AI Labelling from August: What Companies Need to Decide Now

On 2 August 2026, the EU will require labelling of AI-generated content. For many marketing executives, this still sounds like an abstract regulatory question. In reality, it is production logic – and the time to prepare is running out.

What exactly happens in August?

The EU AI Act requires companies that publish AI-generated images, videos, or audio content externally to label them as such – provided they could be mistaken for real. Studies show: 40 to 75 percent of people can no longer distinguish AI-generated images from real photographs. The trend is rising. What this means for the labelling obligation is clear.

The law distinguishes between two obligations. AI providers must make their outputs technically recognisable as synthetic – through watermarks or metadata. Companies that publish this content carry the obligation outward: visibly, unambiguously, at the content itself.

How far does the obligation reach?

Many companies underestimate the actual scope of the obligation. The law is based solely on the overall visual impression.

The only relevant factor is the overall visual impression: whether an average person would take the content to be real. How the image was technically created is irrelevant. Photorealistic product images, synthetic persons in advertising scenarios, AI-placed products in realistic environments – all of this falls under the obligation. Exempt are contents that are obviously stylised or fictional and carry no risk of confusion, as well as content used purely internally. When in doubt: label rather than not.

The real risk lies in the production pipeline

When compliance managers think about AI labelling, they think first about individual decisions: does this image need to be tagged? The actual weak point is elsewhere.

In companies that produce AI-generated assets at scale, the risk emerges in everyday operations: an employee generates an image and uploads it directly – without compliant labelling. Not out of bad intent, but because the process did not prevent it. The liability risk lies with the company as publisher, not with the AI tool provider.

What does this mean in practice for preparation?

Companies using AI in image production should now answer three questions:

Which AI tools are being used, and which outputs go into external communications? Not every use of AI is subject to labelling – but many organisations still lack an overview.

Who ensures that labelling is in place before publication? As long as this responsibility is not clearly assigned, compliance depends on individuals.

Is this requirement integrated into approval processes? A policy that no one sees in daily operations does not protect.

The comparison with the GDPR introduction is apt. In 2018, many companies treated the regulation as an external imposition, delayed implementation, and waited for clarification. The result was a patchwork of retroactively built solutions. Companies that take the same attitude towards the AI labelling obligation today are buying time at the cost of security.

Transparency as positioning

By August 2026, further EU guidelines and a code of conduct on labelling are expected. The basic rule is already clear. Those who act early build trust – with customers, partners, and regulators.

Labelling as a product feature

Open Wonder addresses the labelling obligation not through downstream checklists, but as a fixed part of the generation logic. Every image the platform produces automatically receives a visible label: "AI-generated by Open Wonder". It sits on the asset, not in the process. No manual step, no approval logic that can fail.

This aligns precisely with the requirement in Article 50 of the EU AI Act: providers of AI systems must make their outputs technically recognisable as synthetic. Open Wonder translates that into something marketing teams can actually use, without involving IT or legal for every single asset.

It started with a number: 40 to 75 percent of people can no longer distinguish AI-generated images from real photographs. That number will continue to rise. In this world, AI labelling will not be a competitive disadvantage – it will be the normal state, as self-evident as a legal notice. Companies that build this into their processes today will barely register it as a separate topic in two years' time. Those who wait until August will build under time pressure what can be put in place now with care.

Tim Herzog is co-founder and CEO of Open Wonder (openwonder.com), the AI Brand Operating Platform.

Also published at Meedia.