The EU now wants labels on AI content. I read the rules so you don’t have to

On August 2, the EU AI Act’s transparency rules went live. Article 50 is the one that decides when AI-generated content needs a label.

I generate content with AI every day, including programmatic SEO pages, glossaries and social posts. So instead of the LinkedIn panic posts about them, I read the actual rules: the regulation, the Commission’s guidelines from July 20 and the Code of Practice from June 10.

The short version: this is much narrower than the panic suggests, and one exemption does most of the work. But if your company ships AI content at scale, or you’ve vibe-coded a tool that generates it, there are real obligations now, with fines up to 15 million EUR or 3% of global turnover behind them.

Here’s what actually changed.

First, figure out which hat you’re wearing

The law splits the world into providers and deployers.

You’re a provider if you built the generative tool and a deployer if you use one (ChatGPT, Midjourney, Claude, whatever) in a professional context. A marketing team is a deployer. A founder shipping AI blog posts is a deployer.

If your vibe-coded weekend project generates text or images for other people, congratulations, you might be a provider now. Providers must make sure outputs are marked in a machine-readable format and detectable as AI-generated, such as through watermarks or provenance metadata. Tools already on the market before August 2 got a grace period until December 2, 2026, while new ones need it from day one.

At Parvi, the content and social agents draft and publish content daily, with a growth strategist owning the quality of everything that ships. Keep that setup in mind, because the law is about to reward it.

Blog posts probably don’t need a label

AI-generated text needs a disclosure only when it’s published to inform the public on matters of public interest. These include politics, health, security, environment and consumer safety. A SaaS landing page is not a matter of public interest, no matter what your pitch deck says.

But even then, there’s an exemption that covers most real content operations: no label is needed if the text went through human review and someone holds editorial responsibility for it.

That review has to be real, though. The guidelines say a human with relevant knowledge has to examine the substance and have the power to approve, change or reject the piece. So for instance running spell-check does not count, and neither does skimming the headline while automation publishes the text.

In the big picture, this is about more than compliance. The AI engines deciding visibility reward the same thing the law now rewards: a validated human with substance. Google has spent years tuning its quality systems toward demonstrated experience and real authorship, and LLMs pick their citations from sources that carry credibility signals: real names and expertise that is accountable for the text. I wrote a year ago that visibility is being replaced by credibility. The regulation is now pointing the same direction the algorithms were already going.

So the compliance question for AI text turns out to be the same question I keep writing about as a quality question: is there a human who actually read this and put their name on it?

If yes (which is how I run my own pSEO builds, and literally Parvi’s operating model), you’re fine. If you built a fully automated publish loop where nobody reads anything, the EU would now like a word. So would your readers, if you have any left at this point.

Images follow the deepfake rule

The image rule doesn’t care about public interest. It cares about deepfakes, and the definition is wider than the word suggests: it’s AI content that resembles existing persons, objects, places or events, and looks deceptively authentic.

The July guidelines drew some useful lines here:

Photorealistic AI people fall under this categorisation, including virtual influencers and realistic AI avatars. A purely fictional person can fall outside the definition, but the closer the result is to something that could be a real photo of someone, the more certainly it is in scope.

Content that obviously couldn’t exist is out, and the Commission’s own example is a human-sized capybara. There is a capybara clause in EU law now. Somebody in Brussels got paid to write that sentence.

Illustrations and clearly stylized images are out, and photorealism is what decides it.

From a marketing perspective, the most important line: the artistic exception does not save ads. The guidance says that when informative or commercial purposes predominate, the exception doesn’t apply. If a campaign uses a photorealistic AI product shot or an AI-generated “customer”, it needs a visible label from the first moment a viewer sees the image. Visible means a human can spot the label without any tools. Metadata embedded in the file is not enough, and neither is a label that gets cropped out in a story format.

Color correction, noise reduction and normal retouching are not manipulation. Nobody has to label a lighting fix.

What to do this week

Sort your content into three buckets: AI-generated, AI-assisted and human. You need to know which is which anyway, and your file names don’t currently know, so fix that.

For text, make the human review step explicit: usually this is a named person who actually reads and has the power to reject. Log it. You probably do this already, and now it’s also your legal exemption.

For images, one rule: photorealistic AI people, places or events get a visible “AI-generated” label at the point of first contact. Stylized stuff doesn’t. When in doubt, ask whether a stranger would believe it’s a photo.

Don’t strip provenance metadata in your publishing pipeline. The tools are starting to write it in, and quietly deleting it is a bad move with a fine attached.

If you built a generative tool, the machine-readable marking is on you as the builder, not on your users. It’s due now, or by December if you shipped before August.

Old content is safe. Nothing generated before August 2, 2026 needs retroactive labels. Enforcement here in Finland runs through existing market surveillance authorities, and Traficom acts as the contact point. The powers have been in force since January.

The part I like

I’ve written before that every AI writing tool sounds the same and that AI didn’t kill your SEO, your low-quality content did. The pattern underneath both is the same: the problem was never that a machine made the thing, it was that no human reviewed the thing.

And it shows up everywhere once you start looking. It’s one of the reasons I left VC. I got tired of reading pitches that didn’t say anything anymore, because they were pure slop: generated, polished and sent without anyone caring what was in them.

Article 50 pretty much agrees. Text with a real human taking responsibility needs no label. Slop published by automation with nobody looking needs a warning sticker.

The EU just turned this into a legal category.

For the record: an AI drafted this post, a human reviewed the substance and holds editorial responsibility for it. That human is me, a real domain expert you can verify on LinkedIn. Connecting content to an accountable expert is exactly the signal both the law and the AI engines now reward. That’s the whole compliance story, and it fits in one paragraph.

Mari Luukkainen

Slop Cannon

Ex-VC, full-time builder. Personal record: 349 GitHub commits in one day. Find me on GitHub and at mariluukkainen.com.

Read more about Mari Luukkainen