/

Updated 18 November 2026

2 December 2026: The AI Image Marking Deadline Explained for Website Owners

AI Disclosure: This content includes AI-generated text.

Short answer: on 2 December 2026, the four-month grace period ends for AI tools that were already on the EU market before 2 August 2026. From that date, every image, video, audio or text generator used in the EU must add a machine-readable mark to what it produces. That duty sits with the tool makers (providers). If you run a website, the deadline mostly changes what arrives in your media library, not what you owe. Your own duty, visibly labelling deepfakes, has applied since 2 August 2026 with no grace period. This guide explains who has to do what, and what to check on a WordPress site before December.

Legal status checked on 18 November 2026 against the AI Act as amended by the Digital Omnibus on AI (Regulation (EU) 2026/1744), the Commission’s Article 50 guidelines and the Code of Practice on Transparency of AI-generated Content. Sources are listed at the end.

What actually changes on 2 December 2026

Article 50(2) of the AI Act requires providers of AI systems that generate synthetic audio, image, video or text to make sure the outputs are “marked in a machine-readable format and detectable as artificially generated or manipulated”. Article 50 as a whole has applied since 2 August 2026.

The Digital Omnibus on AI, which entered into force in July 2026, added a transitional rule to Article 111 of the AI Act. Providers whose generative systems were already placed on the market before 2 August 2026 must comply with Article 50(2) by 2 December 2026. According to the Omnibus recitals, the four months exist so that providers can adapt “without disrupting the market”. Systems launched on or after 2 August 2026 never had a grace period.

Three things the December date does not do:

  • It does not postpone the deployer duties in Article 50(4). Labelling deepfakes and certain AI-generated text has applied since 2 August 2026.
  • It does not cover chatbot disclosure (Article 50(1)). That has also applied since 2 August 2026.
  • It does not move the high-risk rules. Those were pushed back separately by the Omnibus, to 2 December 2027 for Annex III systems and 2 August 2028 for Annex I products.

Provider or deployer: which one is your website?

Article 50 splits duties between two roles. Getting your role right tells you which deadline matters to you.

RoleTypical exampleMain image dutyApplies from
ProviderThe company behind an image generator, or a plugin or SaaS that builds its own AI image feature into a product it offers to othersMachine-readable marking of outputs (Art. 50(2))2 Aug 2026 for new systems, 2 Dec 2026 for systems already on the market
DeployerA business or blogger who uses an AI tool to create images for its own websiteVisible disclosure of deepfakes (Art. 50(4))2 Aug 2026, no grace period

Most website owners are deployers. You become a provider when you offer an AI system to others under your own name, for example a WordPress plugin or a web app with a built-in AI image generator, even if it runs on a third-party model behind the scenes. The Commission’s guidelines make clear that Article 50 attaches to AI systems, so application-layer businesses built on someone else’s model can be providers too. If that is you, the 2 December deadline is yours, and you should read the provider section of the Code of Practice now.

For a deeper look at the split, see our guide on whether you are responsible for third-party AI tools on your website.

What “machine-readable marking” means in practice

The AI Act itself does not name a technique. It asks for solutions that are “effective, interoperable, robust and reliable” as far as technically feasible. The Code of Practice, which the Commission and the AI Board confirmed in July 2026 as an adequate way to show compliance, fills in the detail:

  • At least two layers. Signatory providers mark images, video, audio and text with two or more machine-readable techniques, for example digitally signed, tamper-evident metadata (such as C2PA Content Credentials) plus an invisible watermark in the pixels.
  • A detection tool. Providers offer a way to check whether content came from their system, in principle free of charge.
  • Interoperable detection later. The Code asks providers to make their watermark detection interoperable by 2 February 2027.

The reason for two layers matters to website owners. Metadata is easy to lose: re-saving, cropping, converting to WebP or generating thumbnails can drop it or break its signature. An invisible watermark is meant to survive those steps. So after December, images from compliant tools should stay detectable as AI-generated even after your site has processed them, but only through the watermark, not through the metadata your visitors’ browsers can read.

Why the marking does not replace your visible label

This is the point most often misunderstood. A watermark in the file is for machines. Your visitors cannot see it. The Commission’s guidelines state that the provider’s marking does not discharge the deployer’s duty to label.

Under Article 50(4), if you publish an AI-generated or manipulated image, video or audio that is a deepfake, you must disclose that it is artificial. The AI Act defines a deepfake as content that “resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful”. According to the guidelines, the test is objective: intent to deceive is not required, and photorealism alone does not decide it. What counts is whether the subject could plausibly be real.

For a typical business website, that means:

  • Needs a visible label: a realistic AI “photo” of a team, an office, a customer, a product in use, a place or an event.
  • Usually not a deepfake: an obviously illustrated hero graphic, an abstract background, a cartoon, or a fantasy scene nobody would take as real. Labelling these is still good practice and builds trust.
  • Creative works: for evidently artistic, satirical or fictional content, a lighter disclosure is enough, as long as it does not spoil the work.

The label must be clear, distinguishable and visible at the latest when a visitor first sees the content (Article 50(5)), and it must meet accessibility requirements. A note in your privacy policy or footer is not enough. Our guide on where to place AI disclosures covers placement in detail.

The EU icons you can use

The Commission has published a free set of EU icons for labelling AI-generated content: a basic icon, a “fully AI-generated” icon and a “partially AI-modified” icon, each in black, white and 50% transparent versions, in SVG and PNG. Using them is optional, but the labelling duty is not. The Commission’s display guidance says the icon should be visible at first exposure, not hidden behind overlays, and embedded in the content where possible, or shown through an equivalent overlay in the user interface.

A pre-December checklist for WordPress sites

  1. List your AI image sources. Note every tool your team uses to create or edit images (image generators, AI features in design apps, AI stock image services, AI fill or background removal). Ask each vendor, in writing, how and from when they mark outputs under Article 50(2).
  2. Check whether you are a provider anywhere. If your site or product lets visitors generate images with AI, you may be a provider of that feature. Talk to whoever built it about marking before 2 December.
  3. Find the deepfake candidates in your media library. Look for realistic people, places, products and events that were AI-generated or heavily AI-edited and published on or after 2 August 2026. According to the guidelines, content published before that date does not need to be labelled retroactively, though labelling it anyway costs little.
  4. Add a visible label to each one. A small badge on the image, or a caption directly under it, with the EU icon or plain words such as “AI-generated image”.
  5. Do not strip provenance on purpose. Some optimisation plugins and hosts remove embedded metadata to save bytes. That is not illegal for deployers, but keeping provider marks intact helps anyone verifying your images later. Check one of your own resized images to see what survives.
  6. Write down your process. A short internal note saying who decides what gets labelled and how is cheap, and the Code of Practice expects deployers to have proportionate internal processes.

How EU AI Act Ready helps

Our free plugin, EU AI Act Ready, covers the deployer side of steps 3 and 4. It scans new uploads and your existing media library for signs of AI generation (metadata, EXIF and file name patterns), flags likely AI images, and shows a visible label to visitors. You can mark or unmark any image yourself, because no automatic check is perfect. Images whose metadata was already stripped need that manual check. The plugin does not write C2PA or watermarks itself. Machine-readable marking is the provider’s job under Article 50(2).

What happens if you get it wrong

Breaches of Article 50 fall under the middle fine tier in Article 99: up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher. For SMEs and start-ups the lower of the two amounts applies. National market surveillance authorities enforce the rules. In Germany that is the Bundesnetzagentur. In practice, regulators are expected to start with guidance and corrective orders for small sites, but a missing label on a realistic AI image is an easy finding. See EU AI Act fines and penalties explained for the full picture.

FAQ

Do I have to watermark AI images on my website by 2 December 2026?
No, not as a website owner using someone else’s tool. Machine-readable marking is the provider’s duty. Your duty is the visible label for deepfakes, which has applied since 2 August 2026.

My image tool already adds Content Credentials. Am I covered?
Not for visitors. Content Credentials and watermarks are machine-readable. If the image is a deepfake, you still need a visible label on your page.

Do I need to label AI illustrations and icons?
Usually not by law, if nobody would take them as real. Labelling them anyway is a simple way to be transparent.

What about AI-written text?
AI-generated text published to inform the public on matters of public interest must be disclosed, unless a person with editorial responsibility reviewed it. Product descriptions and fiction are generally outside that rule. Our guide to disclosing AI-generated content covers text in detail.

Sources

This article is general information, not legal advice. For decisions about your own business, talk to a lawyer.

Practical guide: How to label AI-generated images in WordPress.