Practical guide · Article 50 of the EU AI Act

How to label AI-generated advertising, channel by channel, before 2 August.

From 2 August 2026, identifying advertising generated or altered with AI is mandatory across the European Union. This guide explains, in operational terms, which pieces need a label, how to label them channel by channel, what machine-readable marking is and who is accountable for it. No legal jargon.

Written by EficiencIAl Studio, an AI-native production company that delivers masters already labelled and documented

In one sentence: Article 50 of the EU AI Act requires, from 2 August 2026, that advertising generated or altered with AI be identifiable across the European Union, and that is met in three layers: a visible label adapted to the channel, machine-readable technical marking inside the file, and documentation proving the criteria applied.

Regulation
Regulation (EU) 2024/1689, Art. 50
Mandatory from
2 August 2026
Applies to
Brands, agencies and production companies
Scope
Pieces broadcast in the EU, wherever they're produced
The three layers
Visible label · technical marking · documentation
Penalties
Up to €15 million or 3% of turnover

Post-Brexit note: the obligation is tied to where a piece is broadcast, not to where it's produced or the advertiser's nationality. A UK brand running a campaign in EU markets has to label it exactly like an EU-based advertiser would.

What the law requires

Three obligations, told without jargon

Article 50 of the AI Regulation introduces transparency duties that land squarely on advertising. Reduced to the operational essentials, here they are.

1 · That it's visible

The audience must be able to recognise that an advertising piece was generated or altered with AI: a clear identification, adapted to the medium where it airs.

2 · That a machine can read it

The file must carry technical marking, metadata and content credentials, that declare its artificial origin even if the visible label disappears after recompression or cropping.

3 · That bots identify themselves

If the campaign includes a chatbot or virtual assistant, the user has to know they aren't talking to a person.

In June 2026, the European Commission published its Code of Practice on marking and labelling AI-generated content. The rules of the game are already written.

First decision

Which pieces need a label and which don't

The right question isn't "did we use AI?" but "did AI generate or substantially alter what you see or hear?" This is the practical test we apply.

CaseLabel?Why
Video, image or voice generated from scratchYesThis is the central case the law addresses: synthetic content the audience could mistake for real.
Real footage substantially altered (scenes added, faces modified, cloned voices)YesThe alteration changes what the viewer believes they are seeing or hearing.
Synthetic people or digital doubles that appear to be realYes, with reinforced clarityThis is deepfake territory, the case that will draw the most scrutiny.
Minor tweaks: colour correction, audio clean-up, stabilisation, removing a stray cableUsually notThese are editing-assistance functions that don't substantially alter the content.
Grey area: generated backgrounds, shot extensions, generative fillCase by caseIt depends on how much of the piece is generated. This is where documenting your reasoning protects you.

House rule: when in doubt, label it and write down why. A well-integrated label doesn't penalise creativity; getting called out in public for skipping it does.

Second decision

How to label on each channel

The law demands clarity, not one fixed formula. This is our recommended practice, channel by channel.

ChannelRecommended visible labelAlso
Television and VODA brief, legible on-screen caption at the start or end of the pieceA mention in the broadcast documentation the production company delivers
Social networks and platforms (Meta, TikTok, YouTube)The platform's own AI-disclosure toggle, switched onVisible identification on the piece itself too: the toggle doesn't travel with the video if it's reused elsewhere
Web and displayA mention in the creative or on the campaign's landing pageTechnical marking in the published files
Digital out-of-home (DOOH)On-screen labelling built into the piece, legible at the distance the format is viewed fromDocumented criteria per screen version
Audio, radio and podcastA spoken or written mention in the materials accompanying the spotMetadata in the audio master

The layer almost everyone forgets

Machine-readable marking

The visible label is half of compliance. The other half lives inside the file.

What it is

Metadata and content credentials embedded in the master that declare its artificial origin so a system can detect it, even if the piece is recompressed, cropped, or loses its on-screen caption along the way.

How it's handled in practice

It's applied when exporting the final masters, once per version, as part of the post-production workflow. Built in properly, it takes minutes per piece; bolted on afterwards, it forces you to reopen the whole production.

Third decision

Who is accountable for what in the chain

If the piece airs in European territory, the obligation reaches the whole chain, wherever the content was generated.

The brand

It's the one broadcasting and the one that answers to the public and the regulator. It must be able to show the campaign went out compliant: a written record and rationale.

The agency

Devises the campaign and commissions the production. It's on the agency to ask about labelling and rights before signing off, and to set it out in the contract.

The production company

Executes and delivers the masters. Its masters must already carry the label, the technical marking, and a record of which shots are generated.

What keeps all this in order is a contract clause that allocates Article 50 responsibility. Ask us by email and we'll send you a ready-to-adapt template.

Before you broadcast

The six-point checklist

What we review on every campaign before it goes out the door.

1 · Inventory

Which pieces and versions make up the campaign, and which ones carry AI-generated or altered content.

2 · Criteria

A labelling decision per piece, with the reasoning in writing, including for the ones that don't get a label.

3 · Label per channel

Visible identification adapted to each medium where that version will air.

4 · Marking the master

Metadata and content credentials embedded in every final file.

5 · Rights

A documented chain of rights for music, voice and image, with the models used on record.

6 · Contract and archive

A signed accountability clause and an archive of broadcast versions, in case anyone asks tomorrow.

The next step, free

Get the checklist and the contract clause template by email

Ask for the six-point checklist and the ready-to-adapt Article 50 contract clause, and we'll send both straight to your inbox. No sales call attached.

Request the checklist by email

Prefer not to deal with it?

We'll take care of it

We produce your piece and deliver it already compliant, or we audit and label the ones you already have: label per channel, technical marking on the master and a compliance record. Fixed quote in 24 hours.

See the compliance service

Frequently asked questions

The questions we get asked most

When does the obligation to label AI advertising take effect?
On 2 August 2026, across the whole European Union, with a sanctions regime attached. That is the date of application for the transparency obligations under Article 50 of the AI Regulation (AI Act). If a piece keeps running after that date, it must comply even if it was produced earlier.
Do I have to label a piece if I only used AI to touch it up?
As a practical rule: minor editing adjustments (colour, audio clean-up, stabilisation) don't usually trigger the obligation. What does trigger it is generating image, video or voice from scratch, and substantially altering real material, such as adding scenes, changing faces or cloning voices. The grey area gets resolved by documenting your reasoning piece by piece.
How do you label a TV commercial made with AI?
The regulation doesn't impose one fixed formula: it requires the content to be clearly identifiable. Our recommended practice on television is a brief, legible on-screen caption at the start or end of the piece, plus a mention in the broadcast documentation the production company delivers.
What is machine-readable marking?
Information embedded in the file itself, in the form of metadata and content credentials, that lets a system detect the piece was generated or altered with AI even if the visible label gets lost along the way. It's the second layer Article 50 asks for, and the one almost everyone forgets.
Don't platforms like Meta, TikTok or YouTube already handle this for me?
Their disclosure toggles help, and you should use them, but they don't cover everything: the obligation on the piece belongs to the advertiser and its chain, the platform doesn't apply technical marking to the master, and the same piece is usually broadcast outside them too, on your website, on television or on outdoor screens.
Who is responsible if a campaign isn't labelled?
The whole chain is exposed: the brand that broadcasts, the agency that devises and the production company that delivers. That's why it's worth fixing by contract who answers for what, with a specific Article 50 compliance clause. Ask us by email and we'll send you a ready-to-adapt template.
What are the penalties for not labelling?
The AI Act's sanctions regime allows for administrative fines that, for breaches of the transparency obligations, can reach up to 15 million euros or 3 percent of global annual turnover. In practice, the damage usually lands sooner through another route: an undeclared AI ad called out on social media burns brand trust within hours.

Sources: Regulation (EU) 2024/1689 (AI Act), Article 50 and its sanctions regime; European Commission Code of Practice on marking and labelling AI-generated content (June 2026). This guide is operational guidance, not legal advice: for specific cases, consult your legal counsel.