How to Label AI Content According to the Law? Ready-Made Formulas for Designers and Archviz Studios.

Publication date: 01-08-2026  |  Update date: 01-08-2026  | Author: Mateusz Ciećwierz

From August 2, 2026, the AI Act imposes new rules for labeling content generated by artificial intelligence. How to avoid legal chaos and not ruin the aesthetics of your projects? In this practical guide, I show which archviz materials actually require a label and which are exempt from this obligation. Here you will find copy-ready labeling formulas for social media, developer materials, and real estate listings. I also explain why it is worth applying the C2PA standard in classic renders to protect against false alarms from AI detectors, and suggest how to secure yourself in one sentence in the contract against liability if the client removes the labels.

AI-generated content? In practice, a single sentence visible next to the image is sufficient, added no later than the moment the viewer sees this image for the first time. Metadata in the file will not replace this, and a footnote in the terms of service even less so. Below you will find ready-to-copy formulas, a list of materials that actually require a label, and the technical layer that is more useful in practice than what the regulations alone might suggest.

Before you label, check if you even have to

The obligation under Article 50(4) of the AI Act does not apply to everything that has passed through a generator. You check three things in order: whether generative AI was involved in creating the material, whether the material resembles something that actually exists or could exist, and whether it reaches a context in which the viewer expects a record of reality. Only three affirmative answers trigger the label.

Most often the second and third questions trip people up, because a visualization of an unbuilt project shown as a project does not pretend to be reality. I have elaborated on this topic in the text whether visualizations need to be labeled as AI-generated. Here we deal with what to do when the answer is “yes.”

Which materials require a label?

Where is the boundary exactly? Where AI touches something existing, and the dividing line is sharper than it seems. Labels are not needed for:

• conceptual renders of unbuilt projects, including photorealistic ones,
• classic renders from V-Ray, Corona or Cycles without the involvement of generative AI,
• denoising, upscaling and texture generation within a pipeline, because only the final material is labeled,
• visualizations created before August 2, 2026.

A label is needed wherever reality appears in the frame:

• virtual staging of an apartment listed for sale or rent,
• enhanced greenery, facade or interior in a photo of an existing building,
• a design inserted into a photograph of an existing street or plot,
• a renovation visualization overlaid on the current state.

Attention! There is one borderline case that happens very often in archviz. If you generate an interior for a specific, existing property and the material goes into an offer, you are already in the second group, even if the project itself is only a concept. What matters is what the viewer can expect, not the stage of the project.

What does a proper label look like?

The European Commission guidelines set two conditions. The information must be clear, meaning noticeable and understandable, and prominent, meaning easy to distinguish from the rest of the content. A label that can be overlooked under normal viewing conditions, hidden in instructions, under menu layers, or in a terms-of-service document that nobody reads does not meet this.

Think of the label next to a painting in a museum. It hangs beside the canvas at eye level and plainly states: oil on canvas, 1889. No one ever thinks of sticking that information to the back of the frame because then it would serve no purpose. Metadata in the file is exactly the back of the frame. It is necessary but does not replace the label, and the guidelines say this explicitly: the applying entity cannot rely on machine-readable marking implemented by the provider, because a person must see the label without resorting to any tools.

The second condition concerns timing. The information must reach the viewer no later than at the first contact with the material, and this applies to each viewer individually. With a single image, it is simple because the label is either in the frame or just beneath it. With video, things get more complicated, because the opening slide does not reach someone who starts watching midway. The guidelines indicate that, in such cases, a single label at the beginning is not enough and it needs to be supplemented at later points.

Ready-to-copy labeling formulas

There is no single mandatory formula, so readability counts more than legal precision. Four variants usually work best, depending on where the material ends up.

For a social media post, in the first line of the description—not in a comment and not after a text collapse: “Conceptual visualization generated using AI.” Remember, however, that platforms like Instagram, TikTok and LinkedIn have their own rules. The text alone in the post description is often not enough—you must also enable the system toggle “Created with AI” (or a similar setting) when publishing, so that the algorithm does not flag your material in an uncontrolled way or, worse, reduce your reach. For material delivered to a client, entered in the slide footer or below the frame: “Conceptual visualization. Material does not represent the current state of the object. Selected elements were generated using AI.” For a sales or rental listing, where the risk is highest: “Interior arrangement generated using AI. The photo does not represent the actual furnishings of the property.” For video, add a title card at the start and a persistent label in the corner of the frame with the short “Generated using AI.”

For foreign clients I keep an English version: “AI-generated visualisation. Conceptual image, does not represent the current state of the building.” Also remember that Article 50(5) refers to accessibility requirements, so contrast and size of the label matter. Six-point gray text on a light background is not a label, it's just an alibi.

A separate situation concerns materials with a clearly artistic, creative or fictional character. There, the obligation is relaxed and boils down to disclosure in a way that does not hinder the reception of the work, for example, in end credits. However, in the archviz industry this exception is extremely difficult to defend. Even the most visionary, artistically opulent visualization usually has a commercial purpose in our case—it sells a project, a service or an architect’s vision. It is safer to assume from the outset that sales material will simply not benefit from this exception.

Metadata, C2PA, and the layer you can't see

If the label fulfills the obligation, why bother with metadata at all? Because they solve another problem that the regulations do not address at all, and which is actually more bothersome in practice.

The C2PA standard, also known as Content Credentials, records in the file a cryptographically signed provenance trace: how the material was created and what happened to it afterwards. This is complemented by the IPTC field digitalSourceType with the value trainedAlgorithmicMedia for generated content. The guidelines explicitly allow such methods and note that the provider is not required to maintain a complete provenance chain but may use it. Visible marks are also not excluded as a supplement because they are meant to make your life easier.

Unfortunately, some platforms strip metadata when uploading a file, so treat this layer as reinforcement, not as a standalone protection. The visible label remains at all times because it is part of the image or description.

What about AI detectors?

This question comes up in every conversation about labeling and has a surprisingly simple answer. Detectors have no legal significance. The obligation depends on how the material was actually created, not on what a detection tool spits out, so a false positive on a 3ds Max render does not trigger anything and you do not need to explain it.

However, there remains the business problem, because a client who runs your visualization through a detector and receives a red alert will simply ask about it. Here comes something few think of: sign your classic renders with Content Credentials as well. That way, you invert the whole situation, because instead of having to justify the use of AI, you show proof that the material was created in 3ds Max. A signed render defends itself.

Archiving, contracts, and who labels upon publication

Materials created before August 2, 2026, do not require retroactive labeling. The Commission encourages supplementing labels on these materials, but it specifies that nobody expects a disproportionate effort like auditing your entire file database. You can leave your old portfolio published online alone. However, beware of commercial reuse. If you pull an unlabelled virtual staging from 2025 out of the archive and a client uses it as the main image in a new sales campaign, the regulator may treat it as a new share of content that now requires a label. By putting the file back into circulation, you are subject to current rules.

It is far more important to determine who presses the button. The guidelines make it clear that a company that merely commissions the creation of material and does not decide whether or how the contractor uses AI is not the applying entity.

In practice, this means that when commissioning a developer, the obligation falls on you, not the client. A listing service that merely distributes third-party content is also not an obligated entity, although the Commission encourages it to maintain the disclosures.

Therefore, it is worth including a single sentence in the contract: the contractor will label materials requiring disclosure pursuant to Art. 50(4) of the AI Act, and the client undertakes not to remove these labels upon publication. It costs one line but closes the most common scenario in which a properly labeled visualization ends up in the listing without the label, cropped to fit the portal’s format.

Finally, one more option. On June 10, 2026, the Commission published a Code of Practice on the transparency of AI-generated content, accompanied by a set of icons for labeling. Participation is voluntary and can serve as proof of compliance, and the icons themselves are simply convenient because they work without needing translation into other languages.

It turns out that the entire procedure fits into two habits: for materials concerning existing properties, you add a single sentence, and everything that comes out of the studio you sign with Content Credentials. The rest is just a matter of consistency.

I hope that labeling no longer seems like yet another bureaucratic requirement and now looks like what it is—a matter of two minutes at export. And if you work with developers or real estate agencies and want to know exactly where the line lies between acceptable photo editing and misleading manipulation, be sure to check out the article on virtual staging in real estate listings. You will find the most frequently asked scenarios laid out there.

Sources on which this article is based

Legal status as of the date of publication. References to the guidelines refer to paragraph numbers in the European Commission document of July 20, 2026. This article is informational and does not constitute legal advice.

• Regulation (EU) 2024/1689 (AI Act) - Art. 3 point 60, Art. 50 paragraphs 4 and 5, and Art. 99
• European Commission Guidelines on transparency obligations under Art. 50 of the AI Act - §14, §16, §§113–116 with example lists, and §§142–143
• Directive 2005/29/EC on unfair commercial practices
• Act of 23 August 2007 on counteracting unfair market practices - Journal of Laws 2007 No. 171, item 1206
• Act of 3 July 2026 on artificial intelligence systems - Journal of Laws 2026, item 1003