Check if your product images are marked as AI-generated.
Drop in a folder of product images and see which ones carry a machine-readable AI marking, which do not, and which lost theirs somewhere in your pipeline. Everything runs in your browser. No upload, no account.
Runs entirely in your browser · nothing is uploaded · up to 200 files
A missing mark is not proof an image is not AI-generated, and a mark present is not by itself proof of compliance. Where an image carries C2PA Content Credentials we verify the signature cryptographically, but we do not check the signer against an official trust list. Invisible watermarks such as SynthID live in the pixels rather than the metadata, so no metadata tool can read them. This is an engineering signal, not legal advice.
Anything unmarked above had to be marked by whoever generated it. Every image On-Model produces carries the marking automatically, on every plan, whether or not you switch on the visible watermark. One less thing to fix before a catalog goes live.
How the checker reads an image.
The checker answers a narrow question precisely, rather than a broad one vaguely.
It reads the file, not the picture
The checker parses the container — EXIF, XMP, IPTC records, PNG text chunks, and any embedded C2PA manifest. It never looks at the pixels, so it cannot be fooled by how an image looks, and it cannot guess either.
It checks the mark is real
Two things count as a machine-readable marking: a C2PA manifest, or an IPTC DigitalSourceType field naming the content as trained-algorithmic media. Where there is a C2PA manifest we verify its signature, because a manifest is only bytes and anyone can write one that lies.
It tells you what is missing
A file with a manifest reference but no manifest was signed and then stripped somewhere downstream. A file with nothing at all has been through a re-encode. Both are common, and both are reported as their own result rather than a pass.
What Article 50(2) actually asks.
The short version, for the person who has been handed this question and needs an answer today.
Providers of AI systems that generate synthetic image, audio, video or text content have to mark the output in a machine-readable format, so that it can be detected as artificially generated or manipulated. The marking must be effective, interoperable, robust and reliable as far as is technically feasible.
Article 50 has applied since 2 August 2026. Generative systems already on the market before that date have until 2 December 2026 to meet the marking requirement. The Digital Omnibus deferred several high-risk deadlines. It did not move these.
On-Model is not a law firm and this page is not legal advice. If a compliance decision turns on it, take the exported record to your counsel. Our longer write-up is here.
Asked, answered.
How do I check if an image is marked as AI-generated?
Drop the image into the checker above. It reads the file’s metadata in your browser and reports whether it carries a machine-readable AI marking — either C2PA Content Credentials or an IPTC DigitalSourceType field. You can check a whole folder at once and export the result as a CSV.
What does the EU AI Act Article 50(2) actually require?
Providers of AI systems that generate synthetic images, audio, video or text must mark the output in a machine-readable format so it can be detected as artificially generated or manipulated. The obligation applies from 2 August 2026, with generative systems already on the market having until 2 December 2026 to comply. The marking has to be effective, interoperable, robust and reliable as far as is technically feasible.
Does a missing mark mean the image is not AI-generated?
No. Metadata is fragile: a CDN re-encode, a screenshot, a social platform upload, or an image editor will often strip it. A missing mark means no mark is present in the file you checked, not that the image came from a camera. It is also common for the original marking to have been correct and to have been removed downstream.
Can this tool detect invisible watermarks like SynthID?
No, and no metadata reader can. Invisible watermarks are embedded in the pixels themselves and need the provider’s own detector. An image can be fully compliant through a pixel watermark and still show no metadata marking here, so a "no marking found" result is never proof that a provider failed to mark its output.
Does a detected mark prove the image is compliant?
Not on its own. Where an image carries C2PA Content Credentials the checker verifies the signature cryptographically, so it will tell you if a manifest was forged or if the image was altered after signing. It does not check the signer against an official C2PA trust list, and an IPTC metadata marking carries no signature at all and so cannot be verified this way. Treat the result as an engineering signal for your own catalog, not as a legal opinion.
Does the checker verify C2PA signatures?
Yes. A C2PA manifest is just bytes, and anyone can write one claiming an AI image was shot on a camera, so reading the claim is not enough. When a file carries Content Credentials the checker runs the official C2PA validator in your browser and reports whether the signature holds, who signed it, and whether the pixels still match what was signed. An image edited after signing comes back as invalid. What it does not do is check the signer against an official trust list, which would mean fetching that list from a third party and breaking the promise that nothing leaves your browser.
Are my images uploaded anywhere?
No. The check runs entirely in your browser using local JavaScript. The image bytes are never sent to On-Model or to any third party, and nothing is stored. You can confirm this by disconnecting from the network after the page loads — the checker still works.
Does On-Model mark the images it generates?
Yes, automatically, on every plan and for every output, independently of the visible watermark. Each generated image carries an IPTC DigitalSourceType of trainedAlgorithmicMedia in an XMP packet, which is the interoperable form third-party tools read. You can verify it by running any On-Model output through this checker.
Which file formats does the checker read?
JPEG, PNG, WebP, GIF and TIFF are fully supported, with best-effort reading for AVIF and HEIF. It reads EXIF, XMP including IPTC DigitalSourceType, IPTC IIM records, PNG text chunks, and C2PA manifest presence.
Guides and comparisons.
Marking AI-generated images under the EU AI Act
What Article 50 asks for, what On-Model does automatically, and where your responsibility as a brand begins.
Quality control for AI product images
The checks worth running on a generated catalog before it reaches a product page.
AI infrastructure for fashion imagery
Why marking, batching and provenance belong in the pipeline rather than bolted on at the end.
Marked by default. Every plan.
On-Model is the AI infrastructure for fashion imagery. Marking is part of the pipeline, not a setting you have to remember.