Do you have to label visualizations as AI? What does the AI Act really say and what changes on August 2, 2026?

Publication date: 01-08-2026  |  Update date: 01-08-2026  | Author: Piotr Kurpiewski

From August 2, 2026, EU regulations (AI Act) governing the labeling of content generated by artificial intelligence come into effect. What does this mean in practice for the architectural and 3D visualization industry? In this article, we break down the creators' obligations and explain why a conceptual render of an unbuilt object is not a deepfake, while virtual staging and photo manipulation are. You will learn how to correctly label materials, who bears legal responsibility in relations with the developer, and why your existing portfolio is completely safe. A mandatory guide for anyone using AI tools in their daily design pipeline.

If you create architectural visualizations, the answer to the most important question is: you do not have to label a visualization of an unbuilt project when presented as a design. The obligation concerns deepfakes, and your concept is not one. It only begins where AI touches something that truly exists, and that is exactly the boundary I will show you in the rest of this article.

What comes into effect on August 2, 2026?

Article 50 of the AI Act gathers four transparency obligations: informing that you are interacting with AI, labeling synthetic content, disclosing emotion recognition systems, and revealing deepfakes. From your perspective, only the last two matter, or rather one.

The EU legislator made several last-minute shifts. The Digital Omnibus, i.e. Regulation 2026/1744, postponed the provisions on high-risk systems to December 2027 and August 2028. However, Article 50 remained untouched and enters into force on the original date. The only concession concerns suppliers of tools that hit the market before August 2, because they have until December 2, 2026 to implement machine labeling.

Which of these obligations apply to you? It depends on which side of the table you sit.

Supplier or user, i.e. who is responsible for what

The regulation divides everyone into two groups. Supplier develops an AI system and makes it available on the market under its own name, such as OpenAI with GPT Images, Google with the Nano Banana model, or any AI rendering app. User uses such a tool in professional activity. You are the latter (unless Sam Altman is reading this, in which case I’m not greeting him).

The split has very practical consequences. Machine labeling, i.e. invisible watermarks and metadata in the file, is an obligation of the supplier under Article 50(2) and you don’t have to worry about it. You are only concerned with Article 50(4), i.e. disclosing deepfakes. A sole proprietorship counts the same as a large firm, because only purely personal non-professional activity is excluded. A holiday card for your brother-in-law is safe, but a portfolio is not.

In the European Commission’s guidelines to Article 50, there is one sentence hardly discussed in the industry but which determines liability in outsourcing. A company that merely commissions the creation of material and does not decide if and how the contractor uses AI is not a user. So if you make visualizations for a developer and choose the tools yourself, the obligation lies with you, not with them. However, this does not mean that the developer is completely freed from responsibility for what they publish.

If a client publishes a deepfake generated by you (e.g. an unlabeled virtual staging of an existing empty apartment), they might avoid sanctions under the AI Act itself, but they will immediately expose themselves to penalties from, for example, the UOKiK for misleading consumers under rules on unfair commercial practices. If an authority comes after the client, they will almost certainly come back to you with a recourse claim. That’s why it’s so important to regulate these matters in the contract and to clearly communicate what the image represents—especially for materials that go directly into the sales offer.

What is a deepfake according to the AI Act?

The definition in Article 3(60) is broader than intuition suggests. A deepfake is content generated or manipulated by AI that resembles existing people, objects, places, entities, or events and which the recipient could wrongly assume to be authentic. As you can see, it’s not just about politicians and celebrities.

The Commission broke down this definition into four criteria that must all be met: similarity, existence of the represented object, belonging to one of the listed categories, and apparent authenticity. And here’s the bad news. The Commission defines objects as realistic material items, explicitly mentioning buildings, and only requires similarity to something that could probably exist. Your photorealistic render meets the first three criteria without any problem. The whole matter therefore boils down to the fourth criterion.

Why a conceptual visualization is not a deepfake

The guidelines say that if content is placed in a context where a typical recipient does not expect it to be real, it does not appear falsely authentic, even if it does not correspond to reality. In such cases, the fourth criterion simply isn’t met.

Think of a scale model. Cardboard models have been displayed in architectural offices for decades, made with great care, with window lights and sponge trees. Yet no one thought to attach a “this building does not exist” plaque, because the context says it all: the model sits on a table in an office, at a 1:200 scale. A photorealistic render shown as a design works exactly the same way, just in a different format.

The Commission itself uses the film production analogy. Generated backgrounds, special effects, and post-production do not make a film appear authentic, because the viewer knows they are watching fiction. Although the guidelines add that high photorealism increases the likelihood of considering content a deepfake, they explicitly point out that photorealism itself is not decisive.

And what about people, cars, and dogs that you add to the scene? “Persons” in the regulation also include realistic AI-generated avatars, so they technically fall into the category. However, the fourth criterion still needs to be met separately, and a character walking through a visualization of an unbuilt plaza is not impersonating anyone or recounting anything.

Keep in mind, however, that the Commission’s guidelines are non-binding, and only the Court of Justice of the EU can issue a binding interpretation of the AI Act. This is an interpretation firmly rooted in the text, but not a judgment.

When must a visualization be labeled?

The boundary lies where AI touches something that truly exists. In practice, this involves at least four situations:

• virtual staging of an empty apartment listed for sale,
• enhanced greenery or facade in a photo of a home from a listing,
• a massing model inserted into a photograph of an existing street,
• a renovation visualization overlaid on the current state of a building.

In each of these cases, the recipient has every right to assume they are viewing a record of reality. The Commission even lists an ad image of a product that may mislead about its actual appearance, features, or use among deepfakes. An apartment in a sales listing is precisely such a product. On the list of content that are not deepfakes, you will find a real product shown on a generated background, provided the ad does not mislead about the product itself. This yields a simple working rule: AI in the background is usually allowed, AI on the offered item requires labeling.

How to label? Visibly and at first contact with the material. Metadata in the file is not enough, because the guidelines explicitly state that the user cannot rely on machine labeling by the supplier. A person must see the label without any tools. A footnote in the terms of service or a grey footer at the bottom of the page is also insufficient. The usual formula works: “Conceptual visualization. The material does not represent the current state of the object. Certain elements were generated using AI.”

Warning! A second regime applies to sales materials that is easy to forget. Retouching real estate photos can be an unfair commercial practice regardless of whether you did it with generative AI or manually in Photoshop. Not even V-Ray saves you here.

What the AI Act does not cover

A classic render is outside the scope and there is no debate here. The Commission’s guidelines explicitly list “rendered frame” as content not subject to the labeling obligation. A rendering engine does not create synthetic content under the regulation—it simply calculates light in a scene you built yourself.

So what about AI detectors that regularly take 3ds Max renders for generator images? They have no legal significance. The obligation depends on how the material was actually produced, not on what a detection tool outputs, so a false positive triggers nothing. However, if a client starts asking questions, the C2PA standard—an origin signature embedded in the file—comes to the rescue. A signed render defends itself.

AI used inside the production pipeline is also out of scope. Denoising, upscaling, or texture generation in a closed production process require nothing, because the guidelines mandate labeling only the final material. And one thing that relieves the most stress: visualizations produced before August 2, 2026 do not require retroactive labeling.

Penalties for missing labeling and who enforces it in Poland

The numbers are impressive. Violating Art. 50 carries a penalty of up to EUR 15 million or 3 percent of global turnover, whichever is higher. For small and medium-sized enterprises, the reverse rule applies, using the lower amount. However, these are caps for the most serious infringements, not a tariff for a one-person studio.

In Poland, oversight is carried out by the Commission for the Development and Security of Artificial Intelligence, abbreviated as KRiBSI. The chair will not be appointed until October, and the first session is scheduled for November 2026, so actual enforcement will begin later than the regulations themselves. KRiBSI also issues individual opinions upon request, and that is the simplest way if you want to have in black and white how the authority views your specific case.

Sources on which this article is based

Legal status as of the publication date. All references to the guidelines refer to paragraph numbers in the Commission's document of 20 July 2026.

• Regulation (EU) 2024/1689 (AI Act) – Art. 3 points 3, 4 and 60, Art. 50, Art. 99, Art. 113 and recitals 133 and 134
• European Commission Guidelines on Transparency Obligations under Art. 50 of the AI Act – § 5, §§ 10–16, § 64, § 68, §§ 113–116, § 153 and § 154
• AI Act Service Desk, Art. 50 – the Commission's official page with the text of the provision and information on changes
• Directive 2005/29/EC on unfair commercial practices – a parallel regime to the AI Act
• Regulation 2026/1744 (Digital Omnibus) in EUR-Lex
• Act on Artificial Intelligence Systems in the Journal of Laws