From August 2026, new EU transparency obligations kick in that will affect every marketing and communications team using generative AI, whether they know it or not, writes Colin Hetherington.
From August 2026, the EU AI Act starts asking a simple question of anyone using generative AI in their work: are you telling people?
Article 50, the transparency chapter of the Act, sets out new requirements around whether the people seeing your content know a machine had a hand in making it. For those working in communications and marketing, that’s the bit worth paying attention to. Here’s what we think it means in practice, written as plainly as we can manage.
The four things Article 50 actually asks for
At the heart of this, there are four obligations:
- AI chatbots and assistants must tell people they’re talking to AI, at first interaction, clearly.
- Companies that build generative AI tools (think Google Gemini, ChatGPT, Canva or Adobe Firefly) must mark their outputs as AI-generated in a machine-readable way, typically through provenance metadata and watermarking rather than anything visible to the eye. This is largely a job for the providers.
- Emotion recognition or biometric categorisation, such as cameras that read facial expressions on digital screens or retail displays to gauge audience reaction, must be disclosed to anyone exposed to it.
- Deepfakes, AI-generated content that looks like a real person, place or event, must be disclosed for what it is. So must AI-generated text presented as informing the public on something in the public interest, unless a human has genuinely reviewed it and taken editorial responsibility.
That last one is where most of the judgement calls live, so it’s worth slowing down on.
“Public interest” doesn’t have a neat legal definition yet, but the test is simple: is this content trying to sell someone something, or is it telling them something they might rely on to make a decision? A product ad is the first. A government page explaining how to apply for a grant is the second. If it’s AI-generated and no human has signed off on it, it needs to be disclosed.
Where Article 50 Actually Applies
Article 50 isn’t a public sector versus private sector divide. It’s built around types of AI use and types of content, which means private sector organisations, advertisers in particular, can face real obligations too. If you use AI to generate a realistic image of a product and present it in an ad as if it were genuine, you’re potentially in Article 50 territory.
For brands and agencies, the line gets crossed when content starts to look like real testimony, real events or information people might use to make a decision, about their health, their money, a service they’re entitled to, or something happening in their community. That’s when disclosure becomes the safer default, legally and reputationally.
For public bodies and semi-state organisations, the exposure is broader, because so much of what they publish is, by definition, public interest information. A body like ESB Networks, providing information about outages, infrastructure, safety and services people depend on, sits in the same category as a Government Department for these purposes. AI-drafted guidance on services, benefits or public health that goes out more or less as written is exactly the kind of content Article 50 has in mind. The way through this isn’t to avoid AI. It’s to be honest that a human reviewed it and stands behind it, and to be able to show that’s actually true rather than a rubber stamp.
If you’re relying on the human review exception, the review needs to leave a trace. That means keeping a record of who reviewed AI-generated content, when, and what changed as a result, even if it’s a simple log rather than anything elaborate. “We have a process” isn’t the same as “we can show our process happened,” and the second one is what actually matters if anyone ever asks.
Worked Examples
Say a public health body publishes a campaign promoting vaccine uptake. AI-generated clinic footage, an AI-generated patient, a voiceover framed as a real person’s experience. This is public interest content. Article 50 applies.
If the patient looks real and nothing signals otherwise, it needs a label confirming it’s AI-generated. If the voiceover sounds like a genuine testimonial, that needs disclosure too. And if the written content alongside it, a FAQ, a guidance note, social captions presenting health facts, was AI-generated without documented human review, you’ll be obliged to disclose that too.
Say you’re producing a social media video for a major infrastructure scheme. AI-generated footage of the finished project, an AI-generated community member welcoming the development, a location styled to look like a real Irish town. It’s a promotional video, not public information, so the public interest disclosure doesn’t apply.
The question that does apply is simpler: could someone mistake this for real? If the community member looks like a genuine person in a genuine place and nothing signals otherwise, that’s a deepfake in Article 50’s terms. Label it. If the style is clearly rendered or illustrative, a label isn’t required, though a credit is still good practice. The voiceover on its own won’t need disclosure unless it’s framed to sound like a real person’s testimony.
The Grey Areas, Honestly
Some of this is still settling, though the picture has sharpened recently. Under Article 50, chatbot disclosure, emotion-recognition disclosure and deepfake disclosure all apply from 2 August 2026, while the revised deadline for watermarking and content-labelling obligations for existing systems is 2 December 2026. The public interest threshold is still being clarified in practice, and what counts as a deepfake versus an obviously stylised AI image is, in places, a judgement call.
Our view: where it’s genuinely unclear, disclose anyway. A small label costs almost nothing and no one is going to be shocked that you’re using AI to help your communications. The absence of one, if someone later asks why your “real” testimonial was AI-generated, costs a lot more. Breaches of Article 50 sit in the same penalty band as the AI Act’s high-risk requirements, fines of up to €15 million or 3% of worldwide turnover. For most organisations the bigger risk is reputational, not regulatory, but both are worth having on the radar.
A Practical Cheat Sheet
| Scenario | Covered by Article 50? | What to do |
| AI voiceover doing straightforward ad narration (no claim to be a real person) | Generally no | Good practice to credit it as AI-generated in the description, especially if the voice sounds human |
| AI voiceover framed as a real person’s testimonial, for example “I studied at DBS and I’d recommend it” | Yes, this presents as a real endorsement | Disclose that the voice is AI-generated, the testimonial itself needs to be genuine regardless |
| AI-generated “customer” testimonial video with a realistic face and voice | Yes, this is a deepfake | Disclose clearly, ideally on-screen, not just in small print |
| AI-edited product photography (background removal, retouching, colour correction) | No, this is assistive editing | Nothing required |
| AI-generated hero image of “people” using a product or service, photorealistic | Likely yes if it could pass as a real photo | Label it as AI-generated, or use a style that’s clearly illustrative |
| Chatbot on your website or social channels | Yes | Tell users they’re talking to AI at first contact, and make it obvious |
| AI-drafted blog post or LinkedIn article, reviewed and edited by a person before publishing | No, editorial responsibility exception applies | Make sure that review genuinely happens and could be evidenced if asked |
| AI-drafted public information content (a state body’s guidance page on a scheme or service) published largely as generated | Yes, likely | Either disclose the AI involvement, or ensure documented human review and sign-off |
| AI-generated illustration in a clearly stylised, non-photorealistic format | Generally no | A small AI icon or credit is still good practice and likely to become standard |
| Software that reads audience emotion or demographics from a camera (digital signage, event activations) | Yes | Inform people it’s in operation, and think hard about whether you want to be doing this at all |
What to do before August
Look at where AI sits in your content production, across imagery, copy, voice and any conversational tools. Decide, channel by channel, what needs a label and what doesn’t, using the questions above as a starting point. If you work with a state body, get clear on what “editorial responsibility” looks like in practice for AI-drafted content, and make sure it’s real, not nominal. And keep an eye on the EU’s Code of Practice, which is heading towards a common AI icon that may become the simplest way to handle all of this.
None of this should change whether you use AI. It should change whether you’re upfront about it. Given how quickly people can tell the difference anyway, that’s not really a cost. It’s just good practice arriving slightly ahead of the law.
Colin Hetherington is co-founder of Common Good, an agency working with public bodies, semi-state organisations and purpose-driven brands on strategy, communications and campaigns.
