Skip to content
imgmarker
Go Premium

AI Labeling Since August 2: What's Actually Required on Your Website

By Cody Krause

Since August 2, Article 50 of the EU AI Act has been in force, and ever since, a precautionary "Created with AI" has been showing up under every other blog image. You can skip most of that. Two cases, however, are overlooked by many site operators: deepfakes and AI-generated text on matters of public interest. When labeling is required, when it isn't, and what the notice actually has to look like.

Since Sunday, Article 50 of the EU AI Act has been in force. Ever since, I've been seeing a precautionary "Created with AI" under every other blog image. Most of these notices are unnecessary. Two cases, however, are overlooked by many site operators, and those are exactly the ones that can end up costing money.

Why the deadline actually held

Back in spring, the word was that the Commission would push everything back. That's only half true. The Digital Omnibus, passed by Parliament and Council in June 2026, moves the obligations for high-risk systems from August 2, 2026 to December 2, 2027. So anyone using AI in hiring or lending decisions gets 16 more months. The transparency rules in Article 50, however, were left untouched by lawmakers. They've applied since August 2, 2026, with no transition period.

There is exactly one grace period. For generative systems that were already on the market before August 2, 2026, the machine-readable marking required under Article 50(2) doesn't kick in until December 2, 2026. That deadline applies to the makers of the tools, not to you as a user. Anyone putting a system on the market after the cutoff date has to mark it from day one.

Provider or deployer: everything hinges on this

The regulation assigns obligations based on roles. Get the role wrong, and you'll end up labeling the wrong thing while missing the right one.

Providers

Providers develop an AI system and put it on the market. Think OpenAI, Adobe, Midjourney, Google. They have to mark their outputs in a machine-readable way and build chatbots so that users can recognize they're talking to AI. This obligation doesn't land on you as long as you're using someone else's tools.

Deployers

Deployers use AI in a professional capacity. That's you, if you're generating images or having text written for your company website. Your list is shorter than most people think: label deepfakes, and label AI-generated text on matters of public interest. That's it.

Private individuals are excluded entirely. Someone who makes a deepfake for fun and shares it in a family chat doesn't fall under Article 50. The moment that same content ends up on your company website, things look different.

The trap of putting your own brand on it

Anyone who offers a third-party AI system under their own name slides into the provider role and inherits all of its obligations. This affects more websites than it sounds like it would. A white-label chatbot that you run under "Your Assistant" on your homepage makes you the provider of that system. The same applies if you self-host an open-source model and build a service on top of it.

Images: three questions and you'll know where you stand

An image only requires labeling if all three of the following apply. The Commission lists them in its FAQ on Article 50: resemblance to a real person, object, or place; plausible existence, meaning what's shown could genuinely exist; and the impression that the image depicts something real.

If even one of the three is missing, you don't need a label. That's why a cartoonish robot in a blog header doesn't count, while a deceptively realistic product photo does.

Four everyday cases

  • Label it: You furnish an empty apartment photo with AI for a listing. Viewers assume the furnishings are real, and then show up to an empty apartment at the viewing appointment.
  • No label needed: Your smartphone's night mode removes noise. Standard editing like this doesn't change what the image is saying.
  • Borderline case: You retouch wrinkles out of a portrait. Once the point of the image shifts, say in a before-and-after comparison for a cosmetic treatment, it tips into requiring a label.
  • No label needed: You generate a dragon flying over Cologne Cathedral. Obviously fantastical, nobody mistakes it for a photograph.

The trade publication docma worked through this distinction using several real cases and offers a rule of thumb that holds up in everyday use:

Does the image still show what was actually in front of the lens?

If you answer yes, the image doesn't need a label. If no, it's worth taking a second look at the three criteria.

Text: the filter is called "public interest"

This is the point most blogs get wrong, in both directions. Some slap a notice under every piece of AI-assisted text, others under none at all.

An AI-written text needs a label if you're publishing it to inform the public, and if it concerns a matter of public interest. In its guidelines, the Commission lists: politics and democratic processes, public administration, the judiciary, fundamental rights, public safety, health, environmental protection, consumer protection, and economic, financial, scientific, and cultural developments that are subject to public debate.

Your recipe for pasta alla Norma doesn't fall under this. Your article about a new tax rule, a medication, a product recall, or the local election does. A guide to retirement savings and a write-up analyzing a new study also belong in this group.

What counts as editorial review

The exemption is what saves most editorial operations. Text that a qualified person has reviewed for content, and for which someone holds editorial responsibility, doesn't need a label. The Commission requires a substantive review for this to apply. A spellcheck, a quick glance, and a checkbox in the workflow aren't enough.

In practice, that means: someone who reads the raw AI text, fact-checks numbers, verifies sources, and rewrites passages satisfies the exemption. Someone who runs the text through a tool and hits publish does not. Keep a record of who checked what and when, even if it's just a field in your CMS. If there's ever a dispute, you'll need to show this, and three years later nobody will remember.

What doesn't fall under the obligation at all

A lot of everyday work is exempt. You don't publish internal notes, quotes, meeting minutes, or emails, so Article 50 doesn't apply. Product descriptions in a shop don't inform the public about a matter of public debate, even if AI wrote them. The same goes for travel tips, recipes, craft tutorials, and the vast majority of service content.

Also exempt is text where AI helped you write without supplying the content itself. Organizing an outline, polishing phrasing, tightening sentences: these helper functions count as standard editing. The Commission's test is whether the tool changed the substance of the content. If you're the one doing the thinking and using AI as an editor, you remain the author.

What the notice actually has to look like

A person must be able to perceive it without technical aids, at the latest upon first contact with the content, clearly and distinguishably. Metadata alone isn't enough to satisfy deployer obligations. A visible label on the image, a notice line above or below the text, a spoken notice for audio.

Since June, the Commission has been providing an EU icon you can use instead of your own wording. It comes from the Code of Practice on Transparency of AI-generated Content, which the Commission finalized on June 10, 2026. Signing on is voluntary, but it helps as evidence if regulators come asking.

For artistic, satirical, or fictional work, a more understated notice is enough, say in the credits or accompanying text. Nobody expects a watermark stamped across an illustration.

Two things that go wrong quickly

The chatbot on your homepage

The obligation under Article 50(1) falls on the provider, who has to design the system so users can recognize they're dealing with AI. The moment you run the bot under your own name, the responsibility falls on you. A single sentence in the first message is enough: an AI assistant is responding here. Where AI use is obvious, the notice isn't required, but I wouldn't push that exception too far with a support widget.

Destroyed provenance data

Converting, cropping, and compressing images often strips out C2PA data and watermarks. That's why the Code of Practice requires providers to use at least two layers of marking, since no single technique is robust enough on its own. Check your image workflow before your CDN optimizes the evidence away. One test is enough: upload a generated image, download it again, and see whether the signature survived the trip.

Who enforces this, and what does a violation cost

In Germany, the Bundesnetzagentur handles market surveillance. This was established by the Bundestag in June 2026 through the AI Market Surveillance and Innovation Promotion Act, known as KI-MIG for short. The agency serves both as the central point of contact and as a complaints office, meaning tips from competitors land there too.

The penalty framework for violations of Article 50 goes up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher. Small and medium-sized businesses are subject to the lower amount. The often-cited 35 million euro figure applies to the prohibited practices under Article 5, not to labeling.

Whether the Bundesnetzagentur will open proceedings or focus on guidance in its first year is something nobody currently knows. So far, there are no publicly documented cases. That said, I'd still expect letters from competitors, since competition law offers a second lever alongside regulatory oversight.

Your to-do list for this week

  1. List every spot on your site where AI contributes to writing or imagery: images, text, chatbot, translations, product descriptions.
  2. Sort your images using the three questions. Only what's left needs a label.
  3. Go through your articles on health, law, money, politics, and consumer topics. Either add a label or document editorial review.
  4. Add a disclosure sentence to your chatbot.
  5. Check whether your image pipeline preserves metadata.
  6. Record who holds editorial responsibility. Without a named person, the exemption for text doesn't apply.

Where I'm uncertain

The Commission's guidelines from July 20, 2026 don't carry the force of law. They show how the Commission interprets the rule, but the final say rests with the European Court of Justice. On borderline cases, like wrinkle retouching or the question of when a how-to article touches on a matter of public interest, legal opinions differ.

If a lot of money or a lot of reach is riding on this for you, get it checked by a lawyer instead of relying on a blog post, including this one.

Sources

Read next

Diesen Beitrag auf Deutsch lesen

Reviewed by the editorial team of MOVA Empire UG (haftungsbeschränkt).