EU AI Act Article 50: transparency duties from 2 August 2026

0:00 EU AI Act Article 50

Second of August, twenty twenty-six. That's the day a set of transparency rules in the EU AI Act started to apply, and they reach far beyond Europe. If you run a chatbot that talks to customers in Paris, or you make a realistic AI video for a client selling in Berlin, they can apply to you, even if your office is in Lahore, Dubai or Manchester. In this lecture you'll learn the four Article 50 duties, who carries each one, what changed in twenty twenty-six, and a screening checklist you can run on every asset.

0:41 Why it matters

Why does it matter? Because the Act applies to providers and deployers outside the EU when the output is used in the EU. Fines for breaking Article 50 can reach fifteen million euros or three percent of worldwide turnover, whichever is higher, with the lower figure for small businesses. But the bigger practical risk for most of you is simpler. EU clients will now ask you to prove you've done it. Agencies that can answer quickly, with evidence, win the work.

1:16 Two roles

Here's the mental model. Think of the Act as a restaurant with two roles. The kitchen, the provider, builds the dish and must put an allergen code on the packaging that machines can read. The waiter, the deployer, serves it to the customer and must say out loud when something could be mistaken for something it isn't. Both have duties. Knowing whether you're the kitchen or the waiter for each tool is the first step. Most marketers are deployers. But if you build and brand your own chatbot, you may be a provider too.

1:57 The four duties

Now the four duties. Paragraph one: providers of AI systems that interact with people must design them so people know they're talking to AI, unless that's obvious. Paragraph two: providers of generative systems must mark outputs in a machine-readable way, as far as technically feasible. Paragraph three: deployers of emotion recognition or biometric categorization must inform people. And paragraph four: deployers must disclose deepfakes, and must disclose AI-generated text published to inform the public on matters of public interest, unless a human editor reviewed it and takes responsibility. All disclosures must be clear, and given at the latest at first interaction or exposure.

2:42 What's a deepfake?

What counts as a deepfake? The Act's definition: AI-generated or manipulated image, audio or video that resembles existing persons, objects, places, entities or events and would falsely appear to be authentic. So a stylized illustration of a coffee cup? Not a deepfake. A photorealistic video of a real founder saying new words, or a real Amsterdam street in a scene that never happened? Deepfake. A cloned voice of a real person? Deepfake. And if the content is evidently artistic, satirical or fictional, the duty is lighter: you disclose in a way that doesn't spoil the work, like a credit at the end.

3:26 2026 changes

What changed this year? Two things. First, in June twenty twenty-six the Commission published the final Code of Practice on marking and labelling AI-generated content. It's voluntary, but it's the practical reference. It covers machine-readable marking for providers, and labeling for deployers, including a common EU icon you can use. Second, the Digital Omnibus on AI gave providers of generative systems already on the market before August second a grace period to December second, twenty twenty-six, for machine-readable marking only. Chatbot disclosure, deepfake labels and public-interest text? No grace period. They apply now.

4:06 Example 1: Riyadh retailer's chatbot

First example, simple. A Riyadh furniture retailer launches an AI chat assistant on its website, and it sells to customers in Germany and France. The assistant's first message says: Hi, I'm the AI assistant for the store. I can help with sizes and delivery. Type human any time to reach our team. They test it monthly in Arabic, English, German and French. That's paragraph one handled, whether they count as provider or deployer, because the disclosure is designed in from the first message.

4:42 Example 2 (illustrative): Lahore agency, Dutch client

Second example, a realistic business scenario with illustrative details. A Lahore agency makes a spring campaign for a Dutch homeware brand. Four assets. AI abstract backgrounds: not deepfakes, no label, keep the metadata. A photorealistic AI customer in a real Amsterdam street: deepfake, so it gets the EU icon and an AI-generated scene line, and they check it doesn't imply a real customer testimonial. The founder's cloned voice-over, with written consent: deepfake, so a spoken AI-generated voice line. And the website assistant: an opening disclosure. Illustratively, when the client's legal team asked for the agency's Article 50 evidence, the agency sent the asset register and screenshots the same afternoon.

5:29 Watch me do it: 30-second ad

Watch me do it. I'm screening a thirty-second video ad for an EU skincare launch. Question one: will it be seen in the EU? Yes. Question two: is it a chatbot or voice agent? No. Question three: does anything resemble a real person, place or event and look authentic? The presenter is an AI-generated woman who doesn't exist, but the lab scene shows the brand's real factory, altered to look bigger. That's a real place shown in a way that didn't happen: deepfake. Question four: evidently fictional? No. Question five: public-interest text? No. Question six: did our tool vendor mark the file? I check the vendor sheet: yes, metadata. Question seven: I add an end card, this video contains AI-generated imagery, screenshot it, and log the row.

6:24 Common mistakes

Common mistakes. Thinking we're not in the EU ends the conversation, when the output is used there. Believing the December grace period covers deepfake labels; it only covers providers' machine marking. Hiding disclosures in terms and conditions instead of at first exposure. Relying on a platform's automatic label to meet your own duty. Forgetting audio: in an audio-only ad, say it. And stripping metadata in your export settings, which undermines the provider's marking you're relying on.

6:57 Ask your vendors

Because you rely on your tool vendors for machine-readable marking, ask them four questions, in writing. One: how do you mark generated outputs, with metadata, a watermark, or both? Two: is marking on by default for our plan, or do we have to switch it on? Three: which of your export or download options remove the marking? And four: do you follow the EU Code of Practice on marking and labelling, and if not, how do you meet Article fifty, paragraph two? Put the answers in your vendor list with the date. If a vendor can't answer, that's a signal to consider another tool for EU-facing work.

7:44 Measure it

How do you prove it's working? Three measures. Every EU-facing chatbot and voice agent opens with an AI disclosure, tested monthly in every language you support. Every EU-facing asset has an Article 50 screen row in your asset register. And your vendor list records how each generative tool marks outputs, metadata, watermark or both, and when you last checked. With those three, you can answer a client or a regulator in minutes, not weeks.

8:16 Recap + try this now

Recap. Article 50 applies from the second of August, twenty twenty-six, and follows your output into the EU. Providers disclose chatbots and mark generated content; deployers label deepfakes and unreviewed public-interest text. The grace period to December is only for providers' machine marking. Use the Code of Practice and the EU icon as your reference. Try this now: add the seven-question screen to your campaign brief template and run it on three assets you're working on this week. Next, we open up Content Credentials and watermarks.