Your AI marketing has to come clean. From 2 August 2026.

A friend messaged me last week. She runs marketing for a mid-size brand. "We made this video with an AI avatar," she said, "looks totally real. Do we have to put a sticker on it?"

I read the message twice. Then I went and re-read Article 50 of the EU AI Act.

The short answer: yes, almost certainly. And it's not a sticker.

Why the AI Act cares about your marketing

The AI Act sorts AI systems into risk buckets. For marketing teams the bucket that bites is "limited risk," and the role that bites is the deployer. The Act's worry is simple, and honestly reasonable: people should not mistake a machine for a human, or a fake for a recording.

So if you publish something an AI made or seriously altered, the people on the receiving end have to be able to tell. That's the whole game.

When does this actually show up at your desk? Three places.

1. Synthetic or manipulated media. Images, audio, video, especially the photorealistic kind, or a voice that sounds like a real person. The dreaded deepfake.
2. Synthetic text. LinkedIn posts, blogs, web articles that go out without a human editor touching them.
3. Chatbots and voice assistants on your site, where the company may count as a "provider," not just a deployer.

The rule of thumb: at the very first moment someone meets the thing, they should know it's AI.

"Deepfake": what does it actually mean?

This word gets thrown around loosely. The Act has a precise line.

AI-generated or AI-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.

Read that slowly. Two things matter.

Does it look authentic? Not "did you mean to fool anyone." That's irrelevant. The question is whether a normal viewer, given the context, could be fooled. And you have to think about your actual audience, not some abstract average user. If kids, or older people, or people with low AI literacy might see it, the bar is higher, because they're more likely to get duped.

And the resemblance has to be objectively recognisable. Persons means realistic humans, alive or historical, dead people count too. Objects, places, even animals. Events: a realistic-looking scene of, say, your product being used.

Honest question to ask yourself about every asset: could a reasonable person think this really happened? If yes, label it.

The escape hatch that won't save you

There's a carve-out in the law for "evidently artistic, creative, satirical, fictional" work. There, you only have to disclose in a way that doesn't ruin the piece.

Sounds great, until you remember what marketing is for. Marketing is almost always informative or commercial. That single fact kicks you out of the carve-out. So don't lean on this one. It is, for most of us, a door that doesn't open.

There's another, smaller door that does open. Minor, technical edits. Background cleanup, lighting tweaks, colour correction, noise reduction, compression, accessibility fixes. Those don't need a label. The law cares about authenticity, not whether you ran a photo through Lightroom.

The text question, and the answer you'll like

Here's where marketers breathe out.

AI-generated text has to be labelled only when all of these are true: it's published to a large, unconnected audience, its job is to inform, and the topic is a matter of public interest. Something society actually needs to debate, like the economy, politics, science, culture.

And even then, there's the big exemption. If a human reviewed it, and a real person or company holds editorial responsibility for what goes out, you don't have to label it.

Think about that for a second. Your blog post that an AI drafted and an editor signed off on? No label. Your LinkedIn caption that someone actually read before hitting publish? No label. The marketing department's job, day to day, is almost always already inside this exemption. As long as the review is real, not theatre.

A private email, an internal memo. Those aren't "published" at all. Out of scope entirely.

So when does a marketer actually have to slap "AI wrote this" on text? Rarely. Honestly, rarely.

The chatbot rule, and a quiet trap

If you put a chatbot on your site, or a voice assistant, Article 50(1) kicks in. People have to be told they're talking to an AI, at the moment they start talking to it. A written line ("You're chatting with an AI"), a permanent badge, a voice prompt at the start of a call. Any of these can work.

The trap is the role shift. When the company commissions or builds that bot, it can cross from "deployer" into "provider" under the Act. That brings more obligations, not fewer. The IT team that installs the bot and the marketing team that asked for it need to know that.

And here's the bit most people miss: "first contact" isn't once. It's every viewer's first contact. A deepfake video republished somewhere else six months later still needs its label visible to whoever sees it then. If it's a live stream, you don't just warn people at the top. You warn them again mid-stream, because someone just joined.

Where the label goes, and what it looks like

The European Commission's guidelines are blunt here: the label goes on the content. Not buried in metadata. Not in a caption someone can crop out in two seconds. On it.

For chatbots there's no fixed format. Written notice, visual badge, audio cue, whatever fits. The only hard rule is that it has to be perceivable, including by people with disabilities. If a blind user can't tell your bot is AI, your label isn't done.

The fine

This is the part that focuses the mind.

Up to €15 million, or 3% of global annual turnover from the prior year, whichever is higher. Plus, on top of that, separate actions under the GDPR or consumer-protection law if your AI content misleads someone or processes their data wrongly.

That number is the reason "we'll get to it later" is not a strategy.

The part nobody says out loud

Here's what I keep coming back to. Most teams I see are treating this as a labelling problem. It isn't. It's a process problem.

You don't solve it with one training. You solve it with three small, boring habits.

Standard label texts for each channel (website, social, video, audio) so nobody invents a new one each time. File naming that flags every asset as "AI-generated," "AI-modified," or "Human-reviewed" before it ever reaches the CMS. And a 30-second pre-publish check: is this a deepfake? Is this text unreviewed? Is the bot disclosed?

That's it. Three habits. Run them every time and the legal risk collapses to almost nothing.

And content you published before 2 August 2026? Grandfathered. No retroactive labelling required. Though the Commission gently suggests you label the old stuff anyway, because the whole point of this law is trust, and trust is built by doing slightly more than the minimum.

The real headline

Transparency isn't a watermark. It's how your team operates on a Tuesday afternoon when nobody's watching.

The brands that get this right won't be the ones with the slickest AI labels. They'll be the ones whose process is so normal that nobody on the team even thinks about it anymore. The label is just the visible tip. The iceberg is the workflow underneath.

Get the workflow right. The labels take care of themselves.

And if you're that friend with the photorealistic AI video: yes, label it. The deadline just passed.