Labelling AI Content: What the EU AI Act Requires — and What It Does Not

The short answer

The EU AI Act does not require a label for "somehow AI-assisted" work. It requires one in specific situations: when people interact with an AI system without recognising it, and when synthetic content looks deceptively real. For marketing that means: label chatbots and AI avatars, label photorealistic generated images and video — but not draft copy, research, translation suggestions or analysis work.

The distinction that matters

The decisive question is not "was AI involved?" but "can a person be deceived?".

The line is also practical. Labelling everything devalues the label. A notice that appears on every page stops being read — and protects nobody.

What a good label looks like

Three requirements: visible, understandable, at the moment of perception. A line in the legal notice meets none of them.

• Chat and avatar: a notice before the first interaction, not in the fine print. Keep it plain: "You are talking to an AI assistant." No marketing tone.
• Images and video: a badge on the asset plus one line in the caption. What matters is that the label travels with the asset when it is shared.
• Audio: an announcement at the start or end depending on format.

Metadata belongs alongside. Content Credentials (C2PA) write cryptographically signed provenance into the file and make it traceable later how an asset came to be. They do not replace the visible label, but they survive reuse better than a caption — and they are the format platforms increasingly rely on.

What invisible watermarking realistically delivers

Embedded watermarks in image, audio and text output are useful but not full protection: cropping, recompression, screenshots and further processing degrade them to different degrees. They are worthwhile for internal traceability and platform checks; they do not stand alone as a compliance argument. The robust combination is a visible label for humans, C2PA metadata for systems, and an internal asset register for evidence.

The process that carries it

Labelling rarely fails on technology, almost always on process. What works:

1. Decide at the source. When an asset is created, record whether it was generated, edited or captured. Reconstructing that later is close to impossible.
2. A field in the DAM, not a spreadsheet. Label status belongs on the asset, not in a parallel list.
3. Automate the output. If the field is set, the CMS renders the badge automatically — manual discipline does not scale.
4. Bind agencies and freelancers. Deliverables include provenance, or they are not accepted.
5. Spot checks over full audits. A monthly look at 20 random assets exposes process gaps more reliably than any self-declaration.

What labelling does to trust

The common fear that an AI notice devalues content barely holds up in practice. What damages trust is the later discovery: a generated testimonial presented as real. Open labelling is therefore less a compliance cost than an insurance policy — it removes precisely the risk that gets expensive when it goes wrong.

Next steps

Take stock: which published assets fall under the transparency obligation, which do not, and where is the notice missing today? How we handle it ourselves is documented on our AI transparency page; the concepts are explained in the glossary under content credentials and EU AI Act.