AI-Generated Images on Your Belgian Website (2026)

You have been using DALL-E for blog post headers for a year. A new client asks whether you need a licence. For most ordinary marketing images the answer is no. The four risk layers below explain when that answer changes and what Article 50(4) of the AI Act adds from 2 August 2026.

In Belgium the copyright framework for AI images sits in Book XI of the Code of Economic Law (WER), not in a separate Copyright Act as in the Netherlands. Article XI.165 WER requires "an own intellectual creation of the author". The Court of Cassation has tightened that criterion in Cass. 26 January 2012, C.11.0108.N: mere prompt output from a generative model usually fails that threshold and falls outside Belgian copyright. What can be protected is the composition, layout or post-editing in which the user adds creative choices.

On the personal-data and portrait-rights side the supervisor is the Gegevensbeschermingsautoriteit (GBA). For AI images in which a real person appears recognisably, the GBA Recommendation 01/2024 on the use of AI applies in conjunction with EDPB Opinion 28/2024 on AI models. Belgian courts also apply the Constitution (Article 22 on privacy) and Article 8 ECHR for portrait-rights claims. That is a different framework from the Dutch one, where Article 21 of the Auteurswet handles portrait rights on a specific works-law basis.

The enforcement route in Belgium is also organised differently. No Autoriteit Persoonsgegevens, no Reclame Code Commissie, but a combination of GBA procedures, Test-Achats as the consumer body with collective complaint standing under Book XVII WER, and the FOD Economie / SPF Économie as market authority. For AI Act elements that are not about personal data (such as the generic labelling duty in Article 50), the FOD Economie becomes the Belgian market-surveillance authority once Regulation 2024/1689 applies to providers and deployers from August 2026.

The short answer

You can publish generic AI marketing images in Belgium just fine. The risk sits in four places: images that reproduce a specific existing work, images with a brand logo or stock watermark, images with recognisable real persons and images that look like real photographs. A Getty-style demand letter for an AI image is rare but possible when the output reproduces an existing photograph or carries a Getty watermark. Otherwise the worst that happens is you discover you do not own the copyright in the image you "made".

What the judges actually decided

The most-cited ruling is Getty Images (US) Inc & Ors v Stability AI Ltd [2025] EWHC 2863 (Ch), UK High Court, 4 November 2025. You can read the full judgment on the Judiciary site. The popular headline was "Stability AI wins on most points". That is roughly accurate but the holdings are narrower than the headline suggests. For Belgian readers this is UK-specific context, not settled EU law.

Getty withdrew its primary copyright-infringement claim mid-trial because it could not prove that the relevant training had taken place in the UK. The court therefore did not decide whether training a diffusion model on copyrighted work amounts to infringement. The reading "the judge said training is fine" is wrong. The judge did not answer that question.

The court did rule on secondary infringement under sections 22, 23 and 27 of the UK Copyright, Designs and Patents Act 1988. The Stable Diffusion model weights are not a "copy" of the training images. Trademark infringement under sections 10(1) and 10(2) of the Trade Marks Act 1994 succeeded partially for a limited set of synthetic images carrying the Getty or iStock watermark.

For Belgium the open questions matter more than the UK holdings. The Infopaq test (Case C-5/08) determines originality under Book XI of the Code of Economic Law. Robert Kneschke v LAION in Germany produced a first-instance decision in 2024 favourable to LAION on a TDM exception, with an appeal expected. Like Company v Google Ireland Limited (Case C-250/25) is a pending preliminary reference to the CJEU. The GBA follows EDPB Opinion 28/2024 on AI models for the GDPR side of the question.

The four risk layers for your business

Layer 1. Does it reproduce a specific recognisable existing work?

Usually not. Sometimes yes. Diffusion models occasionally produce near-copies of training images, and prompts that name a living artist, a recent film or a protected comic character raise that risk significantly. A prompt for "a friendly dentist practice" is low-risk. A prompt for "in the style of [living photographer]" is not. A reverse image search on hero images and ad visuals usually finds any pre-existing originals within a minute.

Layer 2. Does it contain a brand, logo or watermark?

This is the layer the Getty ruling was really about. Stable Diffusion produced images with visible Getty and iStock watermarks, and the UK court held that this amounted to trademark infringement on those specific outputs. The same logic applies under Belgian trademark law (Benelux Convention on Intellectual Property) if your AI image contains a recognisable brand logo, sports-club logo, film-studio mascot or stock-watermark fragment.

Watermarks are the easiest mistake to spot. Zoom in to 200% in the corners of every hero image before publication. If a demand letter does arrive, the response pattern follows the same shape as for ordinary stock-photo claims.

Layer 3. Does it contain recognisable real persons?

A photograph of a recognisable real person is personal data under Article 4(1) GDPR. An AI image that strongly resembles a recognisable real person triggers the same controller obligations under Article 5 GDPR. Belgian portrait rights are grounded in Article 22 of the Constitution (privacy), Article 8 ECHR and the case law around image rights. The person depicted can object to publication absent express consent. The GBA regularly intervenes in publications without consent.

The practical rule is the same as for ordinary photography. Want to publish an image that resembles a specific real person? You need consent. AI does not change that rule. Without consent, explicitly prompt for "a fictional person not resembling any existing individual" and verify the output before publication.

Layer 4. Do you own the copyright in what you made?

Probably not, in practical terms. The US Copyright Office held in March 2023 that pure AI output without sufficient human creative input is ineligible for copyright protection. The Infopaq test (Case C-5/08) requires "an author's own intellectual creation". Pure prompt output usually fails that threshold.

That ownership gap is less bad than you might think. Most SMBs use marketing images to communicate something, not to license. The practical consequence is that any competitor can reuse the same image. The wider question of who is liable when AI helps build a Belgian website is more layered, and the answer is rarely "the AI vendor".

Article 50 of the AI Act, applicable from 2 August 2026

The EU AI Act applies from 2 August 2026 for the transparency obligations in Article 50 of Regulation (EU) 2024/1689. Two paragraphs touch your website images.

Article 50(2) imposes a marking duty on the provider of a generative AI system. Midjourney, OpenAI, Stability and the rest must mark their output machine-readable as artificially generated. That is not your obligation as website owner.

Article 50(4) imposes a labelling duty on the deployer (you) if the AI content is a deepfake under Article 3(60). The definition is narrow.

An Antwerp real-estate agent who produces "virtual staging" of an actual property in a way that could pass for a photograph of that specific room sits in the red quadrant and needs an "AI-generated" label. A bakery placing a generic loaf on the menu page does not. The maximum fine under Article 99 is the higher of EUR 15 million or 3% of worldwide annual turnover, but Article 99(6) directs member states to consider lower fines for SMBs. Realistic Belgian enforcement against an SMB will sit far below those maxima. Our full AI Act guide for website owners covers the other paragraphs of Article 50.

A provisional agreement of 7 May 2026 on a Digital Omnibus for AI may give providers of pre-existing generative AI systems a transition window to 2 December 2026 for the marking under Article 50(2). The application date of Article 50 itself (2 August 2026) does not shift.

Practical rules for your business

A short checklist for every Belgian SMB using AI imagery on the website.

1. Avoid "in the style of [living artist]" prompts. Raises copyright risk without much creative gain.
2. Zoom in to 200% in the corners of every AI image before publication and look for stray watermarks, brand logos and signature fragments.
3. If you need an image of a person, prompt for "a fictional person not resembling any existing individual". If you want a real person, get written consent. Belgian portrait rights do not vanish because AI made the image.
4. Keep your prompt logs. A folder with "prompt + date + tool" is cheap insurance if a question ever comes up.
5. If you publish an image that looks authentic of a person, place or event, label it as AI-generated. A short caption suffices and keeps you safe under Article 50(4).
6. Do not claim copyright in AI images in marketing contracts. You may have nothing to assign.
7. For important images (campaign hero, packaging, legal documents) use a licensed library or a human illustrator. Our recommendations for safe free stock photo sources are a starting point.

Our free compliance scan tests GDPR, cookies, accessibility and image rights. It does not yet check directly whether AI images are legitimate. AI image checks are on the roadmap.

What this article does not tell you

It does not tell you whether the AI vendor was allowed to use the training data. That question is open in the EU and is being fought out in Germany, in California and at the CJEU. It does not tell you that an AI image is automatically safe. Each of the four layers can go wrong. And it does not tell you what happens if your AI vendor is held liable for its training. The pass-through to you via their terms is untested.

The other cluster articles cover related questions. AI website liability under Belgian law sits next to this article.

Frequently asked questions

Can I get a Getty-style demand letter for an AI-generated image?

It can happen, but only if the AI output reproduces a recognisable existing photograph or carries a Getty or iStock watermark. Getty v Stability AI confirmed this in November 2025. For ordinary AI marketing images without watermarks and without recognisable real subjects, the risk of a demand letter is small.

Do I own copyright in the AI image I made for my website?

Probably not. The CJEU Infopaq test requires an author's own intellectual creation. Book XI of the Belgian Code of Economic Law follows the same line. A competitor can reuse the same image. That is more a marketing concern than a legal risk.

Do I have to label AI images on my website under the AI Act?

Only if the image is a deepfake under Article 3(60), meaning an image that depicts a real person, place or event and could come across as authentic to a reasonable viewer. The Article 50(4) labelling duty applies from 2 August 2026. Generic AI marketing imagery without recognisable real subjects falls outside.

What does the Belgian DPA say about AI images of real persons?

The GBA follows EDPB Opinion 28/2024 on AI models. AI images in which a real person is recognisable are personal data under Article 4(1) GDPR. Publishing them on a website without consent triggers the same controller obligations as ordinary photography.

Is Midjourney safer than DALL-E or Stable Diffusion for business use?

No relevant legal difference today. All three rely on comparable training-data practices and their terms of use place the risk on the user. Choose on quality and licence terms, not on supposed safety.

Further reading

The AI cluster pieces this article touches:

• AI website liability under Belgian law. The hub article on GDPR, EAA and cookie liability when AI helped build.
• The AI Act for website owners. Article 50 in detail, including chatbot disclosure and the editorial-control exception.
• AI-generated code and open-source licences. The sister article for code rather than images.
• The Product Liability Directive 2024/2853 in Belgium, applicable 9 December 2026.

This article is technical analysis, not legal advice. The author is not your lawyer and not your registered controller. For a binding view, speak with one of those two.