WHAT IT READS
Images, video and audio
Google's announcement says you can check whether an image, video or audio file was made with AI from Google or its partners. eWeek and IT-Connect both note text is not on the list.
GOOGLE DEEPMIND, OCTOBER 2026
SynthID: what it is, how it works and the detector
Google opened its SynthID Detector to the public on October 7, 2026. It reads images, video and audio from four AI companies. It does not read text, even though Gemini has watermarked its answers since 2024. Here is the whole picture, with sources.
Scan my text free Updated
gemini.google.com
Today
Write a short caption for our spring sale post.
Message Gemini
MacBook Pro
9:41
ClaudeCleanLocal scanPASTED FROM GEMINI
SCANNING
…hidden characters
U+200Bzero-width space1
U+00A0no-break space1
U+00ADsoft hyphen1
Characters cleaned. SynthID lives in word choice, so this scan cannot see it.
9:41
ClaudeCleanLocal scanPASTED FROM GEMINI
SCANNING
…hidden characters
U+200Bzero-width space1
U+00A0no-break space1
U+00ADsoft hyphen1
Characters cleaned. SynthID lives in word choice, so this scan cannot see it.
Short answer
SynthID is Google DeepMind's invisible watermark for AI-generated images, video, audio and text. Media gets a signal woven into pixels or sound; text gets a statistical pattern in word choice, with no hidden characters. Since October 7, 2026, anyone can check images, video and audio at synthid.com. Text cannot be checked there.
NEW THIS WEEK
SynthID Detector is public: what does it check?
The portal Google previewed in May 2025 for a waitlist is now open worldwide in English. It answers one narrow question about one file at a time.
SYNTHID.COM, OCTOBER 2026
Upload a file, get one answer
synthid.com
SynthID DetectorSigned in
Check a file for a SynthID watermark
Works on media from companies that use SynthID: Google, OpenAI, NVIDIA, Kakao.
Drop an image, video or audio fileOne file at a time
No file chosenWaiting for upload
Result appears here
Not a general AI detectorMacBook Pro
WHOSE AI
Google, OpenAI, NVIDIA, Kakao
Google names OpenAI, NVIDIA and Kakao as partners whose media it can now verify, with Apple to follow. A file from any other generator gets no reading at all.
ACCESS
Sign-in and a daily cap
IT-Connect reports sign-in with a Google, Apple or OpenAI account. Launch coverage also describes a daily limit on checks.
THE ANSWER
Detected or not, nothing more
Per Search Engine Journal, a hit says SynthID was detected and the media was made or edited with AI, and may have been edited further since.
This is not a general AI detector. We can only detect media from the above companies, who have adopted SynthID technology.
Scale is the reason this matters. Google says SynthID has marked over 180 billion images and videos plus 240,000 years of audio, and that its checks in Search, the Gemini app and Chrome already handle over a million requests a day. If you want to know whether a chat answer came from Gemini, that is a different job, covered in our Gemini AI detector guide.
THE BASICS
What is SynthID?
One brand name for four watermarking methods, built by Google DeepMind and switched on across Google's generative products.
IMAGES
Woven into the pixels
DeepMind describes a watermark embedded directly into the pixels, invisible to the eye, and built to survive cropping, filters and lossy compression.
AUDIO
Hidden in the spectrogram
Audio is converted to a spectrogram, marked, and converted back. Google says you cannot hear it and that it holds through MP3 compression.
VIDEO
Every frame marked
For Veo, SynthID marks every frame of a generated video, building on the image method, so the signal is spread across the whole clip.
TEXT
A pattern in word choice
For text, SynthID adjusts the probability of the words the model picks. Nothing is inserted, so there is no character to find or delete.
~20M
Gemini responses in the live test of SynthID Text, with no significant drop in ratings
SynthID is one vendor's system. For how it compares with other approaches to marking AI text, see our AI text watermarking overview, and for the vendor-by-vendor picture, the ChatGPT vs Claude watermark comparison.
THE PART YOU CAN SEE
SynthID hides in word choice. Stray characters hide in plain sight.
ClaudeClean cannot read SynthID. It finds the zero-width spaces, no-break spaces and soft hyphens that ride along when you copy AI text, and removes them in your browser.
- 300 rewrite credits a month
- 20,000 characters per request
- .txt, .docx and .pdf files
FOUR METHODS
How does SynthID mark images, audio, video and text?
Each content type gets a different technique, a different launch date and a different answer to the question of who can check it.
FOUR KINDS OF SYNTHID
Same name, four different tricks
Images
First shipped: August 2023, Imagen on Vertex AI (beta)
- How it is marked
- A trained model embeds the signal directly into the pixels. A second model reads it back.
- Built to survive
- Cropping, filters, color changes and lossy compression such as JPEG, per DeepMind.
- Weakened by
- Heavy edits and dedicated attack research (see the robustness section).
- Who can check it
- Anyone at synthid.com, plus the Gemini app, Search and Chrome.
| Content type | How it is marked | First shipped | Public check |
|---|---|---|---|
| Images | Signal embedded in the pixels by a trained model | Aug 2023 (Imagen, Vertex AI beta) | synthid.com, Gemini app |
| Audio | Mark added to a spectrogram, inaudible | Nov 2023 (Lyria) | synthid.com, Gemini app |
| Video | Every frame marked | May 2024 (Veo) | synthid.com, Gemini app (under 90 s) |
| Text | Word choice nudged by tournament sampling | May 2024 (Gemini app and web) | No public checker for Gemini text |
Swipe to compare →
Sources: Google DeepMind's SynthID page and launch posts, Google's October 2026 announcement and the Gemini help page.
Not the same thing as the visible Gemini logo
Gemini images can also carry a small visible sparkle mark in the corner. That is a separate, visible label, and cropping it off does not touch SynthID. We cover the visible one in the Gemini watermark remover guide.
UNDER THE HOOD
How does SynthID Text work?
The text method is the most interesting one, because it is the only part Google has open-sourced and the part other labs have borrowed.
SYNTHID TEXT, ONE WORD AT A TIME
How a tournament picks the next word
PROMPT SO FAR
Tomorrow the weather will be ____
STEP 1. THE MODEL SUGGESTS 8 CANDIDATES
sunnywarmmildbrightclearlovelycoolcalm
ROUND 1. COMPARE g-VALUE LAYER 1
sunnyg=0warmg=1
mildg=1brightg=0
clearg=1lovelyg=1
coolg=0calmg=1
ROUND 2. COMPARE g-VALUE LAYER 2
warmg=1mildg=0
clearg=0calmg=1
ROUND 3. COMPARE g-VALUE LAYER 3
warmg=0calmg=1
“calm” is printed. It won all three rounds, so it carries g=1 on every layer. A detector with the key reruns the g-values over the whole text: marked text scores high more often than the coin-flip average.
The Nature paper by Dathathri and colleagues calls the method tournament sampling. The model proposes candidate tokens as usual, a keyed function scores each one, and the candidates play knockout rounds until one remains. Only the sampling step changes, so the model itself is untouched and detection does not need the model at all. Google's developer docs describe it as a logits processor applied after Top-K and Top-P.
- Watermarking key
- A private list of random integers. Hugging Face suggests 20 to 30 of them. Whoever holds the key can generate and detect the mark; nobody else can.
- g-function
- A pseudorandom function that turns the key, the recent context and a candidate token into a score. Unmarked text averages out like coin flips; marked text scores high more often.
- Tournament layers
- The number of knockout rounds per token. The Nature paper uses 30 by default; more layers carry more signal per word.
- ngram_len
- How many preceding tokens feed the g-function. The Transformers default is 5, and it must be at least 2.
- Bayesian detector
- The open-source detector is trained per configuration, with 10,000 or more examples recommended, and returns watermarked, not watermarked or uncertain.
- Logits processor
- The hook in the generation loop where SynthID runs. It adjusts which token is sampled without retraining the model.
Anyone can try it: the code lives on GitHub, inside Hugging Face Transformers as SynthIDTextWatermarkingConfig, and in a public demo Space. Open-sourcing the method did not publish Gemini's key, which is why the code cannot check Gemini output.
WHO CAN CHECK WHAT
Can anyone check text for SynthID today?
Media checks are open to the public. Text checks are not, and that gap is where most confusion starts.
CHECKING IN THE GEMINI APP
An image gets a reading. Text does not.
9:41
Gemini
Was this made with Google AI?
Ask Gemini
- CoversImages, video and audio made or edited by Google AI
- NeedsA signed-in Gemini account, one file per check
- Size100 MB or less, video under 90 s, audio under 1 hour
- TextNot listed on the help page
A “no” is narrow. Google says a negative means the file was not made or edited by Google AI, and that other AI systems could still have made it.
Can I check an image, video or audio file?
Yes, two ways. Upload it at synthid.com, or ask in the Gemini app whether it was made with Google AI. The Gemini route only recognizes content from Google AI tools.
Can I check Gemini text?
We found no public tool that does. The 2025 portal announcement said text detection would roll out to early testers, but the public site in October 2026 takes images, video and audio, and the Gemini help page lists the same three.
Can I check text from my own model?
Yes, if you watermarked it yourself with the open-source SynthID Text and trained a detector for your key. Google's docs describe keeping that detector private, offering it through an API, or publishing it.
Can I check Claude text?
Not with any SynthID tool. Claude uses Anthropic's own key, and Anthropic's detection API is in private preview for eligible organizations. Our Claude watermark detector guide covers what exists.
- 01
Open synthid.com
Go to Google's public SynthID Detector site, open worldwide in English since October 7, 2026. - 02
Sign in
Use a Google, Apple or OpenAI account, as reported at launch. - 03
Upload one file
Pick an image, video or audio file. Text is not accepted by the public version. - 04
Read the result
A hit reads as SynthID detected, meaning a SynthID partner's AI made or edited the file. It may have been edited after that. - 05
Treat a miss as narrow
No signal does not mean human. The file may come from an AI that does not use SynthID or may have been edited heavily.
HOW WE GOT HERE
SynthID timeline, 2023 to 2026
Three years from an image beta to a public detector, with text open-sourced in the middle.
Aug 29, 2023
SynthID starts with images
Google DeepMind launches SynthID in beta for Imagen on Vertex AI, with one model to embed a watermark in the pixels and another to identify it.Nov 16, 2023
Audio joins with Lyria
Music from the Lyria model gets an inaudible watermark that survives noise, MP3 compression and tempo changes.May 14, 2024
Gemini text and Veo video
Google extends SynthID to text in the Gemini app and web, and to every frame of Veo video.Oct 23, 2024
Nature paper and open source
The SynthID Text method is published in Nature, and the code ships in Hugging Face Transformers v4.46.0 and Google's Responsible GenAI Toolkit.May 20, 2025
Detector portal for early testers
Google announces the SynthID Detector with a waitlist for journalists, media professionals and researchers, and says over 10 billion pieces of content carry the mark.Aug 14, 2026
Claude gets a SynthID-style watermark
Anthropic starts watermarking Claude text with a version of the SynthID-Text approach, using its own key and detector.Oct 7, 2026
The detector opens to everyone
Pushmeet Kohli announces public access at synthid.com for images, video and audio from Google and partners including OpenAI, NVIDIA and Kakao.
PUBLISHED NUMBERS
How robust is SynthID?
Google says no signal is completely robust. Independent labs have measured where SynthID Text holds and where it breaks.
87.6% to 5.4%
SynthID Text detection at 5% false positives, before and after a DIPPER paraphrase
WHAT EDITS DO TO SYNTHID TEXT
Rewording hurts it. Light swaps barely do.
Detection rate before and after a DIPPER paraphrase
True positive rate at a fixed false positive rate. Gemma-2B-IT with SynthID, 200-token answers.
5% FPR, untouched87.6%
5% FPR, paraphrased5.4%
1% FPR, untouched66.5%
1% FPR, paraphrased1.5%
TRANSLATION
Translation weakens it
Google's own launch post warned that translating text lowers detector confidence, and the Han numbers show how much.
LENGTH AND FACTS
Short and factual text is weak
The docs say the mark is less effective on factual responses, where there are fewer good word choices to nudge. Short answers carry less signal too.
IMAGES
Image attacks exist
The Register reported that the UnMarker attack from the University of Waterloo cut SynthID image detection from 100% to around 21% in later testing.
The Nature paper itself says generative watermarks are weakened by edits such as LLM paraphrasing and are exposed to stealing, spoofing and scrubbing attacks. For the deep dive on rewording, read does paraphrasing remove a Claude watermark; the same research applies because Claude's mark follows the same approach.
THE HONEST VERSION
Can you remove SynthID?
People search for this a lot. The answer depends on whether you mean a file or text, and in both cases there is no button that deletes it.
What published tests say weakens it
- A full paraphrase of text. The DIPPER rewrite took SynthID Text detection from 87.6% to 5.4% in the DAMAGE study.
- Translation round trips. F1 fell to between 0.711 and 0.819 depending on the language, per Han et al.
- Burying a short marked passage. Mixed into ten times as much other text, detection AUC dropped to 0.62.
- Targeted image attacks. Research attacks like UnMarker degraded image detection, though they need real compute and expertise.
What does not remove it
- Deleting characters. SynthID adds no hidden characters to text, so a character scanner finds nothing to delete.
- Light synonym swaps. With 30% of words swapped, F1 stayed at 0.987.
- Cropping or compressing media. Google built image, video and audio marks to survive crops, filters and lossy compression.
- Removing the visible logo. The corner sparkle on Gemini images is a separate label; SynthID stays in the pixels.
ClaudeClean cannot see or promise anything about SynthID
Our tool works on text only. It removes invisible characters locally and offers an optional rewrite that goes to OpenAI. It does not read any watermark, and no detector result is promised. If your school, publisher or employer asks you to disclose AI use, a rewritten draft does not change that.
SYNTHID AND CLAUDE
Is Claude's watermark SynthID?
It is built on the same published idea, with a different key and a different owner. Here is what Anthropic's own page says.
Claude’s text watermark is a version of the SynthID-Text approach published by Google DeepMind in a Nature paper in 2024.
SEPARATE KEY
Anthropic runs its own detector
Detection needs Anthropic's key. Its detection API is in private preview for eligible organizations, and Claude is not on the SynthID Detector partner list.
NO CHARACTERS
Nothing is added to the text
Anthropic says there are no hidden characters. Any zero-width spaces you find in Claude text came from somewhere else, as our Claude watermark guide explains.
EDITS
Light edits leave it, full rewrites remove it
Anthropic says light editing probably won't remove the watermark completely, while a complete rewrite where every word is replaced will.
WHAT CLAUDECLEAN DOES
Clean the characters you can see, decide on the rest
ClaudeClean works on any pasted text, from Gemini, Claude, ChatGPT or a person. It handles characters and style, never watermark detection.
01
> source text
Paste or upload
Paste text or drop a .txt, .docx or .pdf file. Up to 2,000 characters works without an account.
02
scan report
3 zero-width1 soft hyphencleanable signals2Scan in your browser
Zero-width spaces, no-break spaces, soft hyphens and lookalike letters are flagged and removed locally. The scan does not send your text anywhere.
03
before → after
clearusefulanswerfocusedhelpfulresponse
Copied ✓Rewrite if you want to
An optional Light, Medium or Strong rewrite reworks AI habits with a before and after diff. It sends text to OpenAI and is best effort.
PRICING
Character cleanup is free. Pro adds rewrites and files.
Free accounts get 10 credits a month. Pro is $5/month with 300 credits and 20,000 characters per request. No plan reads or removes SynthID.
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.
READ RESULTS CAREFULLY
Common mistakes with SynthID results
The detector answers a narrow question. These are the ways people stretch it too far.
MISTAKE 1
Reading no signal as human-made
Google's Gemini help page says a negative means the file was not made by Google AI, and that other AI systems could still have made it.
MISTAKE 2
Treating it as a general AI detector
The site itself says it only reads media from companies that adopted SynthID. A file from a generator that never adopted it will not register, whatever made it.
MISTAKE 3
Uploading an essay
The public portal does not take text, so a screenshot of an essay is only checked as an image. Text classifiers like Turnitin are a different technology entirely.
MISTAKE 4
Assuming a hit means fully AI
A hit means AI made or edited the media. Google's result wording notes the file may have been edited further, so a person may have shaped the final version.
WATCH
Google DeepMind on the science of SynthID
Posted a week before the public launch, with the team behind the method.
SOURCES
Where these facts come from
Google's own posts first, then the paper and code, then reporting and independent research.
Google expands SynthID Detector for AI content
Oct 7, 2026 post by Pushmeet Kohli: public access, partners OpenAI, NVIDIA and Kakao, scale figures.
View sourceGoogle DeepMind
SynthID
Overview of image, video, audio and text watermarking and what each is built to survive.
View sourceNature
Scalable watermarking for identifying large language model outputs
Oct 2024 paper behind SynthID Text: tournament sampling, the 20 million response Gemini test, limits.
View sourceHugging Face
Introducing SynthID Text
Transformers v4.46.0 integration, key and ngram settings, the trained Bayesian detector.
View sourceGoogle AI for Developers
SynthID: Tools for watermarking and detecting LLM-generated text
Logits processor design, detector outputs, deployment options and stated limitations.
View sourceGoogle DeepMind
Watermarking AI-generated text and video with SynthID
May 2024 rollout to the Gemini app and Veo, and the warning about rewrites and translation.
View sourceGoogle Gemini Help
Verify AI-generated images, videos, and audio
What the Gemini app check covers, size limits, and how to read a negative result.
View sourceSynthID Detector announcement, 2025
May 2025 preview portal for early testers, 10 billion items watermarked, NVIDIA and GetReal partnerships.
View sourceSearch Engine Journal
Google opens SynthID watermark detector to public
Quotes from the site: not a general AI detector, the wording of a positive result.
View sourceIT-Connect
Google opens its AI image, video and audio detector to everyone
Sign-in options and the absence of text files in the public version.
View sourcearXiv (Pangram Labs)
DAMAGE: Detecting Adversarially Modified AI Generated Text
SynthID Text detection before and after a DIPPER paraphrase at 5% and 1% false positive rates.
View sourcearXiv (Queen's University)
Robustness Assessment and Enhancement of Text Watermarking for Google's SynthID
Synonym substitution, copy-paste and back-translation results for the open-source SynthID Text.
View sourceThe Register
Image watermarks meet their Waterloo with UnMarker
University of Waterloo research and the later SynthID image test result.
View sourceAnthropic
How Claude marks AI-generated text
Claude's watermark as a version of SynthID-Text, no hidden characters, private detection API.
View sourceFAQ
SynthID FAQ.
SynthID is Google DeepMind's invisible watermark for AI-generated content. It marks images, video and audio at the signal level, and marks text by nudging which words the model picks. A matching detector looks for the pattern later.
Since October 7, 2026, anyone. Google opened synthid.com to the public in English worldwide. Launch coverage says you sign in with a Google, Apple or OpenAI account and that there is a daily cap on checks.
Not the public version. Google's announcement and launch coverage list images, video and audio only. The 2025 preview promised text for early testers, but the public site does not take text files, and the Gemini app check does not cover text either.
Yes. Google DeepMind said in May 2024 that SynthID text watermarking was added to the Gemini app and web experience. The mark changes word probabilities, so there is no hidden character to find in the text.
Not by deleting anything, because nothing is attached. Published tests show a thorough paraphrase cut SynthID Text detection from 87.6% to 5.4%, while swapping 30% of words for synonyms barely moved it. Google says no signal is completely robust, and no method is guaranteed.
Related, not the same. Anthropic says Claude's text watermark is a version of the SynthID-Text approach from Google DeepMind's 2024 Nature paper. Anthropic runs its own key and detector, which is in private preview, and Claude is not a SynthID Detector partner.
For images and audio, yes. OpenAI's help center says its images carry Content Credentials plus a SynthID watermark and its audio carries SynthID. ChatGPT text uses OpenAI's own textGrain watermark in the EU, not SynthID.
No. A negative only means no SynthID signal was found. The file could come from an AI that does not use SynthID, could be edited heavily, or could be text, which the public tools do not read. Google says the detector is not a general AI detector.