The EU AI Act's transparency rules, Article 50, apply from 2 August 2026. The short version: if your marketing team uses AI to produce content that could pass for real, the job of flagging that has moved from the platform to you. The European Commission adopted its final guidelines on 20 July and its Code of Practice on Transparency of AI-Generated Content in June, so this is still hot off the press, and I would recommend every marketing leader does some proper digging on it.

Quick health warning before we start: I am a marketer, not a lawyer, and this is not legal advice. It is my read of where this lands for marketing teams. The Commission's Article 50 text and the guidelines are both public if you want the source material.

The big shift: responsibility is moving to the producer

Here is my summary of what is actually changing. Up to now, we produce images, we can use AI, and the platforms and their systems will often pick those up and flag them. Flagging AI has effectively been the platform's job. What the transparency rules do is pass that responsibility to the producer of the content. If you are producing images with AI that could pass for a real person, place or event, you now need to say so. It is no longer just down to the platform to discover it and flag it. It is down to you.

The Act splits duties between providers of AI systems and businesses deploying them, and responsibility travels along that chain rather than sitting neatly with your tool vendor. Which is exactly why this lands on marketing teams and not just on OpenAI.

What is actually in scope

The rules are aimed at content where you are not 100% sure what you are looking at. If you are an advertiser using AI to produce photorealistic content, say a fake human modelling clothes, that needs to be labelled as AI-generated. Realistically, if AI is in your production process and the output is photorealistic, a person, a place, a product in a place, you are now responsible for declaring it.

Worth being precise here, because the test is narrower than “we used AI”. The duty on you as a deployer is about deep fakes, which the Act defines as AI-generated or manipulated image, audio or video that resembles real people, objects, places, entities or events and would falsely appear to be authentic. A stylised illustration, an obviously synthetic graphic, an abstract background: different case. Photorealism and the potential to be mistaken for the real thing are what trigger it.

A few important edges to that:

  • It is not retrospective. Content created before 2 August 2026 does not need to be labelled after the fact, though the Commission does encourage back-labelling where you reasonably can. Anything going forward that will appear in the EU does.
  • UK-based does not mean out of scope. If your content can appear in front of EU citizens, and online it usually can, you should be paying attention.
  • Images are not the whole story. Article 50 also covers chatbots, which must make clear they are AI at the first interaction, emotion recognition and biometric categorisation tools, which require telling the people exposed to them, and AI-written text published to inform the public on matters of public interest.
  • The editorial exemption is a real bar. AI-written text on public-interest topics does not need labelling if it has had genuine human review or editorial control and a person or organisation holds editorial responsibility for publishing it. The Commission is explicit that this means deliberate examination by someone with the relevant knowledge and the authority to approve, change or reject the piece. A spell-check or a grammar pass does not count.

Plenty of the detail will be up for debate, and I am sure it will not be clear-cut in practice. But the direction is clear.

What the label actually looks like

This is the bit most of the coverage skips, and it is the question every creative team asks first: what do we actually put on the artwork? Alongside the guidelines, the Commission has published a Code of Practice on Transparency of AI-Generated Content. It is voluntary, and following it is not by itself conclusive proof of compliance, but it is the clearest published answer to that question. Ignore it and you are on the hook for demonstrating compliance some other way.

It comes with a ready-made EU icon, free to use, no attribution needed. There are two main versions: one for content that is fully AI-generated, one for content that has been AI-modified. The design rule is that a capitalised “AI” does the heavy lifting visually, with a word next to it saying which of the two it is. User testing across member states found the version carrying that word was significantly clearer to people than the icon on its own, which is a useful hint for anyone thinking of designing their own.

Placement is specified almost as tightly as the mark itself:

  • Images: somewhere clear with nothing overlapping it. The top right corner is the worked example.
  • Video: at the start, then again at intervals, so it survives someone clipping a fragment or grabbing a screenshot. At minimum after any break.
  • Text: at the top, near the headline, or in the byline block. You are allowed to label only the part that is AI-generated rather than the whole piece.
  • Audio: a short spoken disclaimer at the start, plus a visual label wherever there is a screen to put one on.
  • Artistic and creative work: a lighter touch is permitted, in the credits or the accompanying description, so the disclosure does not wreck the thing you made.

The principle running through all of it: someone should notice the label at first exposure, without hovering, clicking or paying close attention.

The bit nobody has told your studio about

The other half of Article 50 lands on the AI providers rather than on you. They have to mark their outputs so machines can read them, and the Code expects two layers: signed metadata recording that the content is AI-generated, plus an imperceptible watermark inside the content itself.

Which means the files coming out of your AI tools already carry that provenance data. Two consequences for a marketing team:

  • Do not strip it. Providers are expected to prohibit deliberate removal of those markings in their terms of use. In practice most metadata loss is not deliberate at all: a resize, a re-compression, a CMS upload or an export preset quietly drops it. Worth finding out what your own pipeline does to a file.
  • Detection is going public. Providers have to make detection tools available, generally free of charge, so anyone can check whether a piece of content came out of their system. Assume “would anyone actually know?” stops being a useful question.

A label is not a get-out-of-jail-free card

Google has also introduced AI disclosure labels for adverts, so you can mark an ad as produced with AI. Useful, but it is not a catch-all. Slapping the label on does not make the ad legally compliant. You still cannot produce random images that do not represent your product. There is a whole accountability piece here, and the label is one part of it, not a substitute for it.

"But we could always do this with Photoshop"

I get it. Image manipulation is not new, and none of this existed for Photoshop. The difference is that AI can do it at vast scale, very quickly, and you no longer need to be an expert. You used to have to be genuinely good at Photoshop to manipulate an image convincingly. Now it is prompt-driven. That is why regulators are moving, and honestly, I think it makes sense.

How long will this actually matter?

Here is my view, and time will tell. I think this is partly a covering-the-bases phase while people are saying “I don't like AI images.” In two years' time, that is largely what they are going to get anyway, and I suspect it becomes a non-issue, the same way people once did not want to buy things on their phones. Remember that? Nobody even considers it now. A lot of this is an adoption curve, helping people get confident and familiar with AI content.

There will always be a need for human content, because there are moments where you just want that authenticity. But AI is getting so much better that you will increasingly get authenticity of the product through AI too. It will keep the product exactly as it is and change the bits around it, backgrounds and settings, rather than the thing itself.

On enforcement: I do not know how strict they are going to be, and I would expect a period of everyone getting used to the idea. But as ever with regulation, they may well look to make examples, particularly of bigger organisations or anyone who deliberately flouts the rules. The fines go up to 15 million euros or 3% of worldwide annual turnover, whichever is higher, so it is worth being mindful in the short term while user sentiment is still “I don't want AI images, or I need to know if it is AI.”

The practical bit: build it into your sign-off process

What I would recommend for marketing teams is simple:

  • Identify where AI is used in your production process.
  • Review that content, and decide whether you need to disclose it against the rules.
  • Build it into sign-off. You already check creative through compliance on product, pricing and language. Add another column to the Trello or Asana board: are we declaring whether this is AI, and has it been checked?
  • Name who signs off the writing. If you publish AI-assisted articles on public-interest topics, that editorial exemption depends on a real person holding editorial responsibility. Write down who that is, their role and their contact details, and publish them. It is a short policy document, not a project, and it doubles as the author transparency that search engines and AI answer engines are looking for anyway.
  • Check what your image pipeline does to metadata. One test file through your normal resize, export and upload route will tell you whether provenance data survives.

That extra layer of checking is a quick, practical step while we all familiarise ourselves with the detail of the policy.

And should you rush to get everything published before the 2nd of August? I would not recommend that either. Do it legitimately.

Quick answers

Does every AI-generated image need a label?

No. The duty on you as a deployer is about deep fakes: content that could be mistaken for something real. Obviously synthetic, stylised or illustrative work sits outside that.

We are a UK business. Does this apply to us?

If your content is put in front of people in the EU, work on the basis that it does. Online, it usually can be.

Do we have to relabel everything published before August?

No. It is not retrospective, though the Commission encourages back-labelling where it is straightforward.

Does using the EU icon make us compliant?

It is the clearest route, but the Code of Practice is voluntary and following it is not conclusive proof on its own. Skip it and you need to show you met the obligation another way.

Does AI-assisted copywriting need a label?

Only where the text is published to inform the public on matters of public interest, and even then not if it has had genuine human review and someone holds editorial responsibility for it. Most product and brand copy is a different question.

What are the penalties?

Up to 15 million euros or 3% of worldwide annual turnover, whichever is higher.

Where this comes from

Everything above is my reading of three public documents, all worth going to directly if this affects you: Article 50 of the AI Act itself, the European Commission's guidelines on transparency obligations adopted on 20 July 2026, and the Code of Practice on Transparency of AI-Generated Content, which is where the icon, the placement rules and the marking detail come from. Still not legal advice, and still worth a conversation with someone qualified if you are making a call on a big campaign.

If you want a structured look at where your marketing stack sits against this, it is exactly the kind of thing I do through AI audits and consultation, and with the team at Sleeping Giant Media on the content and search side. Or just get in touch and tell me what your setup looks like.

Luke