AI Content Labelling Is Now the Law in the EU.

On 2 August 2026, Article 50 of the EU AI Act became applicable.

Marking AI generated content stopped being a platform preference and became a legal obligation, backed by administrative fines of up to 15 million euro or 3% of total worldwide annual turnover, whichever is higher. For small and medium enterprises the cap is the lower of those two figures.

The headline is not the interesting part. The paperwork underneath it is.

Somebody in your organisation now has to be able to say which asset used AI, at which stage, whether the change was substantial, and who approved publishing it. Most approval processes have no field for any of that, and no named owner for the decision.

That is a workflow change, not a policy document.

What Article 50 actually asks for

The article splits the work between two roles, and a brand is almost always the second one.

Providers, meaning whoever builds and supplies the generative system, have to mark synthetic audio, image, video and text in a machine-readable format so it is detectable as artificially generated or manipulated. That is the watermarking and metadata layer, and it happens inside the tool.

Deployers, meaning the organisation using that system in its own professional activity, carry the visible half. A deployer whose system generates or manipulates image, audio or video amounting to a deepfake has to disclose that the content was artificially generated or manipulated. AI generated text published to inform the public on matters of public interest is covered too, unless it went through human review and a named person or company holds editorial responsibility for it.

Article 50(5) sets the timing. The disclosure has to reach the person in a clear and distinguishable manner, at the latest at the time of first exposure. That points at the asset, not at a note somewhere else.

A voluntary Code of Practice on marking and labelling was finalised on 10 June 2026 and accepted in July as an adequate way to show compliance. It carries a common EU label set any deployer can use.

The exemption most brand content will rely on

Not everything AI touches is caught. The marking obligation does not apply, in the regulation’s own words, “to the extent the AI systems perform an assistive function for standard editing or do not substantially alter the input data provided by the deployer or the semantics thereof”.

Colour grading a clip. Cleaning up audio. Subtitles from speech that was genuinely said. Tidying the grammar in a caption. That is standard editing, and it sits outside.

Generating a product shot that was never photographed, cloning a voice, swapping a face, writing the script the whole video is built on. Different category.

Most real production sits between the two. The exemption is a test somebody has to apply, not a list somebody can look up, which makes it a question about how AI is used inside the content team rather than a question about the law.

Who is caught, and the honest answer for a UK brand

Article 2 pulls in providers placing systems on the Union market wherever they are established, deployers established or located in the Union, and providers and deployers in a third country “where the output produced by the AI system is used in the Union”.

That last clause is the one UK, UAE and US brands keep asking about, and it does not resolve into a clean yes or no on the text alone. A London brand with European followers is not obviously the same case as one running EU facing accounts through an EU entity.

NBK is a social operations partner, not a law firm, and any team with real EU exposure should take its own advice. What we would say is narrower. Build the record either way, because it costs very little and it is the only thing that makes the question answerable.

What is not covered, and what did not move

Three things are worth knowing before anyone starts auditing the archive.

• Content generated before 2 August 2026 does not have to be labelled retroactively. The Commission encourages it where possible, but does not require it.
• The obligation on AI generated text turns on publication, so something drafted before the date and published after it is a different case from a video generated and posted in June.
• Generative systems already on the market before 2 August 2026 have until 2 December 2026 to meet the marking obligation. That is a supplier deadline, but it explains why your tools may not be marking their output yet.

The transparency date itself did not move. When the EU pushed its high-risk classification deadlines back into 2027 and 2028, Article 50 stayed where it was.

Enforcement is a separate question. As of mid-June 2026, only nine member states had designated both of the national authorities the regulation asks for, and six had designated neither. That says something about the pace of the next year, though it is not a reason to ignore a live obligation.

The record most content teams do not have

Nothing in Article 50 asks a brand to keep a log. The obligation is to disclose, and to disclose clearly. But you cannot disclose consistently across a few hundred assets a month without knowing which ones need it, and you cannot know that unless it was recorded when the tool was used.

Ask most content teams which of last month’s posts used a generative tool and at what stage, and the answer is a chat thread and somebody’s memory. That was survivable when labelling was a platform toggle.

What the asset log should capture

Six fields, added to whatever your team already uses to move an asset from brief to publish.

• Tool: which system touched the asset, named specifically rather than logged as AI.
• Stage: script, image, voice, video, edit, subtitles or caption. One asset can carry several.
• Change: assistive editing or a substantial alteration, decided by a person rather than assumed.
• Output: whether the finished asset contains generated image, audio or video, since that is what the deployer rules turn on.
• Disclosure: what the viewer was shown, and where on the asset it appeared.
• Approver: the name of the person who signed the publish, not the name of the department.

None of this needs new software. It needs six columns on the approval process you already run, and a rule that an asset does not reach scheduled until they are filled in.

Who signs it off

The third field is the one that gets skipped, because it is a judgement rather than a fact.

Somebody has to decide whether a generated element substantially altered the material or merely assisted the editing, and that person should be named in advance rather than found afterwards. It should not be the editor who made the asset, for the same reason the person writing a claim does not approve it.

Put the decision with whoever already owns the final publish approval.

What good looks like

Imagine a brand shipping forty assets a month across four platforms. Three used a generative tool for something beyond standard editing: one product visual, one voice-over, one sequence of B roll.

Each carries its disclosure on the asset itself, each has a named approver in the log, and the other thirty-seven are recorded as assistive editing with a one-line reason. Nobody held a meeting about it.

NBK treats this as an operations change, because that is what it is. The regulation did not change what good content is. It changed what a finished asset has to carry with it.

Teams that already run social as a system absorb it quickly, because they have an asset log, a defined publish gate and a named approver. Teams without a real approval process tend to find the gap in a more expensive way. The rule is rarely the hard part. The missing system to attach it to is.

Next step

Pull last month’s output and answer three questions: which assets used a generative tool, at which stage, and who approved them. If that takes more than an hour, the gap is in the workflow rather than in anyone’s understanding of the law.

If your social process is slowing down good ideas, or cannot answer a question like that one, NBK can help rebuild the workflow behind the content.