Europe Delayed Its AI Law. The Ad Labelling Rules Stayed

The European Union spent the spring softening its own AI law, industry got the postponement it wanted on the expensive obligations, and most of the compliance coverage moved on. Then August 2 arrived and a different set of rules became enforceable exactly on schedule. Among them are the ones that decide whether an AI-generated advertisement has to say so.

What switched on in Brussels on August 2

Three things took effect on August 2, 2026. Article 50, the transparency chapter, now binds providers and deployers of AI systems. The AI Office gained enforcement powers over general-purpose models. And the penalty regime came fully online, with the tier covering transparency failures reaching 15 million euros or 3 percent of worldwide annual turnover, whichever is higher.

Article 50 splits into four duties. A provider whose system interacts directly with a person has to say so, unless it is already obvious in context. A provider whose system generates synthetic audio, image, video, or text has to mark the output so it is machine-readable and detectable as artificially generated. A deployer publishing a deepfake has to disclose that the content was artificially generated or manipulated, with a lighter notice permitted for plainly artistic, satirical, or fictional work. A deployer publishing AI-written text on a matter of public interest owes a disclosure too, unless a human reviewed it and someone carries editorial responsibility. Sidley's breakdown of the provider and deployer split is the clearest short read on which duty lands where.

The obligations Brussels chose to postpone

August 2 was supposed to be a much bigger date. It was the original deadline for the Annex III high-risk regime: risk management, data governance, technical documentation, logging, human oversight, and conformity assessment for AI used in employment, education, credit, essential services, and biometrics. Those obligations moved to December 2, 2027, and the Annex I regime for AI embedded in regulated products slid to August 2028. Gibson Dunn's account of the omnibus agreement walks through the full trade.

Transparency did not move. The single concession is a four-month grace on the machine-readable marking duty for generative systems already placed on the market before August 2, which buys those providers until December 2, 2026. Every other piece of Article 50 is live now, and so is the money.

Why an EU rule reaches a Montreal marketing desk

Article 2 of the AI Act reaches providers and deployers established outside the Union whenever the output of the system is used inside it. A Canadian wealth firm serving impressions to prospects in France sits inside that sentence regardless of where its head office is, and so does a Toronto agency cutting synthetic video for a client with European customers.

This is the second time in three months that an ad-side AI disclosure rule has landed on North American marketers from a regulator they do not report to. New York's synthetic performer law took effect June 9 at 1,000 dollars for a first violation and 5,000 after. Brussels arrived on August 2 with a ceiling four orders of magnitude above that. Neither regulator asked whether the advertiser thinks of itself as an AI company.

Labelling is a generation problem

The operational trap in Article 50 is where the duties sit. Machine-readable marking attaches to the provider of the generative system, so the signal has to be written at the moment of generation. Deepfake disclosure attaches to the deployer, so the advertiser has to know at publication time which assets in a campaign contain synthetic humans, synthetic voices, or manipulated footage.

That second requirement is where most marketing stacks fail. A campaign assembled across a dozen tools produces assets whose provenance nobody recorded. The image came from one model, the voiceover from another, the headline from a third, and the person approving the final cut has no register of which pieces are synthetic. Answering a supervisory authority six months later requires that record to have existed when the asset was made. Rebuilding it afterward is guesswork with a 15 million euro ceiling attached.

Provenance metadata is the intended fix, and the standards work is real. The Coalition for Content Provenance and Authenticity, backed by Amazon, Google, Meta, Microsoft, and OpenAI, publishes a specification for attaching certified origin information to a file. Metadata is also fragile. Any pipeline that re-encodes, crops, or re-uploads an asset can strip the signal without anyone noticing, which is why the marking duty and the disclosure duty are written as two obligations instead of one.

Where the disclosure has to live

For a regulated advertiser, Article 50 lands on a review process that was already the slowest part of the operation. A wealth firm's legal queue was built to check performance claims, testimonials, and jurisdictional disclosures. Adding a per-asset synthetic content determination on campaigns that now generate forty variants at a time makes it longer without making the answers better.

The version that works keeps the record where the asset is made. That is the argument behind LeadLord, the AI marketing platform built for wealth management and other regulated industries: the compliance state of a piece of creative belongs inside the draft rather than reconstructed at the end. A system that generated an image knows it generated the image. It can carry that fact into the campaign object, attach the notice the deployer owes, and hand a reviewer an asset that already declares what it is, in the same pass that checks which return figure the copy can substantiate. When one platform assembles the campaign and also makes each piece, the provenance record is a byproduct of the work. When a firm stitches together six point tools and an outside agency, the record has to be rebuilt by hand, and in practice it is not built at all.

What to watch before December 2

Three dates matter more than the rest. December 2, 2026 closes the grace period on machine-readable marking, and any provider still shipping unmarked output past that point is exposed. The harmonised standards and the Commission's code of practice on marking are still in draft, so firms are designing against statutory text rather than a technical specification. And the first enforcement action aimed at an advertiser rather than a model provider will settle whether national authorities read Article 50 as a filing requirement or a live one. Canada, running its own transparency consultation until September 23, will be reading the same signals.