← All articles

AI-generated images on WordPress: is your site erasing the marking?

Our test of 19 October 2026: WordPress keeps the original's C2PA marking, but none of the 28 resized versions keep it. What Article 50 asks of you before 2 December, and what it asks of the provider.

Dots Papers cover for the article on the marking of AI-generated images in WordPress, Artificial Intelligence section

In brief

  • Article 50 of the AI Act requires providers of AI tools to mark the images they generate in a machine-readable way. For tools placed on the market before 2 August 2026, this marking becomes mandatory on 2 December 2026.
  • On 19 October 2026 we uploaded three AI-generated images to a test WordPress: the original file keeps its marking, but of 28 resized versions of images carrying a C2PA manifest, none kept it.
  • Yet these resized versions are what WordPress shows your visitors most of the time.
  • Machine-readable marking remains the provider’s job, and the provider must also add an invisible watermark designed to survive resizing. What falls to you is the visible label on images that could pass for real.
  • Keep the originals, never remove a marking on purpose, and label what needs labelling: that is the core of compliance on a WordPress site.

To comply with the marking of AI-generated images on WordPress, you first need to know what WordPress does with it. Our test of 19 October 2026 shows that the original file keeps its marking, but the resized versions, the ones the site displays, lose it: of 28 versions of images carrying a C2PA manifest, none kept it. Marking is the obligation of the tool’s provider; yours is the visible label on deepfakes, and it does not depend on any metadata.

How we tested On 19 October 2026 we generated two images with an AI tool that adds a C2PA manifest, in JPEG and PNG, and prepared a third carrying only the IPTC “trainedAlgorithmicMedia” statement. We uploaded them to a test WordPress 7.0.2, once with the Imagick image library and once with GD, then looked for the marking in the original file and in each of the seven sizes generated. The texts were reread the same day. Sources are listed at the end of the article.

What does Article 50 require for AI-generated images?

Two obligations, aimed at different people.

  • The tool’s provider must mark the images its system generates, in a machine-readable format detectable as artificial (Article 50(2)). For systems placed on the market before 2 August 2026, this marking is required from 2 December 2026.
  • Whoever publishes the image, the deployer, must disclose deepfakes: an image generated or manipulated so that it resembles existing people, objects or places and could falsely appear authentic (Article 50(4)). This obligation has applied since 2 August 2026.

A shop using a subscription image generator is in principle a deployer: the invisible marking is not its job, the visible label is. We set out this split, and the cases where you become a provider without knowing it, in what Article 50 of the AI Act asks of your chatbot.

Does WordPress keep the marking of AI-generated images?

The original file, yes. The versions WordPress makes from it, no. At each upload, WordPress creates several sizes of the image, from the thumbnail to the large version, and those are what it inserts into pages, with the variants offered to the browser depending on the screen. Here is what we found.

Image uploaded Library Original file The 7 sizes generated
JPEG with C2PA manifest Imagick Marking kept Lost, 7 out of 7
PNG with C2PA manifest Imagick Marking kept Lost, 7 out of 7
JPEG with IPTC statement only Imagick Marking kept Kept, 7 out of 7
JPEG with C2PA manifest GD Marking kept Lost, 7 out of 7
PNG with C2PA manifest GD Marking kept Lost, 7 out of 7
JPEG with IPTC statement only GD Marking kept Lost, 7 out of 7

Three lessons. The C2PA manifest, the most widespread standard among generators, survives in no resized version, whatever the library. An IPTC statement added on its own survives with Imagick, which keeps certain metadata blocks by default, and disappears with GD, which removes them all. And in the generator’s image, the IPTC statement was stored inside the C2PA manifest: it went with it.

Copying the manifest into the smaller versions would not solve anything. A C2PA manifest is bound by hash to the file it describes: applied to a resized image, it would no longer match, and a verification tool would flag it as invalid. Only a compatible tool can issue a new manifest that cites the original.

Is it serious, and who is responsible?

Less than it seems, for two reasons.

First, marking does not rely on metadata alone. The Commission’s code of practice on transparency of AI-generated content asks signatory providers to combine signed metadata with an invisible watermark written into the image itself, and to make it robust to common processing, including recompression, change of file format and resizing. Our test does not measure that watermark, which can only be read with the provider’s detection tool.

Second, the same code distinguishes deliberate removal from good-faith processing. Signatory providers commit to prohibiting, in their terms of use, the intentional removal or tampering of markings by deployers, and the code encourages signatories who disseminate content to the public to preserve these metadata. In our reading, WordPress’s automatic resizing is not intentional removal. Stripping the marking from an original to pass a generated image off as a photo is.

How do you comply on a WordPress site?

  1. Keep the originals. WordPress keeps them, with their marking. Do not enable an optimisation option that replaces or recompresses the original without checking what it does to metadata.
  2. Never remove a marking on purpose, whether with a metadata-cleaning plugin or by re-saving the image in software before uploading it.
  3. Label what could pass for real. A generated model, a realistic avatar, an embellished product: the statement “Image generated by artificial intelligence” goes in the caption or on the visual, visible as soon as the image is. In WordPress, the Image block’s caption does the job.
  4. Ask your provider how it marks its images, by metadata and by watermark, and whether it provides a detection tool.
  5. Run the test on your own site. Upload a generated image, download one of its sizes from a published page, and put it through a “Content Credentials” verifier. In five minutes you will know what your installation keeps.

What this article is not Legal advice. It describes Article 50 of Regulation (EU) 2024/1689, its partial postponement by Regulation (EU) 2026/1744 and the Commission’s code of practice as we read them. Our test covers a WordPress 7.0.2 with no image optimisation plugin: a different setting or plugin may give a different result.

What this says about the bigger picture

The marking of generated images is designed as a chain: the provider marks, platforms carry, the public checks. Our test shows where this chain breaks on an ordinary site, without anyone meaning it to: at the most mundane step, resizing. That is why the rules do not rely on a single technique, and why the visible label remains, for whoever publishes, the only guarantee they fully control.

That work, from the inventory of visuals to the statements to display, is what we do through our artificial intelligence expertise.

Sources

Texts reread on 19 October 2026.

  1. Regulation (EU) 2024/1689 on artificial intelligence (AI Act), Article 50, EUR-Lex. View
  2. Regulation (EU) 2026/1744 of 8 July 2026, the digital omnibus on AI, marking deadline of 2 December 2026 for systems placed on the market before 2 August 2026. View
  3. European Commission, code of practice on transparency of AI-generated content, section 1, measures 1.1 (signed metadata and invisible watermark), 1.2 (non-removal of markings) and 3.3 (robustness). View
  4. European Commission, questions and answers on the transparency obligations of Article 50. View

FAQ

Does WordPress remove the C2PA marking from AI-generated images?

In the resized versions, yes. In our test of 19 October 2026 on WordPress 7.0.2, the original file kept its C2PA manifest, but none of the 28 resized versions of C2PA images had kept it, with Imagick as with GD.

Who has to mark an AI-generated image: me or the tool?

Machine-readable marking falls to the tool’s provider (Article 50(2)), from 2 December 2026 for tools already on the market before 2 August 2026. You, who publish the image, must disclose with a visible label those that could pass for authentic.

Do you need a WordPress plugin to comply?

Not for the obligation that falls to you: the visible label goes in the caption or on the visual. Be wary instead of plugins that clean metadata or recompress originals, and check what they do to the marking.

Does a resized image also lose the invisible watermark?

Not necessarily. The code of practice asks providers for a watermark written into the image itself and robust to resizing and recompression. Our test does not measure it: it can only be read with the provider’s detection tool.

Can Dotsland help?

Yes. We inventory your generated visuals with you, test your installation and add the statements where they are needed. Start with the five steps above, then get in touch.

Want to apply this to your own business?

Ask us about this topic →

Further reading

Leave a comment

Your email address will not be published. Required fields are marked *

7 − 7 =