AI PROVENANCE
AI watermark checker
Enter a site to see whether its images carry an AI watermark, whether it was built with an AI site builder, and whether its text reads like a chatbot. Or drop in an image, or paste text. Every result names the evidence behind it, and says plainly where the evidence runs out.
We fetch the page from our server, read the images it shows for AI records, and look for the fingerprints AI site builders leave. The address is not stored or logged. Pages that sit behind a login cannot be checked.
- No sign-up
- Images and text checked in your browser
- Nothing stored
- Reads C2PA Content Credentials and IPTC AI labels
- Recognises Lovable, v0, Bolt, Framer, Wix and other builders
- Finds hidden characters in text, and removes them
A positive result is evidence. A clean one is not: most AI output loses its records the moment it is re-saved or uploaded. And no tool can reliably tell who wrote a piece of text.
HOW IT WORKS
What an AI watermark can prove
An AI watermark is a mark a generator leaves in what it makes so that anyone can later tell it was made by AI. Since 2 August 2026 the EU AI Act and California’s AI Transparency Act have required generators to add one. For images, video and audio the marks are standardised: C2PA Content Credentials and the IPTC “trained algorithmic media” label. For text, the EU rule applies too, but no standard mark exists yet that an outsider can check. So the honest answer to “is this AI watermarked?” depends on what you are looking at — and a clean result is only ever an absence of evidence.
| SIGNAL | WHERE IT LIVES | WHAT A HIT PROVES | SURVIVES RE-UPLOAD? |
|---|---|---|---|
| C2PA Content Credentials | Inside the image or video file | A signed record naming the app that made or edited it. The strongest mark there is. | Often not — most social networks and CDNs strip it |
| IPTC digital source type | XMP metadata in the file | The generator declared the file “trained algorithmic media” — made by AI. | Often not — lost on re-save |
| Generator settings | PNG text fields, EXIF Software | Made with that tool: Stable Diffusion, ComfyUI, Midjourney and others. | Lost on re-save |
| AI site-builder fingerprints | The page’s HTML and hosting | Built with Lovable, v0, Bolt and similar. Hand-written and AI-written code look identical. | Yes, while the site stays on that builder |
| Hidden characters | The text itself | The text was copied from somewhere. Not who wrote it. | Until someone cleans it |
| Writing-style patterns | The text itself | Nothing on its own. A reason to look closer. | — |
| SynthID and other pixel watermarks | Woven into the pixels or audio | Cannot be read here. Only the owner’s detector can see it. | Designed to survive editing |
The longer version, with how to check each one yourself: how to check whether something is AI-watermarked.
Sources: C2PA technical specification · IPTC digital source type vocabulary · EU AI Act, Article 50
QUESTIONS
Before you start
Is my website AI watermarked?
Enter its address above and you will see. A website itself carries no single AI watermark, so the check looks in the three places one can show up: the images on the page, which may carry C2PA Content Credentials or an IPTC “made by AI” label from the generator that made them; the page’s code, which shows whether it was built with an AI site builder like Lovable, v0 or Bolt; and the text, which is checked for AI writing patterns. Images uploaded through most CMSs and CDNs have their records stripped, so a clean result is common even on sites that use AI images.
Do AI companies have to watermark what they make?
Increasingly, yes. Article 50 of the EU AI Act, in force since 2 August 2026, requires providers of generative AI to mark synthetic images, video, audio and text in a machine-readable way; systems already on the market have until 2 December 2026 to add the marking. California’s AI Transparency Act, operative since the same date, requires providers with over a million monthly users to embed a hidden disclosure in the images, video and audio they generate. For media, that mostly means C2PA Content Credentials and the IPTC “trained algorithmic media” label — exactly what this reads. For text, no standard watermark exists yet that anyone outside the model maker can check.
Does ChatGPT or Claude text contain a hidden watermark?
Not one you can check. Neither OpenAI nor Anthropic publishes a text watermark that outsiders can detect, and Google’s SynthID text watermark can only be read by Google. The hidden characters many checkers look for — zero-width spaces, narrow no-break spaces — are real and this tool finds and removes them, but they are not a deliberate watermark: they come from copying text out of chat apps, and just as often out of Word, Google Docs or a PDF.
Can it tell whether text was written by AI?
Not reliably, and neither can anything else. Claude does not watermark its text, and Google’s SynthID watermark in text can only be read by Google. What this does is count the phrasings chat models overuse and show you each one it found. That is a style guess. Human copywriters trip it and edited AI text passes it, so it is labelled as a heuristic and never shown as a percentage.
It found nothing. Does that mean a person made it?
No. Most websites, CDNs and social networks re-compress images when they are uploaded, and re-compressing throws the records away. A clean result means nothing was declared, not that nothing was generated. A positive result is the one that tells you something.
Are the Content Credentials verified?
We read what the record says but do not check its cryptographic signature, so a forged record would read the same. For a verified reading, use the Content Authenticity Initiative’s checker at contentcredentials.org/verify.
What happens to the address I enter?
Our server fetches that page and up to twelve of its images, reads them, and sends the result back to you. Nothing is stored and the address is not logged or sent to analytics. Image and text checks never leave your browser at all.
Why does a watermark remover make this?
Because the two fit together. Our removers take the visible badge off footage you generated, and they deliberately leave the invisible C2PA record in place. This is the tool that reads that record.