/

Updated 19 November 2026

Content Credentials (C2PA) and WordPress: What Survives Your Media Library

AI Disclosure: This content includes AI-generated text.

Short answer: the hidden AI marks that image generators add (Content Credentials, also called C2PA, and the IPTC “created using generative AI” tag) usually survive in the original file you upload to WordPress. They are much less likely to survive in the resized copies WordPress makes, and those copies are what most visitors actually download. That is fine for your legal position, because the machine-readable marking duty is on the tool makers, not on you. What you owe visitors under the EU AI Act is a visible label on realistic AI images. This guide shows what WordPress keeps, what it drops, and how to test your own site in five minutes.

Checked on 19 November 2026 against the WordPress core source code (trunk), the C2PA technical specification 2.2, the IPTC Digital Source Type vocabulary, Recital 133 and Article 50 of the AI Act, and the Commission’s Code of Practice on transparency of AI-generated content. Sources are at the end.

Why this matters now

Article 50(2) of the AI Act requires providers of generative AI systems to mark images, audio, video and text in a machine-readable way, so that it can be detected as AI-generated. The rule has applied since 2 August 2026. Systems already on the market before that date have a transitional period until 2 December 2026. After that date, you should expect more of the images you download from AI tools to carry these marks. Our 2 December deadline guide covers who that deadline applies to.

Recital 133 of the AI Act lists the techniques providers may use, alone or combined: watermarks, metadata identifications, cryptographic methods for proving provenance and authenticity, logging methods and fingerprints. For a website owner, two of these matter in practice, because they live inside the image file:

  • Content Credentials (C2PA). A signed “manifest” embedded in the file. In JPEG files it sits in special APP11 segments. It records how the image was made, for example a c2pa.created action with the digital source type “trainedAlgorithmicMedia”.
  • IPTC / XMP metadata. A plain metadata field, Digital Source Type, set to the IPTC value trainedAlgorithmicMedia, which IPTC names “Created using Generative AI”.

A third kind, invisible watermarks, is woven into the pixels themselves. It does not depend on metadata, so it is outside the scope of this guide.

What WordPress does to an uploaded image

When you upload a JPEG, WordPress keeps your original file and then creates several smaller copies (thumbnail, medium, large and any sizes your theme registers). Since WordPress 5.3, it also does this to very large images: if the image is wider or taller than 2,560 pixels, WordPress creates a “-scaled” copy and uses that as the main file, keeping the original separately. You can change the threshold with the big_image_size_threshold filter.

Every one of those copies is a new file, re-encoded by an image editor. WordPress tries Imagick first and falls back to GD (the wp_image_editors filter lists them in that order). The two behave differently:

Image editorWhat it keeps in resized copiesContent Credentials (C2PA)IPTC Digital Source Type (XMP)
Imagick (ImageMagick)Only the profiles named icc, icm, iptc, exif and xmp. Everything else is removed by strip_meta().Not on the keep list. Even if it were copied, the manifest’s hash would no longer match the resized pixels.Usually kept, because XMP is on the keep list
GDNothing: GD rebuilds the image from pixels and writes a new file with no metadata handlingLostLost
Your original uploadUntouchedKeptKept

You can see which editor your site uses under Tools → Site Health → Info → Media Handling (“Active editor”).

Why a resized copy can never keep valid Content Credentials

The C2PA specification ties a manifest to the exact file through a “hard binding”: one or more cryptographic hashes of the content. Resizing, recompressing or converting to WebP changes the bytes, so the hash no longer matches. A verification tool will then report the credentials as missing or invalid, even if a copy of the manifest were still in the file. Only a tool that edits the image and signs a new manifest can keep the chain intact, and the WordPress image editors do not do that. The specification also describes “soft bindings”, such as invisible watermarks and fingerprints, precisely because they can match content “even if the underlying bits differ”.

Practical result: on a typical WordPress site, the original in wp-content/uploads may still carry valid Content Credentials, while the 300, 768 and 1024 pixel versions shown in your posts do not.

Which file do visitors actually get?

  • Images in posts and pages. Even if you pick “Full Size” in the block editor, WordPress adds a srcset list of smaller copies, and the browser usually loads one of those. On a phone it is almost always a resized copy.
  • Featured images, sliders and backgrounds. Themes and page builders request a specific registered size, so these are resized copies too.
  • Images over 2,560 pixels. Even “Full Size” is the “-scaled” copy, not your original.
  • Image optimisation plugins and CDNs. Many have a setting to remove metadata to save bytes, and some convert to WebP or AVIF. Both remove or break Content Credentials. Check the plugin’s settings for options such as “strip metadata” or “keep EXIF/XMP”.
  • Social previews. Facebook, LinkedIn and others download your Open Graph image and process it again on their side.

Is losing the marks a legal problem for you?

Under the AI Act, generally not. Article 50(2), the machine-readable marking duty, applies to providers of generative AI systems: the companies behind the image generator. A website owner who uses those tools is a deployer, and Article 50 contains no duty for deployers to preserve or add machine-readable marks. The Commission’s Code of Practice reflects this split: Section 1 covers providers and marking, Section 2 covers deployers and labelling.

Your duty is the visible one. Under Article 50(4), if you publish an AI-generated or manipulated image that could falsely appear authentic (a deepfake), you must disclose that, clearly and at the latest at first exposure (Article 50(5)). Hidden metadata does not meet that duty, because visitors cannot see it. That applies whether or not the metadata survived.

Two cautions. First, if you build and offer your own AI image feature, for example a generator on your site, you may be the provider of that system and the marking duty may then be yours. See responsibility for third-party AI tools. Second, deliberately removing provenance from AI images to pass them off as real is the opposite of what Article 50 aims at. Even if Article 50 itself does not forbid it, it is a poor idea, and other rules such as consumer protection law can apply to misleading content.

Test your own site in five minutes

  1. Generate a test image with a tool that adds Content Credentials, and confirm the downloaded file shows them in a public verifier such as the Content Credentials “Verify” tool.
  2. Upload it to your Media Library and note the file URL.
  3. Download two files: the original (the plain URL, or the URL without “-scaled” if WordPress scaled it) and one resized copy (the same name with a size such as “-1024×1024”).
  4. Check both in the verifier. Expect the original to pass and the resized copy to show no credentials.
  5. To check the IPTC field, open both files in a metadata viewer (for example ExifTool) and look for Digital Source Type. With Imagick it should usually remain in the resized copy; with GD it will be gone.
  6. Repeat once with your image optimisation plugin or CDN switched on, using the URL a visitor actually receives.
  7. Write the result in your compliance notes with the date. It explains your setup if anyone asks.

What to do about it

  • Label realistic AI images visibly, everywhere they appear. This is the part that is your legal duty, and it works regardless of metadata. Our guide to labelling AI images in WordPress shows four ways, and the EU AI icons guide shows the optional official icons.
  • Keep the originals. Do not delete original uploads to save space. They are your evidence of where an image came from.
  • Don’t strip metadata on purpose. If your optimiser offers it, keep XMP and IPTC data, so the Digital Source Type field can survive in the resized copies.
  • Don’t treat “no metadata” as “not AI”. A scanner that finds nothing in a resized or optimised copy proves nothing. Check the original upload, or the source the image came from.
  • Watch your stock images. Stock libraries increasingly include AI images. Their marks may disappear on your site just like your own.
  • Leave hidden provenance to the toolmakers. Re-signing Content Credentials after resizing would need a C2PA signing setup. That is not something a typical business website needs for AI Act compliance.

Our free EU AI Act Ready plugin checks uploads for AI signals in metadata, EXIF data and file names, and adds visible labels to flagged images, including featured images and page builder images. Because detection reads metadata, scan images when you upload them, and review flags yourself. The plugin comparison notes that some other plugins also write an AI marker into the file itself, which ours does not do.

FAQ

Does WordPress support Content Credentials?
WordPress core stores and serves your original file unchanged, but it has no C2PA features: it does not read, display, verify or re-sign manifests, and its resized copies do not carry valid credentials.

Can I stop WordPress from resizing AI images?
You can turn off the 2,560 pixel scaling with the big_image_size_threshold filter, but WordPress will still create the smaller sizes your theme uses, and browsers will still load them. Serving only originals would make pages slower without changing your legal duty. The visible label is the better fix.

If my image has Content Credentials, do I still need a visible label?
Yes, if it is a deepfake under Article 50(4), meaning it could pass as real. Machine-readable marks are for software and platforms; the label is for people.

What changes on 2 December 2026?
Generative AI systems that were already on the market before 2 August 2026 must then also mark their output. Your own deployer duties under Article 50(4) have applied since 2 August 2026 and do not change on that date.

Sources

This article is general information, not legal advice. Written by the EU AI Act Ready team, 19 November 2026.