AI Watermark Checker

Three different things get called 'AI watermark'

Most confusion in this area comes from one word covering three unrelated mechanisms with completely different verification stories. Separating them is the whole job.

Signed provenance — C2PA Content Credentials — is an open standard. Anyone can validate it with public tooling, which is why this site can check it locally and honestly.

Perceptual watermarks such as Google's SynthID modify the media itself in ways designed to survive compression, cropping, and re-encoding. The signal lives in the pixels, not in a metadata field, and detection requires the provider's own detector. No third-party page can read it, and any page claiming to is not doing what it says.

Ordinary metadata strings are just text fields. A generator may write its name into EXIF or XMP. This is trivially readable and equally trivially editable or removable, which makes it a hint and never proof.

• Signed provenance (C2PA) — open, cryptographic, independently verifiable, easily stripped.
• Perceptual watermark (SynthID and similar) — robust to editing, invisible, only verifiable by the provider.
• Metadata string — readable by anyone, writable by anyone, removed by most platforms.

Why 'watermarked' rarely means 'entirely machine-made'

Providers apply marks to output their systems produced, and that output is often a modification of something a person made. A generative fill that removed a lamppost, an upscale, a background extension, a style transfer — all of these can carry a mark.

When a check reports a signal, the accurate statement is that a provider system touched this asset. Moving from there to 'this image is fake' skips several steps and is frequently wrong.

The provenance model handles this properly through actions and digital source types, which is why reading those fields matters more than reading the top-level badge.

What a negative result is worth

Very little, and this deserves stating bluntly because it is where most misuse happens. A negative result means this particular check found nothing. It does not mean the image is authentic, human-made, or unmodified.

Consider what has to go right for a positive result: the generator has to write a signal, the format has to carry it, every intermediate tool has to preserve it, and the detector has to be available to you. Any break in that chain produces a negative on genuinely synthetic content.

Statistical AI-image classifiers try to fill this gap by guessing from pixels, and their published accuracy is worth reading carefully. It degrades sharply on compressed, cropped, or screenshotted images — exactly the images people actually need to check — and their false positives fall hardest on real photographs that happen to look unusual. This site does not run one, because presenting a guess as a measurement is the specific harm we are trying to avoid.

Where regulation is pushing this

The European Union's AI transparency obligations require providers of generative systems to mark synthetic output in a machine-readable way and require certain deployers to disclose synthetic content. This is driving broader adoption of exactly the kind of marking this tool reads.

The direction of travel is good for verification: more content will carry provenance over time, and a negative result will slowly become more informative than it is today. But adoption is uneven across providers and jurisdictions, and stripping remains trivial.

Plan for a long transition. For now, the reliable posture is to treat provenance as one input among several rather than as an oracle.

A practical checking order

When you actually need an answer about a specific image, this sequence extracts the most information for the least wasted effort.

• Get the most original file you can. A screenshot or a re-share has usually lost everything worth reading.
• Validate any C2PA credential first — it is the only cryptographic evidence available to you.
• Read the ordinary metadata as supporting context, treating every string as editable.
• If you suspect a specific provider, use that provider's official detection flow.
• Reverse image search for earlier or higher-resolution copies, which often surfaces the real origin faster than any watermark check.
• Ask the source for the original. This works more often than people expect and beats every technical measure.