Article 50 of the EU AI Act is now in force, and it decides which AI images, videos and music tracks need a label. See who must mark files, who must disclose them, which dates still matter through December 2026, and how to label your own content without slowing down your workflow.
Every AI image, clip and track you publish in Europe now sits under a rule with a date attached. Article 50 of the EU AI Act became applicable on 2 August 2026, and the EU AI Act labeling requirements it contains decide whether a synthetic photo needs a visible tag, a hidden watermark, or both. If you make marketing visuals, run a blog, edit video or release music with generative tools, this is the part of the law that reaches your desk first.
The short version: companies that build generative models must mark what their systems produce, and the people who publish that output must disclose it in specific situations. The details around that sentence are where most teams get stuck, so this article walks through each piece in plain language, with dates, examples and a workflow you can copy.
💡 This article is general information, not legal advice. Check your own situation with a qualified lawyer before you rely on any reading of the rules.
Article 50 sits in the part of the AI Act about transparency. The idea is simple: people should be able to tell when they are dealing with a machine, or looking at something a machine made. It has several paragraphs, and most of them matter to anyone who creates images, video or audio.
Here is the split, paragraph by paragraph:
Two of these do the heavy lifting for creators. Paragraph 2 is a technical duty that sits with the toolmaker. Paragraph 4 is a visible duty that sits with whoever publishes.
The law uses two roles, and mixing them up is the most common mistake.
| Role | Who it is | Main duty |
|---|---|---|
| Provider | The company that develops a generative model or system and offers it | Mark outputs in machine-readable form and make them detectable |
| Deployer | Anyone using such a system in a professional setting: agency, publisher, brand, studio, business owner | Label deepfakes and public-interest AI text so people can see the disclosure |
Purely personal, non-professional use falls outside the deployer definition. A freelancer selling visuals to clients, a company blog or a YouTube channel that earns money is a different story. One organization can also be both: a studio that fine-tunes its own model and then publishes what it makes carries the duties of each role.
Timing is the part people ask about first, and it has a twist.
The transparency obligations in Article 50 apply from 2 August 2026. From that day, new generative systems placed on the market must mark their outputs, and deployers must label deepfakes and public-interest text. The Commission published its final Article 50 guidance close to that date, and national authorities together with the EU AI Office handle enforcement.
Generative systems that were already on the market before 2 August 2026 got extra time for the machine-readable marking duty in Article 50(2). Their deadline is 2 December 2026, which means that as of early October the grace period is still running for older systems.
| Date | What happens |
|---|---|
| 10 June 2026 | Commission publishes the final Code of Practice on marking and labelling AI-generated content |
| 2 August 2026 | Article 50 obligations apply, including deployer labels for deepfakes and public-interest text |
| 2 December 2026 | Grace period ends for machine-readable marking on generative systems already on the market before 2 August |
💡 According to the Commission's final Article 50 guidance, content created before 2 August 2026 does not need to be labeled retroactively. Your archive is safe. Your next publish is not.
The AI Act says what must happen but not exactly how. The Code of Practice on marking and labelling AI-generated content fills part of that gap. The Commission published the final version on 10 June 2026. It was drafted by independent experts in a multi-stakeholder process facilitated by the AI Office, and it is voluntary.
Signing it is meant to give providers and deployers a recognized way to show they meet the duties. They can also rely on other, equally adequate means, so the Code is a route and not a gate.
The Code starts from a blunt observation: no single marking technique can fully meet the AI Act's requirements. Signatories are asked to use at least two layers of machine-readable marking. In practice that means combining approaches such as:
Each layer fails in a different way. Metadata disappears when someone takes a screenshot. A watermark can weaken after heavy compression. Stacking them means one failure does not erase the trail.
Providers are also asked to offer detection tools so that deployers, users, authorities, researchers and media organizations can check whether a file is synthetic.
Machine-readable marks are invisible to the audience. The visible half of the system is the label. The Code addresses how and where to display it, including the use of a publicly available EU icon or an equivalent label. Early drafts floated a short two-letter tag such as AI, KI or IA depending on the language, plus a distinction between content that is fully AI-generated and content that is only AI-assisted.
Whatever the final design, the principles are consistent:
A badge in the corner does nothing for a podcast or a radio spot. Where a visual label is not feasible, such as audio-only content, the Code contemplates alternatives, including a short audio disclaimer at the start. A line like "This voice was generated with AI" counts for more than a long caption nobody hears.
The Act defines a deepfake as 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. Two tests sit inside that sentence: does it look like something real, and could a viewer believe it?
Some quick sorting:
When a piece sits on the line, label it. A tag costs a second of attention. A dispute costs far more.
Article 50(2) does not apply to systems that perform an assistive function for standard editing, or that do not substantially alter the input data. Exposure, color balance, cropping and dust removal belong in that bucket. Generating a new face, swapping a background with invented content or turning a photo into a different scene does not.
Upscaling is a gray area. A tool such as Topaz Image Upscale that only adds resolution sits close to standard editing, but if it invents details that were never in the original, treat the result as generated.
For content that is part of an evidently artistic, creative, satirical or fictional work, the deployer's duty shrinks. The disclosure only has to say that generated or manipulated content exists, and it must be done in a way that does not spoil the display or enjoyment of the work. The Commission's material describes this carve-out as one to read strictly, so it protects a film still or a satirical illustration, not an ad that borrows the look of a documentary.
AI-generated text has its own exception. If a human reviewed the text or an editor holds editorial responsibility for it, the public-interest disclosure does not apply.
Rules are easier to follow once you attach them to a real workflow. Here is how the duties show up for each format.
PicassoIA hosts a long list of text-to-image models, including Google Imagen 4, GPT Image 2, Seedream 5 Pro, Nano Banana Pro and FLUX 2 Pro. Each model provider decides how its own outputs are marked, so the invisible layer can differ from one model to the next.
What you control is the visible layer:
Remember that labels are about the audience, not the tool. A stock-style picture of an invented person in a coffee shop does not mislead anyone about a real event, while a lifelike photo of a named public figure at a protest does. Ask what a reasonable viewer would believe, then label to match.
Video raises the stakes because motion makes fakes more convincing. Models like Veo 3.1, Seedance 2.0 and Kling v3 produce footage that can pass for filmed material, including synchronized sound.
For clips that look like real footage, show the label in the first seconds and keep it readable, then repeat the disclosure in the description or caption. Use the platform's own AI toggle when it exists, but do not rely on it alone, because the legal duty sits with you.
Generated music and voice are easy to forget because they have no pixels. Lyria 3 Pro, ElevenLabs Music, Stable Audio 2.5 and MiniMax Music 2.6 all create full tracks from a prompt, and a voice model such as ElevenLabs v3 can narrate a script in seconds.
A generic instrumental track is not a deepfake, and the provider carries the marking duty. A voice that imitates a real person is different: that is deepfake audio, and the deployer needs a clear disclosure, which for audio means a spoken line or a visible note wherever the file is shown.
You do not need a compliance department to build good habits. You need three small ones.
Write down which model made each asset, the date and the prompt or source file. A shared spreadsheet is enough. If a client, platform or authority ever asks where an image came from, you answer in a minute instead of a week. It also helps you sort assets later into "deepfake risk", "clearly fictional" and "plain editing".
Do not remove watermarks, scrub metadata or run files through tools that wipe them just to make the file look cleaner. Deployers are expected to preserve what the provider embedded. If your pipeline compresses, resizes or converts files, test that the marks survive, and when they do not, rely on a visible label and a clear caption.
A label nobody sees is a label that does not work. Keep the wording short and plain:
| Format | Weak label | Strong label |
|---|---|---|
| Image | Tiny gray text in a footer | "AI-generated image" in the caption plus a visible corner tag |
| Video | A line buried in the description | On-screen tag for the first seconds, repeated in the caption |
| Audio | Nothing at all | A spoken line at the start: "This voice was generated with AI" |
| Article text on public matters | A note on a separate policy page | A visible line near the top, unless a human editor takes responsibility |
A quick pre-publish check works well as a routine:
Penalties for breaking the Article 50 transparency duties can reach EUR 15 million or 3% of total worldwide annual turnover, whichever is higher. For small and medium-sized businesses, the lower of the two figures applies. National market surveillance authorities and the AI Office handle enforcement, so a complaint can arrive from several directions.
The fine is only part of the picture. Platforms can remove unlabeled synthetic media, clients can cancel contracts, and a viral fake that carried no label can damage a brand faster than any regulator. Labeling is cheap insurance.
There is also a practical angle for teams that work with agencies. Contracts increasingly ask who labels what, who stores the originals and who answers if a regulator writes. Settle those questions in writing before a campaign launches, since the AI Act places duties on each role separately and a vague handoff leaves both sides exposed.
The best way to get comfortable with the rules is to practice them on real work. Open Picasso IA, pick a model, and run a full test: generate a photorealistic image, record it in your log, add a visible label and a caption, then repeat the exercise with a short video clip and a music track. Within an hour you will have a routine you can reuse on every project.
You can even use a language model such as Claude Sonnet 5 to draft short disclosure lines in several languages, then review them yourself before publishing. Try one image, one clip and one track this week, label each of them properly, and keep the log. Compliance gets easier every time you do it.
