Table of Contents(17)
Since 2 August 2026, EU AI Act Article 50 has required professionals to visibly label published AI-edited images that would falsely appear authentic. Providers must add machine-readable marks; tools on the market before then have until 2 December 2026.
Free AI Virtual Staging — Stage Any Empty Room in Under a Minute
Stage vacant rooms with AI, usually in under a minute. 13 design styles, 15 room types, listing-ready photos for real estate, rentals and hospitality. 6 free credits, no credit card required.
EU AI Act Article 50 in Short
Article 50 of the EU AI Act, Regulation (EU) 2024/1689, sets transparency rules for AI systems that talk to people or generate content. For AI-edited images, it splits the work in two:
- Professional users ("deployers") label. Since 2 August 2026, a business or professional using AI to create or edit an image that counts as a "deep fake" must disclose that it was artificially generated or manipulated. The Act defines a deep fake as AI-generated or manipulated content that resembles existing people, objects, places, entities or events and would falsely appear authentic or truthful. The disclosure must be clear and come at the latest when people first see the image. A virtually staged photo of a real apartment can meet that definition.
- Tool providers mark. Providers of generative AI systems must mark their outputs in a machine-readable format and make them detectable as AI-made. For systems placed on the market (first made available on the EU market) before 2 August 2026, this duty applies from 2 December 2026.
- Invisible marks do not replace the label. The European Commission's Guidelines on Article 50 say deployers "cannot rely on the machine-readable marking embedded in the content by the provider".
- What to do now. Find the AI-edited images that would falsely appear authentic, label each one visibly, say what changed and keep the original. The checklist lists every step, and Roomagen's "AI label on downloads" setting can add an "AI" badge to your downloads.
This guide is for estate agents, property photographers, interior designers, architects, hotels and holiday rentals, furniture and e-commerce teams, and developers using image APIs. It covers images only: Article 50(3), on emotion recognition and biometric categorisation, and the rules for AI-generated text are outside its scope.
About the images in this guide: the cover image is AI-generated, and the example image further down is AI-edited.
What Article 50 Requires for AI Images
| Paragraph | What it requires | Who | Applies from |
|---|---|---|---|
| 50(1) | Tell people they are interacting with an AI, unless that is obvious | Providers of chat assistants and other interactive systems | 2 August 2026 |
| 50(2) | Outputs "marked in a machine-readable format and detectable as artificially generated or manipulated" | Providers of generative systems | 2 August 2026, or 2 December 2026 for systems placed on the market before 2 August 2026 |
| 50(4) | "Disclose that the content has been artificially generated or manipulated" when it is a deep fake | Deployers | 2 August 2026 |
| 50(5) | Give this information "in a clear and distinguishable manner at the latest at the time of the first interaction or exposure" | Providers and deployers | 2 August 2026; for marking information, the Article 50(2) dates |
- Tools versus chat. The Commission's Guidelines list "AI-assisted product visualisation (e.g., virtual try-on, room staging, product rendering)" among systems that do not interact directly with people, so Article 50(1) concerns chat assistants rather than staging tools.
- Standard editing. The marking duty does not cover "an assistive function for standard editing", such as minor cropping or colour correction. Removing, replacing or inserting objects in a way that changes an image's meaning needs marking.
- Status. On 20 July 2026 the Commission approved the content of its Guidelines on Article 50 and said it would formally adopt them once all language versions are available. They are non-binding. The AI Act's Article 50 Code of Practice, officially the Code of Practice on Transparency of AI-Generated Content, is voluntary: the Commission says not signing it "does not constitute non-compliance with the AI Act". In July 2026 the Commission and the AI Board assessed the Code as adequate; signatories can rely on its measures to demonstrate compliance, while non-signatories must show that their own measures are adequate.
Providers and Deployers: Who Does What
- Provider: whoever develops an AI system, or has one developed, and offers it "under its own name or trademark, whether for payment or free of charge" (Art. 3(3)). When you edit images with Roomagen, Roomagen is the provider, and Google's Gemini image models generate the results. According to the Guidelines, a provider may rely on its model provider's marking to the extent that this marking is compliant, but stays responsible for showing that it complies.
- Deployer: whoever uses an AI system "under its authority", except in "a personal non-professional activity" (Art. 3(4)). An agency, photographer, designer, hotel or shop that uses AI image tools for work is a deployer.
Special cases:
- Personal use. A homeowner trying furniture ideas for their own home has no deployer duty (Art. 2(10)), but the provider still marks the output.
- Chains. A company stays the deployer when freelancers operate the system on its behalf and under its control. When images pass from photographer to agency to portal, the Guidelines expect whoever applied the label to take "proportionate measures" so it reaches the audience, for example by contract.
- Portals. Platforms that only host other people's AI content are not deployers of it, although the Guidelines strongly encourage them to preserve marks and labels.
- Outside the EU. The Act covers providers and deployers elsewhere "where the output produced by the AI system is used in the Union" (Art. 2(1)(c)). The Guidelines count "posting deep fakes on the globally accessible internet" as foreseeing such use, but not content that reaches the EU through channels a deployer can neither foresee nor control.
- API integrators. Under the Guidelines, a company that modifies an existing generative system and puts it into service under its own name becomes the provider of the new system. For products built on a third-party image API, such as Roomagen's image API, this depends on the facts.
- EEA. As of 24 October 2026, Iceland, Liechtenstein and Norway were still reviewing the AI Act for incorporation into the EEA Agreement.
Key Dates: 2 August and 2 December 2026
| Date | What happens |
|---|---|
| 1 August 2024 | The AI Act enters into force |
| 27 July 2026 | The Digital Omnibus on AI, Regulation (EU) 2026/1744, enters into force |
| 2 August 2026 | Article 50 applies: chat notices, deep-fake disclosure, and marking for systems placed on the market from that day |
| 2 December 2026 | Marking and detection for generative systems placed on the market before 2 August 2026 (Art. 111(4)) |
| 2 February 2027 | Watermark-detection interoperability, for Code of Practice signatories only |
- 2 February 2027 is not the marking deadline. The Commission proposed it in November 2025, but the Omnibus adopted on 8 July 2026 set 2 December 2026, only for the marking duty of systems already on the market.
- Older content. Images generated before 2 August 2026 do not need to be labelled retroactively, although the Commission encourages it where possible.
Who Enforces Article 50, and What Are the Fines?
Article 50 is mainly enforced by national market surveillance authorities; the EU AI Office has a limited role. Fines can reach EUR 15 million or 3% of worldwide annual turnover, whichever is higher (Art. 99(4)(g)), but for SMEs, including start-ups, the lower of the two applies (Art. 99(6)). Member States set the penalty rules, which may also include warnings (Art. 99(1)), and anyone with grounds to suspect an infringement can complain to a market surveillance authority (Art. 85).
When Is an AI-Edited Image a "Deep Fake"?
The EU AI Act's deepfake definition, spelled "deep fake" in the Act, is in Article 3(60). It covers "AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful". The Guidelines add:
- Buildings count. "Objects" include buildings and consumer goods, and "places" are realistic locations. Resembling something that "can plausibly exist" is enough for the "existing" test; the image must also falsely appear authentic or truthful.
- Context decides. The test is objective and looks at the image, where it is shown and what the audience expects. No intent to deceive is needed, and photorealism alone is not decisive.
- Small edits may not count. Removing a passer-by, lighting adjustments, colour correction and noise reduction may not create a deep fake, but this "depends on the context". In product ads, replacing a background "for clearly aesthetic purposes" is also likely to have only a minor impact.
- Product images. An AI image of a product in an ad that misleads about how it really looks, for example by making it "more appealing or with improved quality than in real life", is a deep-fake example. A real product against an AI-generated background is not, "as long as the ad is not likely to mislead".
The Guidelines do not discuss property listings, but the Commission's page on the EU labelling icons uses authentic photographs of an empty apartment furnished using AI as an example for its "Partially AI-Modified" icon. The same page says: "Not all AI-generated or manipulated content needs to be labelled."
Listings are unlikely to count as creative works. Article 50(4) allows a lighter disclosure for "evidently artistic, creative, satirical, fictional or analogous" works. The Guidelines exclude works "mainly motivated by functional or technical considerations" and content that is "exclusively informative or commercial", so listing, booking and product photos are unlikely to qualify.
Which AI Edits Need a Visible Label?
Virtual staging is the clearest case: the Commission's own material on the EU AI Act uses it as an example. Other edits are less settled. "Likely" means the edit can make an image of a real property or product look authentic or truthful when it is not.
| Edit | Label under Art. 50(4)? | Why |
|---|---|---|
| Virtual staging of an empty room | Likely | The Commission's icon example is an empty apartment furnished with AI |
| Emptying a room, removing objects or decluttering | Likely when the room looks different; unclear for small clean-ups | An emptied or decluttered room shows a state that does not exist; removing a passer-by is a minor-edit example in the Guidelines |
| Renovation previews and landscaping of a real property | Likely when shown as the property; less likely when clearly presented as a proposal | The image shows a state that does not exist; context decides |
| Day-to-dusk and sky replacement | Unclear | Lighting changes and aesthetic background replacements in ads are minor-edit examples, but an ad image that misleads about how a product really looks is a deep-fake example; for a property, the sky and view can be part of what is sold |
| Exposure, colour and clarity fixes | Unlikely in most contexts, if nothing else in the image changed | Lighting adjustments, colour correction and noise reduction are minor-edit examples |
| 3D plans, redrawn sketches and plan renders | Unclear | A drawing rarely passes as an authentic photo, but a plan or render presented as the real layout could falsely appear truthful if AI changed it; photorealism makes a render more likely to count |
| Furniture and product images | Likely if the product looks different from reality; unlikely for a real product on an AI background | Both are examples in the Guidelines |
If in doubt, label. A short, accurate label costs little, while an undisclosed deep fake published for work breaches Article 50(4). A designer or architect showing a proposal to a client who knows it is one is in a different position from an agent publishing the same image as a listing photo. The Guidelines give no example for design proposals, but they say content shown to an audience that does not expect it to be authentic may not be a deep fake. So present a proposal as a proposal wherever it may be shown (see our guide to AI tools for interior designers).
How to Meet the EU AI Act Labelling Requirements
Under the EU AI Act, image labelling is the deployer's job, and the Act sets the result rather than the format.
- Each deep fake, each viewer. The disclosure must be clear and distinguishable at the latest when someone first sees the image (Art. 50(5)). The Guidelines apply this to each output and each person who sees it, so label each image that needs a label, not just the listing as a whole.
- Visible, not buried. It should be perceivable without tools; text hidden in terms of use, manuals or menus does not count.
- A label inside the image travels. A caption may work on your own website, but a label inside the image stays with the file when it is reshared or downloaded. The Commission says its EU icon "must be visible when content is reshared or downloaded".
- "AI" in a corner is one option. The Code of Practice asks signatories to build labels around the capitalised acronym "AI" (in the national language where language laws rule out English), with the top-right corner as an example placement. The EU icons are optional, and using them "does not establish legal compliance by itself".
Say what changed in plain words. The Commission reports that its icon did better in user tests with a text label such as "modified". For example: "Virtually staged with AI", "AI-edited: furniture removed", "AI visualisation of a proposed renovation" or "Sky edited with AI". Repeat the disclosure in the caption or alt text for screen-reader users, and see our disclaimer wording examples for more.
Keep the unedited original, ask portals and clients to keep your label, and check that cropping does not cut it off.
Other rules still apply. Independently of the AI Act, the EU's Unfair Commercial Practices Directive (2005/29/EC) treats a trader's practice as misleading if it deceives, or is likely to deceive, the average consumer about a product's main characteristics. The practice must also lead, or be likely to lead, the consumer to a decision they would not otherwise have taken. The Directive's definition of "product" includes immovable property. Complying with Article 50 does not make an AI image lawful under other law (recital 137), so a label cannot fix a photo that hides damage or invents a view. Outside the EU, rules such as California's AB 723 differ; see our AB 723 guide and disclosure rules by country.
Machine-Readable Marks: C2PA, IPTC and SynthID
The EU AI Act's watermarking requirements are technology-neutral: Article 50(2) asks for marks that software can read and detect, and recital 133 lists watermarks, metadata, cryptographic provenance methods, logging and fingerprints as examples. Providers must also make their solutions "effective, interoperable, robust and reliable as far as this is technically feasible" (Art. 50(2)). Neither the Act, the Guidelines nor the Code of Practice names C2PA, IPTC or SynthID. The Guidelines allow a single technique or a combination, as long as the overall solution is machine-readable, detectable and meets those four requirements. The Code of Practice asks its signatories for at least two layers on images, signed metadata and an imperceptible watermark, for as long as no single technique can meet the requirements on its own.
| Mark | What it is | Where it lives | What can remove it |
|---|---|---|---|
| C2PA Content Credentials | A signed manifest that records how the file was made | File metadata | Re-encoding with a tool that does not support C2PA; any such change makes validation fail |
| IPTC Digital Source Type | A value from IPTC's vocabulary, such as "Created using Generative AI" or "Edited using Generative AI" | XMP metadata | Saving the file without its metadata |
| SynthID | Google's invisible watermark, added "the moment content is created" (Google DeepMind) | The pixels | Google says it is designed to stand up to cropping, filters and lossy compression |
Google says images "created or modified using the listed models are automatically digitally signed by Google LLC", and that "any modification to a C2PA-compliant media file using a non-C2PA tool is considered tampering" (Google Cloud). The IPTC value is not signed, so it is easy to read and easy to lose.
How to Check an Image for AI Marks
Check the original file: tools that re-encode, resize or edit an image often drop its metadata marks.
With the free ExifTool, on a labelled Roomagen download:
$ exiftool -XMP-iptcExt:DigitalSourceType -XMP-xmp:CreatorTool result-ai-label.jpg
Digital Source Type : http://cv.iptc.org/newscodes/digitalsourcetype/compositeWithTrainedAlgorithmicMedia
Creator Tool : Roomagen
In IPTC's vocabulary, compositeWithTrainedAlgorithmicMedia means "Edited using Generative AI" and trainedAlgorithmicMedia means "Created using Generative AI". On a file that keeps Google's Content Credentials, ExifTool can also show what they say ("Created by Google Generative AI", even on a photo edit), but it does not check their signature.
With c2patool or a browser. To check C2PA metadata and its signature, run c2patool from the open-source c2pa-rs project with the official C2PA trust list:
$ c2patool result.jpg trust --trust_anchors https://raw.githubusercontent.com/c2pa-org/conformance-public/main/trust-list/C2PA-TRUST-LIST.pem
A file that keeps Google's credentials reports "Trusted"; a labelled copy reports "No claim found". Without a command line, upload the file to Content Credentials Verify.
With Google's tools. Since 7 October 2026, Google's SynthID Detector has been open to everyone, in English, for images, video and audio made with AI from Google or its partners. The Gemini app can also check a file. According to Google, if no SynthID watermark is detected, the image was not created or edited by Google AI, "but it could have been created by other AI systems".
A missing mark proves little: metadata is easily lost, and some tools add no marks.
What Roomagen Does to Support Labelling
Roomagen supports EU AI Act labelling with machine-readable marks, an optional visible label and notices in the app. The visible disclosure itself stays with you as the deployer.
Machine-readable marks in the file
- SynthID. Results are generated by Google's Gemini image models, which, according to Google, add a SynthID watermark to every image when they generate it.
- Content Credentials. A result that Roomagen saves exactly as Google's model returned it keeps Google's signed Content Credentials. When Roomagen has to write a result as a new file, those credentials cannot come along. This happens, for example, when a result is scaled to the size of your original photo or cropped back to the frame of your floor plan.
- IPTC tag. Since 24 October 2026, every result that Roomagen writes as a new file carries the IPTC Digital Source Type tag, with "Roomagen" as the creator tool and no model name, account data or date. In such files, images made from a plan or sketch with 3D Dollhouse, Visualize Floor Plan, Digitize Sketch or Colorize Floor Plan are tagged "Created using Generative AI". Results of the other tools, which edit a photo, are tagged "Edited using Generative AI".
The "AI label on downloads" setting
The setting sits on the Settings page, directly below Appearance. It is off by default and saved on your account, so it applies on every device.
With the setting on:
- Badge. Every result you download gets a small "AI" badge in its top-right corner: white letters on a semi-transparent dark plate, about 5.5% as high as the image's shorter side. It reads "AI" in every app language and says nothing more, so say what changed in your caption or listing text.
- Coverage. Downloads from the result screen, the Library, Design Companion, Refine and batch ZIPs are labelled. A single labelled download gets "-ai-label" before the file extension; files inside a ZIP keep their names.
- Originals. Your originals in the Library stay unchanged, the labelled copy is not stored, and your own uploaded photos are never labelled.
- Marks. The labelled copy carries the badge and the IPTC tag, but not Google's Content Credentials, because it is a new file. For the unlabelled original, switch the setting off and download again.
- Download buttons only. An image saved another way, such as right-click "Save image", is the unlabelled original.
According to the Guidelines, providers may add perceptible labels as a complementary measure for deployers. The badge is Roomagen's own design, not the Commission's EU icon.
Notices in the app
- EU and EEA card. Users in the EU, Iceland, Liechtenstein and Norway see a one-time card, "Publishing in the EU?", on the dashboard and the result screen, with "Turn on AI label" and "Not now". It blocks nothing, appears only while the label is off and stays away once answered. Users in other countries, including the UK and Switzerland, do not see it.
- Design Companion. The chat editor starts every conversation with "Design Companion is an AI assistant. Its replies and images are generated by AI." The line cannot be dismissed and appears in all nine app languages.
- Terms of Service. They forbid intentionally removing, altering or obscuring "the marks embedded in outputs that identify them as AI-generated or AI-modified", and presenting AI images as unedited photographs where that would mislead. They also ban intimate images of real people made without consent, and child sexual abuse material. They note that EU publishing may require a visible disclosure by the user.
- API. API results come without the "AI" label; the API documentation's AI content marks page describes Google's marks on them and what removes them.
What Roomagen does not do
- It does not decide whether an image needs a label, and it adds no visible "AI" label unless you turn the setting on.
- Turning the label on does not by itself make a publication compliant; your disclosure must be clear where you publish.
- It does not sign its own Content Credentials: files that Roomagen writes anew carry the unsigned IPTC tag.
Checklist Before You Publish in the EU
- Find the AI images that would falsely appear to be authentic photos of a real place, building or product.
- Label each one visibly, on or next to the image, wherever people first see it.
- Say what changed in plain words, and repeat it in the caption or alt text.
- Put the label inside the image if it will be reshared; in Roomagen, turn on "AI label on downloads" in Settings.
- Keep the unedited original.
- Check that the image is fair: a label does not excuse a misleading picture.
- Keep the marks: publish files as downloaded where you can, and never strip metadata on purpose.
- Brief your partners to keep the label.
- Building on an API? Offer a visible label in your product, and check whether you are a provider yourself.
This Is Not Legal Advice
This guide reflects the EU rules and Roomagen's features on 24 October 2026 and is not legal advice. The Guidelines are non-binding, authorities and courts will shape how Article 50 is applied, and questions remain open, such as how the deep-fake test applies to listing photos. For a specific case, ask a lawyer where you publish.
Rules, marks and Roomagen features checked against the official EU texts, Commission pages, IPTC and Google documentation, C2PA tools, and Roomagen's code and live service on 24 October 2026. The cover image is AI-generated and the example image is AI-edited.
Ready to transform your listings?
Try Roomagen's AI virtual staging for free. Upload your first photo and see the result, usually in under a minute.
Start FreeSources & References
- 1.EUR-Lex – Regulation (EU) 2024/1689 (AI Act), Official Journal text: Articles 2, 3, 50, 99 and recitals 133-137
- 2.EUR-Lex – Regulation (EU) 2026/1744 (Digital Omnibus on AI) of 8 July 2026: new Article 111(4), 2 December 2026
- 3.European Commission – Guidelines on the transparency obligations under Article 50 of the AI Act, C(2026) 5054 (content approved 20 July 2026)
- 4.European Commission – Q&A: Transparency obligations under Article 50 of the AI Act (updated 24 July 2026)
- 5.European Commission – Code of Practice on Transparency of AI-Generated Content
- 6.European Commission – Q&A: Signing the Code of Practice on Transparency of AI-Generated Content (updated 29 July 2026)
- 7.European Commission – EU icons for labelling AI-generated content (updated 13 October 2026)
- 8.EUR-Lex – Unfair Commercial Practices Directive 2005/29/EC
- 9.EFTA – EEA-Lex factsheet: Regulation (EU) 2024/1689, under scrutiny for incorporation into the EEA Agreement
- 10.IPTC – Digital Source Type controlled vocabulary (trainedAlgorithmicMedia, compositeWithTrainedAlgorithmicMedia)
- 11.Google Cloud – Content Credentials (C2PA) for Google's generative models (updated 22 October 2026)
- 12.Google AI for Developers – Image generation with Gemini: "All generated images include a SynthID watermark"
- 13.Google DeepMind – SynthID
- 14.Google – Google expands SynthID Detector for AI content (7 October 2026)
- 15.Google – Gemini Apps Help: Verify AI-generated images, videos and audio
- 16.ExifTool by Phil Harvey
- 17.Content Authenticity Initiative – c2pa-rs with c2patool (GitHub)
- 18.Content Credentials Verify
Frequently Asked Questions
Written by
Roomagen Team
The Roomagen team creates in-depth guides about AI virtual staging, real estate photography, and property marketing strategies to help agents and professionals stay ahead.
