You probably don't need the AI-generated label

You do need the editorial process that earns you the exemption. Article 50 requires it. Your copyright does too.

Article 50 becomes applicable on 2 August 2026. The final guidelines were published on 20 July, with just two weeks' notice.

Now the market is confused, and searching.

→ Marketers are mostly worried about the wrong thing. They're checking whether to slap a label on everything AI touched. They should be checking their editorial process instead.

Marketing only own one of Article 50's four rules

These are the 4 obligations when you need to notify that you’re using AI.

If your AI talks to customers, it has to say so. Chatbots, AI agents, anything conversational on your site or in your DMs.
AI tools that generate audio, image, video, or text must embed a marker so the output is detectable as AI-generated. These machine-readable markers are the tools responsibility, and are invisible for you and me.
If you're scanning faces or emotions, say so. This is niche. Most content and marketing work never touches this one.
You must inform your audience about deepfakes and text publications on matters of public interest without human review or editorial control.

The fourth rule is ‘deepfakes and text on matters of public interest must disclose that it's AI-generated’ is the one that matters to you. It's also the one that gets misunderstood the most.

And where your exemption lives.

‘Informing the public on a matter of public interest’ explained

There are 3 gates. If one of these answers is ‘no’, you don’t need the AI-generated label.

Is it published? Is it public? Not a draft, not an internal doc, not a one-off email to a client.

Is it informing, not selling? This gate decides most typical marketing cases with ‘no’.

Ask yourself: Is this to promote or directly sell a product or service? Or would you call this journalism, or brand content to inform that happens to touch a serious subject?

Most of your marketing copy, product descriptions, social posts, newsletters don’t fall under the Article 50's AI-labelling rule. Because this content is more sales focused and not a matter of public interest.

The most complicated one: Does the topic sit in a public-interest category? Politics, public health, justice, the environment, consumer safety, or an economic, scientific, or cultural development or trend people are debating.

Explained by example: two pieces, same author (me), same site (mine), opposite outcomes.

My LinkedIn post curating a podcast episode and a summer book pick.Published, yes. Selling? No. Informing the public on a public-interest matter? No. It's brand voice. Out of scope.
A blog breaking down what these Article 50 guidelines mean for content teams, this piece in fact. Published, yes. Written to inform, not to sell. Sitting inside a named public-interest category. That one's in the gray zone. Probably in scope.

But I have my editorial process in place, I can prove my human review, I’m the named author, so no AI-generated label needed.

→ Stop worrying about the AI-tag. Start worrying about the documented editorial process you don't have in place yet.

→ This safeguards your copyright too.

How you qualify for the AI label exemption

This is where most companies don't comply. Human review and editorial control means someone more than just fact-checking or correcting typos. You need someone with real knowledge of the subject examined the content. This person has the actual authority to approve, change, or kill the piece, and takes responsibility for it.

→ Here's the process you need to qualify: one named person, with real subject knowledge, who can reject the piece and is on record as having reviewed it. Write that down as proof, if anyone ever asks.

If you don't have this documentation, you don't have a defensible position. Just a company hoping nobody asks.

There's a second reason to build this documentation. EU copyright law only protects work with genuine human creative input. The proof of your editorial review for Article 50 is the evidence that protects your copyright too.

Deepfakes: visual, video and audio content

A deepfake, under Article 50, is AI content realistic enough to pass as something real that could exist — and convincing enough that someone might believe it. The rules are stricter.

All three of these need to be true before the watermark requirement applies:

It closely resembles the real thing.
That real thing exists.
It could fool someone into thinking it's genuine.

Text gives you an out when the content is selling, not informing. Deepfakes don't get that out. With text, intent decides. With visual, video, or audio, it doesn't — the rule applies regardless of why you made it.

Explained by example: a dragon and an AI host, same technology, opposite outcomes.

A dragon breathing fire in an ad, however photorealistic, isn't a deepfake. It fails criterion two; dragons don't exist and never have, so there's nothing real for the content to falsely pass off as. No disclosure needed, no matter how convincing the rendering.

An AI-generated host fronting your online masterclass is a different case entirely. It resembles a real person. That kind of person, someone teaching, on camera, speaking directly to an audience, plausibly exists. And if you don't disclose it, your audience has every reason to believe they're learning from someone real. All three criteria met.

Same underlying technology. The difference is what the content claims to be, not how well it's made.

Why marketers panic anyway

If the rule is this narrow, why is everyone still panicking?

The usual coverage flattens it. Most headlines read ‘AI Act requires labelling AI content.’ Full stop. Nobody mentions the four-obligation split, the three gates, or the editorial exemption.

The fine number does the scaring. €15 million or 3% of turnover gets quoted next to ‘AI content’ with no qualifier attached. Nobody checks if it applies to them first.

Legal talk adds to the confusion. Provider and deployer obligations get mashed into one. The invisible watermark is the tool's job. The visible disclosure is sometimes yours. Report both as ‘AI content must be marked,’ and marketers can't tell which part is theirs.

Legal teams default to safe. ‘Label everything AI touched’ is the most secure solution. ‘You're probably exempt if you can prove editorial control’ takes work that's outside of their influence.

And fear sells better than nuance. It pitches easier ‘you must label everything’ than ‘you're probably fine if your process holds up.’

→ None of this is irrational.

It just takes real work to understand the obligation, and right now the ownership of that work hangs in the middle, between legal, marketing, and whoever built the AI tool.

Wrappin it up; build the process, skip the AI label

Most companies think they need an AI label. Don’t worry about the tag. Worry about the process you don't have in place yet. Not only for Article 50. For your copyright too.

Compliance was never going to be a sticker. It was always going to be a decision about who's allowed to say yes.

First. Know Article 50, the obligations, the gates and the exemptions. And separate your deepfake from your text content, as different rules apply.

Second. Run your written content through the three gates: published, informing not selling, sitting in a public-interest category. Flag anything that lands in the gray zone.

Third. Pick named persons as topic experts, with real subject knowledge, who have the authority to reject a piece before it publishes. Write down that this person reviewed it, and what they checked. Usually this is also the author of the piece.

→ None of this is a label. It's a process. Build it once, and use it daily.