The short answer
For images and video, look for Content Credentials (C2PA) and the IPTC “made by AI” label inside the file. For websites, look for the fingerprints AI site builders leave in the page’s code. For text, there is no watermark anyone outside the model maker can check.
“AI watermark” covers several different things, and they are not equally checkable. Some are records written into a file that anyone can read. Some are signals hidden in the pixels that only the company that made them can detect. And for text, despite what a lot of tools claim, there is nothing reliable to look for at all. Knowing which is which is most of the job.
Check a site, image or text
Enter an address or drop in a file. Every result names the evidence it found.
Images and video
This is where AI marking is real and standardised. Since 2 August 2026, the EU AI Act’s Article 50 has required providers of generative AI to mark synthetic output in a machine-readable way — generators already on the market have until 2 December 2026 to comply — and California’s AI Transparency Act has required providers with over a million monthly users to embed a hidden disclosure in the images, video and audio they generate. In practice there are two marks to look for.
Content Credentials (C2PA)
C2PA is an open standard for a signed record embedded in a file: which app created it, what was done to it, and who signed the claim. OpenAI, Adobe, Google and Microsoft attach one to images their tools generate, and a growing number of cameras sign photos the same way. It is the strongest mark there is, because it is signed rather than merely written.
The IPTC digital source type
The IPTC digital source type is a short vocabulary for how an image came to exist. The value that matters here is trainedAlgorithmicMedia — made by a trained model. Its neighbour, compositeWithTrainedAlgorithmicMedia, means a real image with AI-generated parts, the typical result of generative fill. It can live inside a C2PA record or in the file’s plain XMP metadata.
Generator settings
Tools you run yourself often leave something more candid. Stable Diffusion’s popular web interfaces write the entire prompt, sampler and seed into a PNG text field called parameters; ComfyUI stores its whole node graph. Not a formal declaration, but hard to argue with.
How to check
- 01Get the original fileDownload the image itself rather than a screenshot. A screenshot is a new file and carries none of the original’s records.
- 02Read its recordsDrop it into our AI watermark checker, or paste the image’s address. It reads C2PA, the IPTC label and generator settings, and shows you each one it found.
- 03Verify the signature if it mattersWe read what a C2PA record says but do not check its cryptographic signature. For a verified reading, use the Content Authenticity Initiative’s checker at contentcredentials.org/verify.
Pixel watermarks like SynthID
The other kind of image watermark is woven into the pixels themselves. Google’s SynthID is the best known: an imperceptible pattern in images, video and audio from Google’s models, designed to survive cropping, compression and filters.
Websites
A website carries no single AI watermark. What it can carry is evidence of how it was built, and there are three places to look.
The builder
AI site builders leave fingerprints. Sites made with Lovable load its editor script and often show an “Edit with Lovable” badge; v0 sets a generator meta tag; Bolt, Replit, Base44 and Gamma publish to their own domains. You can see these yourself: open the page’s source (Ctrl+U, or ⌘⌥U on a Mac) and search for generator.
Be careful with conventional builders. Framer, Wix and Webflow have all added AI generation, but a site built on them may not have used it, and the code does not say. And code written with an AI assistant, then deployed by hand, looks exactly like code written by a person.
The AI website checker looks for all of these at once, and checks every image on the page while it is there.
The images
Every image on a page is its own file and can carry its own Content Credentials. This is often the most concrete finding on a site — when the records have survived, which is less often than you would hope (see below).
What the page says
Some sites simply say so — “this article was written with the help of AI”. Worth reading the page for, and our checker quotes any such sentence it finds.
Text
There is no AI watermark in text that you can check. Any tool that gives you a percentage is guessing from writing style.
The EU’s marking rule covers text as well as media, but no standard text watermark exists yet that outsiders can check. Claude does not watermark its text. Google has built a text watermark into SynthID, but only Google can read it. The “hidden watermark” many checkers look for — zero-width spaces and odd non-breaking spaces — is real, but those characters come from copying text out of chat apps, and just as often out of Word, Google Docs or a PDF. They tell you text was pasted, not who wrote it.
Style-based AI detectors have a poor record. OpenAI withdrew its own AI text classifier in 2023, citing its low accuracy: at launch it correctly identified only 26% of AI-written text. A Stanford study found that popular detectors misclassified essays by non-native English writers as AI-generated more than half the time.
The one near-certain tell is a pasted chatbot reply — “as an AI language model”, “as of my last knowledge update”. Those are worth noticing, and our checker flags them separately.
Why finding nothing proves nothing
The records inside a file are fragile. Almost every social network, messaging app and content delivery network re-compresses images on upload, and re-compression writes a new file without them. So does a screenshot, a crop in many editors, or a “save as”. Most AI images on the web have lost their marks long before you see them.
That makes the test lopsided. A C2PA record or IPTC label that says “made by AI” is evidence. The absence of one is only the absence of evidence — it is consistent with a photograph and with a generated image that has been re-saved once.
Check one now
IN SHORT
- Images and video: look for C2PA Content Credentials and the IPTC “trainedAlgorithmicMedia” label.
- Pixel watermarks like SynthID can only be read by their owner.
- Websites: look for AI site-builder fingerprints, and check the images on the page.
- Text: no checkable watermark exists. Detectors guess, and often guess wrong.
- A positive result is evidence. A clean one is not.
The AI watermark checker does all of the above in one place. Enter a website, drop in an image or video, or paste text — image and text checks run in your browser and are never uploaded.
Is it AI watermarked?
Check a website, image, video or text. Free, no sign-up.