Europe Delayed Its AI Law. The Ad Labelling Rules Stayed

The European Union postponed major parts of its high-risk AI regime in July. That gave businesses more time for some demanding obligations, but it did not provide a blanket pause for AI advertising. Article 50's transparency rules apply from August 2, 2026, subject to a limited transition for some providers. The practical question is which duty attaches to the system, the content and the business using them.

What switched on in Brussels on August 2

Article 50 now applies to covered providers and deployers. The penalty chapter, apart from Article 101 on general-purpose-model fines, already applied from August 2, 2025. Transparency infringements fall within Article 99's ceiling of €15 million or 3% of an undertaking's preceding-year worldwide turnover, generally whichever is higher. SMEs, including start-ups, and qualifying small mid-cap enterprises receive the lower ceiling for this category. An actual fine depends on the infringement and the statutory factors; an unlabelled asset does not automatically produce the maximum penalty. See Articles 99 and 113.

The Commission's transparency guidance separates the duties by role and use. Providers must inform people about direct AI interaction unless it is obvious in context, and must mark generated audio, images, video and text in machine-readable form so their artificial origin can be detected. Deployers must notify people exposed to emotion recognition or biometric categorisation, disclose deepfakes, and disclose AI-generated or manipulated text published to inform the public on matters of public interest.

Those categories have qualifications. The marking rule excludes specified standard editing or changes that do not substantially alter the input or its meaning. Public-interest text has an exception for human review or editorial control combined with editorial responsibility. Evidently artistic, creative, satirical or fictional work receives an adapted deepfake disclosure, not a complete exemption. Article 50 also contains narrowly framed law-enforcement exceptions; ordinary advertising does not acquire one simply by describing a campaign as educational.

The obligations Brussels chose to postpone

The enacted Digital Omnibus on AI moved Chapter III's principal high-risk requirements and operator obligations to December 2, 2027 for systems classified under Article 6(2) and Annex III, and August 2, 2028 for those classified under Article 6(1) and Annex I. These are the rules covering matters such as risk management, documentation and human oversight. A system's actual classification still matters; mentioning employment, education or credit does not finish that inquiry.

The same amendment added Article 111(4). Providers of generative systems already placed on the market before August 2, 2026 have until December 2, 2026 to comply with Article 50(2)'s machine-readable marking duty. This transition concerns those systems and that provider duty. It does not postpone a deployer's otherwise applicable deepfake disclosure until December.

Why an EU rule reaches a Montreal marketing desk

Article 2(1)(c) reaches third-country providers and deployers where their system's output is used in the Union. A Montreal firm's campaign delivered in France can therefore raise an EU compliance question. The analysis still needs the firm's role, the system's use and the relevant Article 50 category. European customers or impressions are a reason to investigate scope, not proof of an infringement by every participant in the supply chain.

For a marketing team, that makes the handoff between firm, agency and model vendor consequential. Who supplies the system under its own name? Who uses it under its authority? Who controls publication? Answering those questions is more useful than asking whether the advertiser considers itself an AI company.

Labelling is a generation problem

A synthetic voice or person is not automatically a deepfake. Article 3(60) requires resemblance to existing persons, objects, places, entities or events and a false appearance of authenticity or truth. That definition must be applied to the finished creative. A provider's broader marking duty can apply even when the deployer's deepfake category does not.

Capturing provenance during creation is a practical way to support that assessment. Consider a campaign whose image, narration and headline came from different tools. If the reviewer cannot identify their origin or later alterations, deciding which disclosure is due becomes harder. A record of the tools, edits and publication decision gives the review something concrete to work from. This is a workflow recommendation, not a claim that Article 50 prescribes one universal asset register or one instant at which every technical signal must be written.

The Code of Practice provides separate provider and deployer measures. Technical marking needs to be effective, interoperable, robust and reliable as far as technically feasible. A particular metadata format, including C2PA, is not named in Article 50 as the sole mandatory solution. The useful acceptance test is what remains detectable and what the audience sees after the actual publishing process.

Where the disclosure has to live

Article 50(5) requires the information for the people concerned to be clear, distinguishable and accessible, by their first interaction or exposure. An internal provenance record does not by itself give the audience the notice it is owed. Review needs to examine both the asset and the version actually delivered.

For a wealth firm, that work sits alongside review of performance claims, testimonials and jurisdictional disclosures. It is easier to assess the final variant when its origin, edits and proposed notice travel with it. That is the design argument behind Leadlord: keep the evidence and review decision with the creative. Such a workflow can support compliance, but the system's classification, the notice, the publishing channel and the firm's legal obligations still need checking. Combining tools in one platform does not establish legal compliance by itself.

What to watch before December 2

December 2, 2026 ends the specified transition for previously marketed generative systems. A provider relying on it needs a plan for the applicable marking duty; a deployer needs to meet its own current disclosure obligations.

The code is already final: it was published June 10, and the Commission's July 31 update reports about 190 signatories and a positive adequacy assessment by the Commission and AI Board. It is a voluntary, recognised means of demonstrating compliance, with adequate alternatives available. It does not replace the Act or turn every implementation choice into a mandatory technical standard. The Commission's guidance complements it on scope and exceptions.

The next useful evidence will be how those measures perform in real publishing pipelines and how authorities apply the rules to particular facts. A firm can prepare by testing its notices and markings now, recording the basis for exceptions, and making responsibility for each campaign explicit.