EU AI Act Article 50: You Have Three Days to Label Your AI-Generated Media

A few days ago, a friend who runs a brand sent me a screenshot and asked me what to do.

His company had used AI to generate a product intro video. The "spokesperson" in it looked strikingly real — the eyes, the mouth, the voice, all indistinguishable from a live shoot. The marketing team was ready to push it out across the major social platforms.

After watching it, I asked just one question: Did you label it?

He froze for a second.

Label what?

AI-generated content must be labeled. That's what Article 50 of the EU AI Act says. It takes effect August 2, 2026. No extensions.

My friend assumed this was something only big companies had to worry about. Wrong.

What the Law Actually Regulates

Let's be clear: this law doesn't sort AI into "allowed" and "not allowed." It classifies AI systems into risk tiers. Most marketing scenarios land in the "limited risk" tier.

What does limited risk mean?

It means you're not banned — but you have to be transparent.

The legislators' logic is dead simple: when someone scrolls through their phone, watches a video, reads a post, or chats with a customer-service agent, they have the right to know whether they're dealing with a real person or a machine. Otherwise? Identity theft, disinformation, fraud — all of it follows.

So Article 50 does one thing: it requires anyone using AI to let the people affected see at a glance that this is AI.

That simple. And that hard.

The Three Traps Marketing Teams Hit Most Often

Let me unpack the triggering scenarios. See how many apply to you.

Trap one: deepfakes. AI-generated or AI-altered images, audio, or video that look like real people, real objects, real scenes. That "spokesperson" is the textbook case. Lifelike person, lifelike voice — an ordinary viewer can't tell the difference. That alone is enough to trigger the labeling obligation.

Note the key test here: it's not about whether you genuinely intended to deceive. It's about how easily the viewer could be fooled. If your target audience includes the elderly, children, or people who don't really understand AI, the bar is even higher. An older viewer watches that video and thinks it's a real person — you're on the hook.

Trap two: AI text with no human review. This one catches a lot of people off guard. LinkedIn posts, blog articles, web content — if it's AI-generated, published directly, with no human gatekeeping and no one willing to take editorial responsibility, in some cases it must be labeled.

But don't panic — there's a big exemption here.

As long as someone reviewed it before it went out and someone owns the editorial responsibility, you don't need to label it. The number of marketing teams genuinely running "zero-review, auto-publish AI copy" is smaller than you'd think.

Trap three: chatbots. Smart customer-service agents and voice assistants on e-commerce sites and brand websites. The moment a user arrives, before the first message is exchanged, they have to know: you're talking to an AI.

There's a hidden trap here too: your company might end up classified as a "provider" under the Act, not just a deployer. Before IT spins up a chatbot, think this through.

What Doesn't Need Labeling

Not everything that touches AI needs a label. Let me draw you a few bright lines.

Light technical editing doesn't need labeling. Adjusting the background, fixing the lighting, denoising, compressing a file, improving accessibility. These incidental AI-assisted tasks are explicitly excluded by the Act.

Content that's obviously a creative work also carries a lighter labeling burden. Art, satire, fiction, sci-fi — the audience can tell at a glance it's made up, so it doesn't constitute deception.

But here's a landmine marketers should not step on: this exemption almost never applies to you. Why? Because the core purpose of marketing content is information and commerce — to sell, to pitch a product — not artistic creation. The European Commission weighs this heavily when making its determinations.

"First Contact" — An Underappreciated Detail

Article 50, paragraph 5, has a special requirement.

Labeling isn't something you show only to the first wave of viewers when a video goes live.

Every new viewer, at the moment of their "first contact" with the content, must be able to see the label.

What does first contact mean? Say you run a livestream that uses deepfakes. Someone joins five minutes in and misses the opening disclaimer. Now you have an obligation to keep reminding them throughout the stream — you can't just broadcast it once.

When content gets reposted or re-published elsewhere, the label has to travel with it. No hiding it anymore.

How to Label, and Where

This is the hands-on question marketing teams care about most.

The European Commission's guidance is firm: labels go on the content itself.

Buried in the metadata? Not acceptable. Stuffed into a caption that's easy to crop out? Not acceptable. The viewer has to be able to see it, read it, or hear it directly.

The format isn't mandated. A few examples:

A chatbot can pop up "You are interacting with an AI system" at the start of a conversation. An AI-drafted email reply can carry an AI notice at the top. A voice assistant can say "I'm an AI assistant" when a session begins. In a long conversation, remind the user again at intervals.

One more thing that's often overlooked: accessibility. Labels have to be reliably perceivable by people with disabilities too. Germany's accessibility law, the BFSG (Barrierefreiheitsstärkungsgesetz), is tightening in parallel — two tracks moving together.

Don't Bet on One Training — Build a Process

Let me say this upfront: the transparency obligation is not something a watermark can solve.

Watermarks get cropped out, scrubbed away, or simply forgotten. Relying on training to remind employees to "remember to label" — three months later, they'll have forgotten for sure.

Marketing teams that actually make this work do three things:

Prepare standard scripts and label styles for each channel in advance. One set for the website, one for social, one for video, one for audio. Don't improvise at the last minute.

Tag assets in your asset library: "AI-generated," "AI-modified," "Human-reviewed." Before anything goes out, the tag follows the asset — it doesn't get lost.

Add one mandatory, lightweight review step before publishing. Even if it takes two minutes, run through a checklist: Is this content a deepfake? Does it need a label? Has someone reviewed the copy? Think it through, then publish.

Content published before August 2, 2026 isn't required to be retroactively labeled. But the European Commission encourages voluntary labeling — after all, the whole point is trust in the information ecosystem.

What Happens If You Don't Label

Fines. Up to EUR 15 million. Or 3% of your company's global annual revenue for the previous fiscal year. Whichever is higher.

That ceiling isn't just there to scare you. They really will fine you.

And don't think it ends with the fine. If the data was used improperly, GDPR steps in. If an ad is misleading, consumer-protection law and unfair-competition law step in. Several statutes can all come after you at once.

One Last Thing

That friend who sent me the screenshot? He went back, added the AI label to the video, and pushed it out. The performance numbers barely moved. A few longtime customers even left comments saying the disclosure made them feel more reassured.

Transparency in marketing has always been treated as a "cost."

Now it's not just a legal obligation — it's a trust ledger that most of your competitors haven't figured out how to build yet.

Whoever turns labeling into muscle memory first, steps on one fewer landmine. It's that simple.

And may your next AI video be labeled loud and clear.