The label is no longer a nice-to-have buried in a platform setting. From today, August 2, 2026, the EU AI Act's transparency rules start applying to a large part of the AI content supply chain.
That matters to marketers because the rule does not stop at the company that built the image model. If your brand or agency decides to use AI to create an ad, controls how it is used, or approves where the asset goes, you may be a deployer under Article 50. The agency and the advertiser can both carry responsibility for the same campaign.
The practical shift is simple: AI creative needs a review trail before it needs a media budget.
The New Advertising Threshold
The EU Commission's guidance on AI-generated content transparency that Article 50 applies from August 2, 2026. For advertising, the immediate concern is generated or manipulated image, audio, and video content that qualifies as a deepfake.
That word sounds narrower than it is. The test is not limited to a celebrity face swapped into a video. Content can qualify when it appears to show a real person, product, place, event, or object authentically, even if the creator never intended to deceive anyone. A product made to look larger, cleaner, newer, or materially different from reality is more exposed than a purely fantastical illustration.
Small technical edits are a different matter. Colour correction, background cleanup, light retouching, or resizing will not automatically turn every production workflow into a regulatory event. The question is whether the intervention changes what the audience thinks is real.
That distinction is useful, but it is not an excuse to wave everything through. A campaign team that cannot explain what changed, who approved it, and where it will run is already missing the operational control the rule expects.
The Agency Is Not Invisible
Advertising has always split responsibility across a chain. The brand owns the brief. The agency chooses tools and makes the work. The platform distributes it. Production vendors handle assets. AI makes that chain harder to hide because the most consequential decision can happen inside a prompt, a model setting, or a fast creative iteration that nobody recorded.
The new rules treat authority broadly. The party responsible for deciding why and how an AI system is used can count as a deployer, even without technical control of the model. In a normal campaign, that can mean both the advertiser and the agency, each controlling a different part of the process.
This is where the usual contract language falls short. “Client approves final creative” does not answer who selected the model, what reference material went into it, whether the output altered a real product, or whether the label stayed attached after resizing and syndication.
The European Commission's code of practice recommends machine-readable marking and visible disclosure, but an invisible watermark is not a substitute for a human-perceivable label where one is required. A platform's built-in AI marker may help. It does not erase the advertiser's responsibility.
Labels Must Survive Distribution
The most common compliance fantasy is that someone can add a note at the end of the workflow. The creative gets generated, approved, exported, cropped, translated, whitelisted, and distributed. Then a person remembers the label.
That is not a system. It is a hope with a deadline.
The disclosure needs to be clear, distinguishable, and understandable when the content first reaches its audience. It cannot be hidden in metadata alone. It cannot disappear because the same image was resized for a new placement. If a platform strips the visible indicator, the team still needs to know where the obligation moved.
That creates a new production question for every asset: what is the disclosure state, and does it travel with the file?
A sensible asset record should capture the model or tool used, the type of alteration, whether a real subject or product is represented, the markets targeted, the human reviewer, the exact disclosure copy, and the final approved file. This is less glamorous than a prompt library. It is also much more useful when a client, regulator, or platform asks what happened. It is the same reason I keep returning to the prompt injection problem in marketing: the operational weak point is usually the handoff, not the model demo.
This is the same operational problem behind AI marketing measurement failures. Teams keep asking for better reporting after the underlying event has already been flattened into a platform export. Compliance has the same shape. If the workflow does not preserve the decision, the final dashboard cannot reconstruct it.
The Text Exception Is Not Comfort
Article 50 also addresses AI-generated or manipulated text published to inform the public on matters of public interest. That does not mean every AI-assisted headline or product description needs a conspicuous label tomorrow. Human review and editorial control matter, and routine advertising copy will often sit outside this particular disclosure scenario.
But the exception narrows fast around claims involving health, consumer safety, or sustainability. Those are exactly the claims marketers already need to substantiate under other laws and advertising codes. Adding an AI workflow without adding a source and approval workflow is asking for two problems at once.
The better question is not, “Do we have to label this?” It is, “What would a reasonable person believe this asset is showing, and can we prove how we got there?” That question catches more risk than a checklist built around file extensions.
The EU AI Act's official text sets the broader application framework, while specialist legal analysis puts the advertising exposure more plainly: non-compliance can mean fines of up to €15 million or 3% of global annual turnover, whichever is higher. The label does not make an unlawful ad lawful under privacy, intellectual property, personality rights, or consumer protection rules. It is one layer of responsibility, not a permission slip.
What Teams Should Change Now
Start with the asset inventory, not a policy memo. Pull the creative made with image, video, audio, and generative editing tools in the last quarter. Mark which assets show real people, products, locations, or events and which markets they can reach.
Then make four workflow changes:
- Assign authority. Name the person at the brand and the agency who can approve AI use, not just the person who uploads the final file.
- Keep the evidence. Store the source asset, generated version, material edits, tool, reviewer, and disclosure decision together.
- Test the export. Check labels after cropping, compression, translation, platform upload, and dark-mode or mobile rendering.
- Write the client clause. Say who owns the review, who supplies factual substantiation, and what happens when a platform removes or replaces a disclosure.
This is not a reason to stop using generative tools. It is a reason to read the vendor lock-in trap as a workflow problem, not only a procurement problem. It is a reason to stop pretending that generation is the end of production.
The brands that handle this well will not be the ones with the cleverest label. They will be the ones that can show, quickly and without a committee meeting, what an asset claims to depict and who decided that claim was safe.
That is the uncomfortable part of AI-assisted advertising. The model may make the picture. The liability still belongs to people.