers from journalism, media and research.Nov 20, 2025
The Gemini app learns to check images
Users could upload an image and ask if it was made with Google AI. Google said over 20 billion pieces of content had been watermarked by then.Aug 19, 2026
A forum answer on the Gemini API
On Google's developer forum, a reply said Gemini API and AI Studio text is SynthID-watermarked too, reversing an earlier answer in the same thread. No official doc confirms it yet.Oct 7, 2026
SynthID Detector opens to everyone
The public launch covers images, video and audio, in English, worldwide. Text is not on the list.
For how Gemini's approach compares with OpenAI's and Anthropic's text watermarks, see the 2026 watermark comparison. The deeper mechanics of the technology live in our SynthID explainer.
WHO CAN CHECK
Who can check Gemini text today?
Google built the most reliable Gemini detector there is and kept the text part for itself. Everything you can run on a Gemini essay is a style classifier.
THE CHECKER MATRIX
Which Gemini detector reads what?
CheckerMethodTextImage, video, audioPublic
Google's internal SynthID text checkHolds the Gemini key. No public text endpoint.Watermark
Text: YesMedia: YesPublic: No
SynthID Detector (synthid.com)Public since Oct 7, 2026. Images, video and audio, English only.Watermark
Text: NoMedia: YesPublic: Yes
Gemini app: “Is this AI-generated?”Signed-in users. Recognizes content from Google AI tools only.Watermark
Text: NoMedia: YesPublic: Yes
Open-source SynthID TextDetects text marked with your own key, not Gemini's.Watermark
Text: PartlyMedia: NoPublic: Yes
TurnitinThrough schools only. 300-word minimum.Classifier
Text: YesMedia: NoPublic: Partly
GPTZeroStyle-based score, any model.Classifier
Text: YesMedia: NoPublic: Yes
PangramStyle-based score, any model.Classifier
Text: YesMedia: NoPublic: Yes
Originality.aiStyle-based score, any model.Classifier
Text: YesMedia: NoPublic: Yes
CopyleaksStyle-based score, any model.Classifier
Text: YesMedia: NoPublic: Yes
Nine tools, two very different methods. Watermark checks read a hidden key; classifiers judge how the writing reads.
What does the public SynthID Detector check?
Images, video and audio made or edited with Google AI, per Google's October 7 announcement. You sign in with a Google, Apple or ChatGPT account, heise reports, and the interface is English only. Search Engine Journal stresses it is "not a general AI detector".
What does the Gemini app check?
Upload a file and ask "Is this AI-generated?". Google's help page covers images, video up to 90 seconds and audio up to an hour, about 10 checks of each per rolling 24 hours, and says Gemini can only recognize content from Google AI tools. Pasted text is not part of it.
Does a 'no watermark' result clear a file?
No. Google's help page says a missing SynthID mark does not rule out other AI tools, and The Register found the app unreliable on images from ChatGPT.
Is Gemini API text watermarked?
Unclear on paper. Google's product pages name the Gemini app and web. A developer forum reply in August 2026 said API and AI Studio text is marked as well, but it did not say since when, and Vertex AI was not addressed.
Why sites that promise to read Gemini's watermark are guessing
The only key that can test Gemini text belongs to Google, and Google has not opened a text check to the public. A site that reports a "SynthID score" on your Gemini essay is running a classifier and labeling it something else.
BEFORE YOU PASTE IT ANYWHERE
Check your Gemini draft for hidden characters.
ClaudeClean scans any pasted text for zero-width spaces, odd non-breaking spaces and soft hyphens, and removes them in your browser. It does not read SynthID and does not score AI writing.
- 300 rewrite credits a month
- 20,000 characters per request
- .txt, .docx and .pdf files
THE NUMBERS
How accurate are AI detectors on Gemini text?
Independent tests first, vendor tests second. Plain Gemini output is easy to catch. Gemini told to sound like a person is harder.
99 of 99
plain-prompt Gemini 3.1 Pro texts caught by each of Pangram, GPTZero and Originality.ai
0.02
highest false negative rate for Pangram across models including Gemini 2.0 Flash
3.84%
of 495 human texts flagged as AI by Originality.ai, against 0 for Pangram and GPTZero
INDEPENDENT TEST, JULY 2026
How often detectors miss Gemini text
Share of 99 Gemini 3.1 Pro texts missed
Pangram 3.3.20.0% (20 of 99)
GPTZero0.0% (14 of 99)
Originality.ai Turbo 3.0.20.0% (14 of 99)
Told to write like a specific human, Gemini slipped past each detector 14 to 20 times out of 99.
CLASSIFIER VS WATERMARK
Same draft, two very different verdicts
One draft, two kinds of check
Plain Gemini answer
Urban green spaces offer measurable benefits for public health. Studies link access to parks with lower stress, more physical activity and stronger community ties...
Style classifier
Turnitin, GPTZero, Pangram, Originality.ai
Likely AIIn Epoch AI's July 2026 test, Pangram, GPTZero and Originality.ai caught all 99 plain Gemini 3.1 Pro texts.
SynthID text check Google only
Reads the watermark with Google's key
WatermarkedGemini app and web text carries SynthID. Longer answers hold the strongest signal.
MacBook Pro
Classifier: Likely AI. In Epoch AI's July 2026 test, Pangram, GPTZero and Originality.ai caught all 99 plain Gemini 3.1 Pro texts.
Watermark: Watermarked. Gemini app and web text carries SynthID. Longer answers hold the strongest signal.
UNIVERSITY OF CHICAGO, 2025
A clear first tier
Brian Jabarian and Alex Imas tested 1,992 human passages across six genres against GPT-4.1, Claude and Gemini 2.0 Flash. Pangram led; GPTZero and Originality.ai formed a second tier that struggled more on short texts and humanized output.
PANGRAM, VENDOR TEST
99.7% to 99.8% on new Gemini
Pangram reports its version 4 flagged 998 of 1,000 Gemini 3.8 Flash texts and 997 of 1,000 Gemini 3.1 Pro texts. Most misses were poems. It is the vendor grading itself.
ORIGINALITY.AI, VENDOR TEST
96.4% to 98.4% on Gemini 1.5 Pro
Originality.ai tested 1,000 Gemini 1.5 Pro samples with two of its models. The set had no human texts, so it says nothing about false positives.
GPTZero and Copyleaks say their models cover Gemini, but we found no independent Gemini-only results for Copyleaks. Treat any single percentage, including the ones above, as true for one model version on one kind of text.
TURNITIN
Can Turnitin detect Gemini?
It can flag Gemini-style writing like any other AI text. It cannot read the watermark, and it has rules about what it will score.
METHOD
A classifier, not SynthID
Turnitin's AI writing detection looks for patterns common in machine text. Its FAQ lists Gemini among the model families it trains on, but no key from Google is involved.
LENGTH
300 words of prose first
Turnitin needs at least 300 words of long-form prose before it scores. Bullet lists, tables, code and poetry are not what it was built for.
LOW SCORES
An asterisk under 20%
Turnitin reported under 1% false positives per document only for papers with 20% or more AI writing, and about 4% per sentence. Below 20% the report shows an asterisk instead of a number.
Turnitin has not published Gemini-specific accuracy. For the full breakdown of how it scores, how the asterisk works and what to do after a flag, see can Turnitin detect Claude; the mechanics are the same whichever chatbot wrote the draft.
EDITS
What survives when Gemini text gets edited?
The watermark and the writing style are two different signals, and they fade at different speeds. Pick an edit to compare them.
WHAT SURVIVES AN EDIT
Slide the edit, watch both signals
Style fingerprint
Can a classifier still tell which chatbot wrote it?
97.8%
SynthID watermark
Does the key holder still detect the mark?
87.6%
Original: SynthID on Gemma-2B-IT before any edit, at a 5% false positive rate.
The watermark holds up to
- A few changed words. Google's launch post says SynthID survives light edits and mild paraphrasing.
- Cropping out a piece. Google lists cropping pieces of text among the light changes the mark tolerates, as long as enough text is left.
- Copy and paste. The mark is in the words, so moving the text between apps changes nothing.
The watermark weakens with
- A thorough rewrite. Google says detector confidence can be greatly reduced.
- Translation. New language, new word choices; Google names this case directly.
- A model paraphrase. In the DAMAGE study, SynthID detection on Gemma fell from 87.6% to 5.4%.
More published robustness results, including the SynthID Nature paper, are collected in does paraphrasing remove a watermark.
GEMINI'S HABITS
How to tell if text was written by Gemini
Only one peer-reviewed study we found pins down Gemini-specific habits. It is enough to explain why classifiers do well, and not enough to accuse anyone by eye.
ICML 2025
Chatbots have fingerprints
Sun, Yin, Xu, Kolter and Liu trained a classifier to tell ChatGPT, Claude, Grok, Gemini and DeepSeek apart. It was right 97.1% of the time, against 20% for a random guess.
FORMATTING
More italics
In the same paper, Gemini used much more italic text than the other chat models. Markdown habits like this vanish the moment text is pasted as plain text.
OPENINGS
Short, direct openings
The authors found Gemini's outputs more concise, with direct openings. That describes a tendency across thousands of answers, not a rule for any single paragraph.
There is no reliable Gemini word list
We found no sourced list of words that give Gemini away, the way some lists exist for ChatGPT. The ICML study also found that held-out Gemini, Claude and Grok outputs tended to get classified as ChatGPT, which shows how close the styles sit. For general tells that apply to every model, see signs of AI writing.
FALSE POSITIVES
Can a Gemini detector flag human writing?
Yes. Style classifiers score how predictable writing is, and some people write predictably for good reasons.
61.3%
average share of TOEFL essays by non-native writers that seven detectors called AI
30-69%
of human passages mislabeled by an open-source RoBERTa baseline in the Chicago study
WHO GETS HIT
Non-native and formulaic writers
The Stanford team found almost no false flags on essays by US eighth graders, and a majority on TOEFL essays. Simple vocabulary and textbook structure read as machine-like to a classifier.
WHAT TO DO
Bring the process, not a counter-score
Drafts, version history, notes and a short conversation about the work carry more weight than a second detector. Our guide on how to prove you didn't use AI walks through the evidence that helps.
STEP BY STEP
How to check text you think came from Gemini
For teachers, editors and anyone reviewing work. The order matters: context first, scores last.
- 01
Ask before you scan
Ask the writer how the text was made and whether AI helped. Drafts, notes and version history answer more than any score, and many policies allow some AI use. - 02
Check length and type
Classifiers and watermarks both need long prose. A 120-word answer, a bullet list, code or a poem gives weak results either way, so do not lean on a score for those. - 03
Run two different classifiers
Pangram, GPTZero and Originality.ai disagree on edge cases. If only one of two flags a text, treat that as a reason to look closer, not a verdict. - 04
Read the flagged sentences
Most tools highlight the passages that drove the score. Compare them with the writer's earlier work, usual mistakes and voice. - 05
Use SynthID for images, video and audio
If the work includes a Gemini image or clip, upload it to SynthID Detector or the Gemini app. For text there is no public SynthID check, and no website can stand in for one. - 06
Decide with context
Weigh the drafts, the conversation and the written policy. A detector score alone should not settle an accusation, and Turnitin itself says its AI score should not be the sole basis for action against a student.
WHAT CLAUDECLEAN DOES
Paste Gemini text, clean the characters, rewrite only if you choose
ClaudeClean works on text from any chatbot, Gemini included. It cleans and rewrites; it does not judge who wrote what.
01
> source text
Paste or drop a file
Paste a Gemini answer or upload a .txt, .docx or .pdf. Up to 2,000 characters works without an account.
02
scan report
3 zero-width1 soft hyphencleanable signals2Scan locally
Zero-width characters, regular and narrow no-break spaces, soft hyphens and lookalike letters are flagged and removed in your browser.
03
before → after
clearusefulanswerfocusedhelpfulresponse
Copied ✓Optional rewrite
Light, Medium or Strong passes reword typical AI habits and show a before and after diff. Rewriting sends text to OpenAI and is best effort.
PRICING
Clean the text free. Rewrite more with Pro.
The scan and hidden-character cleanup are free. Pro is $5 a month for more rewrite room and files. None of it reads or scores the SynthID watermark.
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.
HONEST LIMITS
What a hidden character scan can and can't tell you
People land here hoping for a Gemini detector. Our scan is something narrower, and it helps to know exactly where the line sits.
The scan can
- Find invisible characters. Zero-width spaces, joiners, byte-order marks, soft hyphens and non-breaking spaces show up with their code points.
- Remove them locally. Scan and clean run in your browser. The text is not uploaded for that step.
- Work on any source. Gemini, ChatGPT, Claude or a human draft pasted through Google Docs: characters are characters.
- Reword on request. An optional rewrite changes word choice, with a diff you review before using it.
The scan can't
- Read SynthID. The mark is a word-choice pattern keyed by Google. No character scan can see it.
- Say whether Gemini wrote something. We do not run a classifier and do not score AI likelihood.
- Promise a detector outcome. No detector result is promised, for Turnitin, GPTZero or anyone else.
- Change your disclosure duty. If a course, journal or client asks you to declare AI help, cleaned text does not change that.
If your question is about the visible sparkle logo on Gemini images, that is a different mark with a different fix, covered in the Gemini watermark remover guide. If you just want the stray characters out of a draft, paste it into the cleaner.
TERMS
Gemini detection terms, in plain words
The words detector sites use loosely, pinned down.
- SynthID
- Google DeepMind's watermarking family. For text it nudges word choices with a secret key; for images, video and audio it hides a signal in the content itself.
- Classifier
- A model trained on human and AI writing that estimates how likely a text is to be machine-written. Turnitin, GPTZero, Pangram, Originality.ai and Copyleaks are classifiers.
- False positive
- Human writing labeled AI. This is the error that hurts students and freelancers, and the one vendors most want to keep low.
- False negative
- AI writing labeled human. Epoch AI's 2026 test measured these at 14% to 20% for Gemini 3.1 Pro text imitating a human style.
- Uncertain
- The third outcome of Google's SynthID text detector, next to watermarked and not watermarked, for text where the signal is too weak to call.
- Content Credentials (C2PA)
- Signed metadata describing how a file was made. Google adds it to Nano Banana Pro images alongside SynthID. It applies to media files, not to pasted text.
WATCH
Google DeepMind on how SynthID marks text
The two-minute video Google released the day text watermarking came to Gemini.
SOURCES
Where these facts come from
Google's own pages first, then independent tests, then vendor claims. Checked October 10, 2026.
Google DeepMind
Watermarking AI-generated text and video with SynthID
May 2024 post adding text watermarking to the Gemini app and web, with the limits on short, factual and rewritten text.
View sourceGoogle AI for Developers
SynthID Text documentation
How the logits processor works, the three detector outcomes and the rewrite and translation caveat quoted above.
View sourceNature
Scalable watermarking for identifying large language model outputs
Dathathri et al., October 2024. Live test on nearly 20 million Gemini responses.
View sourceSynthID Detector opens to everyone
October 7, 2026 launch: images, video and audio, English only, partners including OpenAI and NVIDIA.
View sourceSynthID Detector portal (2025)
May 2025 announcement of the waitlist portal that listed text among supported formats.
View sourceGemini Apps Help
Check if content was made with Google AI
Supported file types, size and length limits, daily check limits, Google-AI-only recognition.
View sourceGoogle AI Developers Forum
Does Gemini API text output carry SynthID watermarking?
August 2026 thread where a reply says API and AI Studio text is watermarked. Not official documentation.
View sourceEpoch AI
AI detectors and false negatives
July 15, 2026 test of Pangram, GPTZero and Originality.ai, including Gemini 3.1 Pro and 495 human texts.
View sourceNBER
Jabarian and Imas, Working Paper 34223
University of Chicago comparison of commercial detectors on 1,992 human passages and four models.
View sourcearXiv
Idiosyncrasies in Large Language Models
ICML 2025. Five-chatbot classifier at 97.1%, Gemini's italics and direct openings, results after rewriting.
View sourcearXiv
DAMAGE: Detecting adversarially modified AI generated text
Masrour, Emi and Spero, 2025. SynthID detection on Gemma before and after paraphrase.
View sourcearXiv
GPT detectors are biased against non-native English writers
Liang et al., 2023. Seven detectors, 61.3% average false positives on TOEFL essays.
View sourceTurnitin
Understanding the false positive rate for sentences
2023 post with the under 1% document and about 4% sentence figures, the 20% asterisk and the 300-word floor.
View sourcePangram
Can Pangram detect Gemini 3.8 Flash?
Vendor test, September 2026: 998 of 1,000 and 997 of 1,000 Gemini texts flagged.
View sourceOriginality.ai
Is Google Gemini Pro content detectable?
Vendor test on 1,000 Gemini 1.5 Pro samples, no human control set.
View sourceSearch Engine Journal
Google opens SynthID watermark detector to public
Coverage of the October 2026 launch, including Google's note that it is not a general AI detector.
View sourceFAQ
Gemini AI detector FAQ.
Two kinds. Google can check Gemini text for its SynthID watermark, but it has no public tool for that. Anyone can run style classifiers such as GPTZero, Pangram or Originality.ai, which guess from how the writing reads and work on any model.
Yes. Google DeepMind added SynthID text watermarking to the Gemini app and web experience in May 2024. The mark is a statistical pattern in word choice, not hidden characters, and it is weaker on short or factual answers.
Not the public version. When Google opened SynthID Detector to everyone on October 7, 2026, it covered images, video and audio. The 2025 waitlist version listed text, but that access was limited to early testers such as journalists and researchers.
Turnitin runs its own classifier and does not read Google's watermark. It needs at least 300 words of prose, shows an asterisk for scores under 20%, and reports under 1% false positives per document only when 20% or more of a paper is AI writing.
On plain prompts, very. Epoch AI's July 2026 test found Pangram, GPTZero and Originality.ai caught all 99 Gemini 3.1 Pro texts. When Gemini was told to imitate a human style, they missed 14% to 20% of them.
Yes. A Stanford study found seven detectors called 61.3% of TOEFL essays by non-native English writers AI on average. In Epoch AI's 2026 test, Originality.ai flagged 19 of 495 human texts, while Pangram and GPTZero flagged none.
Heavy rewriting weakens it. Google says detector confidence can be greatly reduced when text is thoroughly rewritten or translated. In the DAMAGE study, a paraphrase dropped SynthID detection on a Gemma model from 87.6% to 5.4%.
No. It is not an AI detector or a watermark detector. It scans pasted text for hidden Unicode characters in your browser and removes them, and an optional rewrite sends text to OpenAI. No detector result is promised.