Article 50 of the EU AI Act is the part of the regulation most marketers haven’t heard of, and it’s the part that actually reaches their daily work.
The headline obligations everyone’s been bracing for, the ones covering high-risk AI systems, got pushed back to December 2027 after the European Parliament approved a deferral by 423 votes to 57 in June 2026. What didn’t move is Article 50, and it covers exactly the territory a marketing team lives in: chatbots, AI-generated images, AI-written copy, and anything that looks or sounds like a real person but isn’t. It’s enforceable now, and the fines run up to €15 million or 3% of global turnover, whichever is higher.
This isn’t a legal briefing. I’m not a lawyer, and you should treat everything here as a starting point for your own advice, not a substitute for it. But most of what Article 50 actually asks of a marketing team is closer to a UI review than a legal project, and I want to walk through what that looks like with real examples, not just the article numbers.
EU AI Act’s Article 50 is Upon Us
What Article 50 actually requires, in plain terms
Strip away the legal language and Article 50 asks five separate things, each triggered by a different kind of output.
- If a person is talking to your AI, they need to know it.
- If your AI generates content, that content needs a machine-readable mark, not just a visible label.
- If your AI creates something that could pass as a real person, real event, or real photograph, that’s a deepfake disclosure.
- If you’re using AI for emotion recognition or biometric categorisation, people need to be told.
- And if AI-generated or AI-manipulated text is being published on a matter of public interest, it needs its own label.
These are five distinct triggers, not one blanket rule, and mixing them up is where most teams get into trouble.
Example one: the chatbot that “obviously” sounds like AI
The most common mistake I see is a marketing team assuming their chatbot’s disclosure obligation is covered because the bot’s replies are “obviously” AI-generated. Under Article 50(1), the exception for something being obvious only applies when a reasonably well-informed, observant person would recognise they’re dealing with a machine without being told, consider a clearly robotic, synthesised voice, not a fluent, natural-sounding customer service chat. A well-written support chatbot that mimics human conversation almost never clears that bar.
What compliant looks like: a short, visible line at the very start of the conversation, “You’re chatting with our AI assistant,” before the first substantive reply, not buried three messages in and not tucked into a terms of service page nobody opens.
What non-compliant looks like: a chatbot with a human-sounding name and no disclosure anywhere in the conversation itself, relying on a mention in your privacy policy to cover it. A clause nobody reads doesn’t satisfy a disclosure requirement that has to appear at the first interaction.
Example two: the AI-generated blog header image
This is where the distinction between visible disclosure and machine-readable marking actually matters, and it’s the mistake I’d bet most marketing teams are currently making without realising it. A caption reading “image generated with AI” satisfies a person reading your page. It does not, on its own, satisfy Article 50(2), which requires generated images, audio, video and text to carry a machine-readable mark that other systems can detect, not just something a human eye can read.
What compliant looks like: generated images carrying embedded metadata or a watermarking standard like C2PA Content Credentials, alongside your visible caption, so both a person and a platform can identify the content as AI-generated.
What non-compliant looks like: a visible “AI-generated” caption with nothing embedded in the file itself. It looks compliant. It isn’t.
BTW: the image on this blog was created by me, and I haven’t used AI to build it. Just for transparency. 😅
Example three: the AI-written newsletter nobody labelled
Text gets forgotten constantly, because teams mentally file “AI content marking” under images and video and stop there. Article 50(2) covers text just as much as any other modality. If you’re using AI to draft newsletter copy, blog posts, or social captions and publishing that output with minimal human rewriting, it falls under the same marking obligation as an AI-generated image.
What compliant looks like: a documented editorial process showing meaningful human review and revision before publication, which is generally treated as taking the output out of “AI-generated” territory and into ordinary human-authored content with AI assistance. Keep that editorial trail as evidence, not just as good practice.
What non-compliant looks like: publishing model output essentially unedited and treating “we used AI to help write it” as a private implementation detail rather than something that needs to be reflected in your process documentation.
Example four: the synthetic spokesperson in your video ad
If a marketing team generates a video or voice ad using a synthetic likeness, an AI avatar, a cloned voice, or footage that could be mistaken for a real, unaltered recording of a real person, Article 50(4)’s deepfake labelling obligation applies directly, and there is no “obvious from context” exception here worth relying on.
What compliant looks like: a clear, visible label on the content itself stating it’s AI-generated or manipulated, not just in the video description or a caption that disappears after autoplay starts.
What non-compliant looks like: a synthetic spokesperson presented without any on-content label, relying on audience assumption that “everyone knows this kind of thing is AI now.” That assumption isn’t a legal defence.
Example five: the AI-drafted press release on a public-interest topic
Article 50(5) singles out AI-generated or AI-manipulated text published to inform the public on matters of public interest, think policy commentary, public health information, or anything positioning your brand as commentary on a genuinely public issue. This is narrower than routine marketing copy, but PR and content teams touching current events or policy topics should treat it as its own checkpoint, separate from your general content marking process.
The one grace period that’s real, and the trap hiding inside it
There is a real grace period inside Article 50, and it’s narrower than the “everything got delayed” headlines suggest. Systems already on the market before 2 August 2026 get until 2 December 2026 for the machine-readable marking requirement specifically, and only that requirement. Chatbot disclosure and deepfake labelling carry no grace period at all, regardless of when the system launched. And if you ship a new generative feature after 2 August 2026, whatever grace period existed doesn’t apply to it. It needs to be compliant from day one.
The European Commission published a voluntary Code of Practice on marking and labelling AI-generated content on 10 June 2026. Signing it is one route to demonstrating compliance with the marking obligations, not a legal requirement, but it’s worth reviewing if your team is producing generated content at any volume.
If you’re using someone else’s AI tool, check your contract before your code
A lot of marketing teams don’t build their own chatbot or image generator, since they license one. That splits the obligation. The tool provider generally carries the duty to build in machine-readable marking capability. You, as the deployer, carry the duty to make sure disclosure actually reaches your audience and that the marking is switched on, present, and functioning in what you publish.
If your vendor hasn’t implemented marking, your compliance gap becomes yours to explain, not theirs. Confirm this in writing, not by assumption.
A working reference for your team
The table below is the version of this I’d actually hand to a marketing team on a Monday morning, mapped to the touchpoints most content and campaign work actually runs through.
Where to start this week
Don’t start with your model inventory.
Start with your disclosure surface, every chatbot, every AI-assisted piece of published content, every generated image or video that reaches an EU audience.
Most teams find touchpoints sitting in tools that predate their formal AI programme entirely, an AI writing assistant plugged into a CMS months ago, a chatbot vendor swapped in without anyone flagging it as an AI system. Find those first.
The fix for most of them is a UI change and an honest look at your publishing process, not a legal overhaul, but it only works if you know where to look.
This article was drafted with AI assistance and edited and fact-checked by Rui Nunes before publication.
Leave a Reply