C2PA and Content Credentials: how to check what is in an AI image

An AI image can carry its own origin: a machine-readable entry in the file, written by the tool that generated it. These entries are called Content Credentials (technically, a C2PA manifest) and IPTC DigitalSourceType. They are why some AI images in a store can be found automatically, and why this never covers every image. Below: what the file contains, who writes it, how you check it, and where it gets lost.

What is stored in an image file with Content Credentials?

Provenance data sits in three layers of the file’s metadata, invisible in the image itself:

Key point: a marker is a statement by the generating tool, not a measurement of the image. Asked whether C2PA detects AI content, the IPTC FAQ of 8 July 2026 answers: “No … They simply extract the data that was put there by its creator or publisher.” A marker that is found is evidence that the tool recorded an AI origin. A missing marker says nothing. The image may be a photo, an unmarked AI image, or an AI image whose marker was lost along the way.

Who writes the marker: the generator, the image editor, or you?

The generator. The provider (Anbieter) of the AI system is responsible for the machine-readable marking under Article 50(2) of the AI Act (Regulation (EU) 2024/1689). Providers whose systems were placed on the market before 2 August 2026 have until 2 December 2026 to comply with Article 50(2) (Article 111(4), inserted by Regulation (EU) 2026/1744; EUR-Lex retrieved 31 August 2026). Gaps in marker coverage are therefore normal in fall 2026.

The image editor. Tools that modify an existing photo write markers too. Canva sets trainedAlgorithmicMedia for an image generated purely from a prompt (signed by OpenAI, model gpt-image 2.0), but compositeWithTrainedAlgorithmicMedia, signed by Canva, for your own photo with an AI-generated background (measured 1 September 2026). Adobe Firefly embeds no manifest, only an XMP pointer to a cloud manifest at Adobe; a verification tool needs network access to resolve it.

The aggregator writes nothing. With services that bundle many models, the signature comes from whoever actually runs the model, not from the service itself. In 51 files from Magnific/Freepik, Freepik did not appear in a single signature chain. The signers were OpenAI, Google, Byteplus, Black Forest Labs, fal, Ideogram, Recraft, and Microsoft. Two Grok files carried a complete manifest signed by a self-signed “ephemeral CA” without a trust anchor: a verification tool sees the statement but cannot confirm who made it (1 September 2026). Whether a marker is included depends on the individual catalog entry: Flux.1 carried a manifest, Flux.1 Fast did not, despite the same base model.

You. As a merchant, you are a deployer (Betreiber) under Article 3(4) of the AI Act. Your obligation arises not from Article 50(2) but from Article 50(4): the visible disclosure on the image (see Do I have to label AI images in my store?). The Commission guidelines C(2026) 5054 of 20 July 2026 state in paragraph 117 that deployers cannot rely on the marker embedded by the provider. The marker helps you find the images to check for a label; it is not the label.

Which tools write a marker and which do not?

The complete table per tool and model, with signer and what happens to the Shopify copy, is in the guide Which AI image generators write provenance evidence?. Here is the overview: on 1 September 2026, the Easy AI Act Image Labels test bench read 59 files: 47 catalog entries of one aggregator via API, four cross-checks from its web interface, six files from five other tools, and two controls. Of these, 43 carried a provenance marker. The Shopify image editor was tested separately the same day. The numbers apply to that day and those export paths; a quarterly retest is planned, the next after the 2 December 2026 deadline.

One special case: the Shopify image editor delivers no metadata block on the unchanged original (2,014,705 bytes). Whether Shopify sets no marker or loses it in its own processing cannot be determined from the outside (1 September 2026; revisit after 2 December 2026).

How do you check whether an image carries a marker?

1. Use the original, not a copy. Markers survive only as long as nobody re-encodes the file.
2. Read the file with a verification tool. c2patool, the C2PA command-line tool, reads manifests; exiftool shows XMP and IPTC fields such as DigitalSourceType. Without a command line, the free AI image checker and the AI metadata checker on this website read both markers directly in the browser; the file does not leave the device. They cannot see an invisible watermark in the pixels, such as SynthID; that needs the provider’s own tool, see AI watermarks.
3. In Shopify: read the stored original, not the storefront view. For files in the media library (Content > Files), Shopify keeps the untouched original and exposes it through the Admin API as originalSource. On 1 September 2026, 47 of 47 uploaded test images were byte-identical there to the local file, all 34 markers preserved. Visitors mostly get re-encoded copies without markers: via the storefront CDN address, only 5 of 34 markers came through on 1 September 2026; via the image URL that the Admin API returns for display, 0 of 3 on 10 September 2026. Scanning the online store from the outside measures the image pipeline, not the provenance.
4. Know the state “not verifiable.” Collection images uploaded directly to the collection have no media library original and cannot be checked (measured 10 September 2026). The fix: upload the image under Content > Files and select it from there in the collection.

The app Easy AI Act Image Labels reads the original through the Admin API, never the storefront copy; what it cannot check is listed under Limits.

Which export paths destroy the marker?

Every re-encoding of the pixels creates a new file. The C2PA manifest is never carried over; the XMP field with the DigitalSourceType survives only in exceptional cases (in Shopify for PNG, not for JPEG). All measurements below: 1 September 2026, unless noted.

Four rules:

1. Use the tool’s official download button, never a screenshot, never right-click.
2. Choose the highest resolution offered, never “optimized for web”.
3. Keep the native output format; decline any conversion.
4. Upload directly to the store; put nothing in between.

A version with a burned-in icon is a new file too. The free AI Label Generator carries the IPTC data of the original over into the copy; a C2PA manifest only applies to the unchanged file and is not in the copy. The original therefore remains the file with the complete evidence.

Are you allowed to remove markers deliberately? The AI Act does not prohibit third parties from removing them; the Commission recommends preserving them (guidelines C(2026) 5054, paragraph 98; retrieved 31 August 2026). Binding prohibitions come from terms of use and platform rules: the Black Forest Labs Usage Policy prohibits removing or suppressing C2PA credentials and watermarks (retrieved 31 August 2026). Per its help page (support.google.com/merchants/answer/6324350, retrieved 14 September 2026), Google Merchant Center requires the IPTC marker DigitalSourceType on AI-generated feed images and prohibits removing it. Whether Shopify passes the marker from the original file on to Merchant Center is unmeasured as of 14 September 2026; a spot check on a feed image with exiftool is worthwhile.

What does the marker mean for your labelling obligation?

The marker answers the question “Which images come from an AI tool?” for the share of images whose tool wrote it and whose export path preserved it. It does not answer whether an image must be visibly labelled; that is decided under Article 50(4) of the AI Act by whether the image is liable to mislead, for instance a product that looks different from what is delivered (Commission guidelines of 20 July 2026, paragraphs 113 to 116; details in the obligation guide).

Nor does it replace the label: only a label on the image is visible. The official version is the EU icon with the abbreviation “AI” in three variants, exclusively with English wording (Commission page on the EU icons, dated 10 August 2026, retrieved 31 August 2026). The label word in other EU languages is a translation by Easy AI Act Image Labels, not an official file. Variants and placement: The EU icons for AI images. The obligation itself is covered in Article 50 explained.

In practice: markers on the original are evidence and a solid basis for an automatic inventory; results from probability detectors are hints; a missing marker says nothing. If you know an image comes from an AI tool, label it whether or not a marker was found.