Check a file for AI marks

PlentyLabs marks the AI content it makes, so people can tell that AI was involved. Use this page to check a picture, video or sound file for those marks.

It is free and needs no account. Your file stays on your device: the check runs in your browser and nothing is uploaded.

1. What the result means

• PlentyLabs watermark found. The image carries our invisible watermark. We add it to AI images we render, such as exported creatives. The format is ours, but the watermark is not signed, so on its own it cannot prove where an image came from.
• AI label in metadata. The file's metadata says it was made or changed with AI, using the standard IPTC term for this. Many tools write the same label, and anyone can add or remove it, so treat it as the file's own claim.
• Content Credentials. The file has a C2PA manifest, which some AI models add. You can read it with the Content Credentials verify tool.
• No mark found. This does not prove that no AI was used. Screenshots, editing apps and social media uploads often remove metadata. Our watermark survives resizing and re-saving, but not heavy cropping or shrinking an image to a thumbnail.

2. How we mark AI content

The EU AI Act (Article 50) asks providers of AI systems to mark AI-generated content in a machine-readable way, and to give people a way to detect those marks. PlentyLabs uses two kinds of mark:

• Metadata. A standard label inside image, video and audio files that says AI made or changed the content.
• An invisible watermark. A signal hidden in the pixels of images and videos we render. It stays in place when the file is resized, compressed or has its metadata removed.

A file can carry one mark or both, depending on how it was made and exported. Files that come straight from an AI model may also carry marks from the model's maker, such as Content Credentials or Google SynthID.

3. Read the marks yourself

Both marks use open standards and open-source tools, so you do not need this page to read them.

Metadata

An XMP packet with the IPTC property Iptc4xmpExt:DigitalSourceType, set to one of:

• http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia for content AI made
• http://cv.iptc.org/newscodes/digitalsourcetype/compositeWithTrainedAlgorithmicMedia for content AI changed

Any XMP reader shows it, for example exiftool -DigitalSourceType yourfile.jpg. Where the packet sits in each file type:

Watermark in images

We use TrustMark, an open-source watermark from Adobe and the Content Authenticity Initiative, with model variant Q and the BCH_SUPER error-correction schema. That schema carries 40 data bits, which we fill like this:

Decode the image with TrustMark's own decoder. If the 40 data bits start with the marker above, the image carries the PlentyLabs watermark.

Watermark in video

We use VideoSeal 1.0, an open-source video watermark from Meta. Its 256 bits hold one 32-bit unit repeated eight times; take a majority vote for each position across the eight copies. The unit uses the same layout as the image watermark: the marker in bits 0 to 15, the version in 16 to 19, what AI did in 20 and 21, and zeros in 22 to 31.

This page does not read video watermarks, because VideoSeal's detector is too heavy to run in a browser. Use VideoSeal's own detector instead.

4. Limits

• Some files carry only the metadata mark, for example images smaller than 256 pixels on their short side and some video exports.
• Our marks are not digitally signed. They show what a file says about itself, not proof of where it came from.
• A missing mark does not mean a file was made without AI.

5. Questions

Write to hello@plentylabs.com.