Article 50 of the EU AI Act became applicable and enforceable across the EU on August 2. The high-risk obligations that shared that date moved to December 2027 and August 2028 under the AI Omnibus, which entered into force July 27.
The split matters for marketing. High-risk classification applies to hiring systems, credit scoring, and similar decision systems. Article 50 applies to chatbots, synthetic imagery, ad copy, product shots, and AI-generated text published on matters of public interest. That list describes the marketing content pipeline.
The August 4 edition argued that verification cost sets the ceiling on how far agentic execution scales. As of August 2, part of that cost is a documented obligation attached to the output. This edition covers what that requires operationally.
Verification moved from a review step to a property of the asset
Article 50 imposes three requirements that operate together. Providers of generative systems must mark outputs in a machine-readable, detectable format. Deployers must disclose synthetic content depicting real persons, objects, places, entities, or events. Article 50(5) requires that disclosure reach the audience in a clear and distinguishable manner at the latest at the time of first interaction or exposure.
Together these describe a chain of custody. Marking originates at generation. The disclosure has to survive assembly, approval, trafficking, distribution, and platform ingestion, and still be present when a person sees the asset. A single review before launch does not satisfy that requirement. The signal has to travel with the file.
Marketing operations teams know how brittle that chain is. Metadata gets stripped during transcoding and re-encoding. Assets get cropped, resized, versioned, and re-exported. A campaign that clears legal review in March can appear in a regional feed in September with the provenance record gone.
Under the Trust and Verification layer of my AI capability framework, this is the failure mode that appears under Generation: whether the artifact is authentic and attributable. Attribution only holds if the asset carries it.
The Commission has published guidelines and a code of practice covering marking and labelling, and adherence to the code is the recognized route to demonstrating compliance. Generative systems already on the market before August 2 have until December 2, 2026 to meet the marking obligation. That is a four-month window to instrument a pipeline most organizations have not instrumented for this purpose.
It reminds me of what Phyllis Fang, Head of Marketing at Transcend, said when I interviewed her on The Agile Brand podcast: “All that shadow data floating throughout an organization’s tech stack…creates unseen risks that erodes transparency. And as a marketer, as an operator, you don’t always know what data is permissible to use…” The same visibility gap now applies on the output side. A team that cannot say which assets currently in market were AI-generated has a provenance problem with a deadline attached.
The obligation follows whoever controls the AI use
Davis+Gilbert’s July 31 alert covers the scope question that will affect brand and agency contracts first. The Commission’s guidelines confirm that a non-EU advertiser using AI to generate a deep fake for an advertisement displayed in the EU may be treated as a deployer. Physical presence in Europe does not determine the answer. Whether the output reaches an EU audience does.
The definition of “deep fake” is broader than the common US usage. It covers realistic AI-generated depictions of objects, places, animals, events, and human avatars that resemble no identifiable person. The Commission’s examples include an AI-generated product image in advertising or packaging that could mislead an audience about the product’s actual appearance, characteristics, or use. Routine color correction and audio cleanup fall outside the definition. De-aging a performer or simulating part of a performance falls inside it.
The creative carve-out is narrow. Article 50(4) reduces disclosure requirements for evidently artistic, satirical, or fictional work, and the guidelines state that where content combines commercial and creative characteristics, the commercial character generally prevails. Commercial work will rarely qualify.
The text exception is the provision worth planning around. AI-generated text used to inform the public on matters of public interest requires disclosure, and that requirement falls away where the text receives meaningful human review and editorial control and a named human or editorial entity takes ultimate responsibility for it.
This is the Book 2 commitment with a compliance consequence attached. The framework treats humans and AI agents as interchangeable team members on execution and holds humans accountable for direction. European law now assigns a legal effect to that division of responsibility.
It reminds me of what Stephanie Liu, Senior Analyst at Forrester, said when I interviewed her on The Agile Brand podcast: “People don’t like to be deceived…how would you feel if a brand was using AI and didn’t disclose it, and the number one response was deceived or angry or upset?” The regulation codifies a preference consumers already held.
Labeling arrives on top of a relevance problem it does not solve
Optimizely released the latest results from its Marketer’s Survival Guide on August 4, based on a survey of 1,000 UK consumers and 100 UK marketers. Three-quarters of consumers say they regularly receive marketing that is irrelevant to them. Sixty-nine percent get duplicate messages or emails. Sixty-one percent feel overwhelmed by the volume. Forty-two percent disengage when content feels irrelevant, and 35% become more likely to unsubscribe after a poor experience.
The marketer-side figures show the mechanism. Sixty percent say they launch campaigns without the time or data to optimize them. Fifty-four percent move campaign to campaign without evaluating performance. Sixty-six percent say managing multiple tools and platforms creates unnecessary work. Optimizely’s Tara Corey summarized the result as AI helping marketers produce more of the same.
Read alongside Article 50, the operational picture is straightforward. The regulation requires an accurate origin signal on the output. The survey says a large share of that output already fails to earn attention. Labeling a message the recipient did not want makes its origin legible at the moment they are least receptive to it.
My fifth Agile Brand Principle puts the customer relationship and the experience at the center of the work. Disclosure obligations give that principle an enforcement mechanism. They also raise the cost of volume without raising its value.
The next disclosure surface is a brand-configured agent
MarTech reported on August 4 that OpenAI appears to be testing a ChatGPT ad format where a click opens a business-specific conversation in place of a landing page. The described workflow crawls a company’s website to generate a business profile, then lets advertisers configure an agent using custom instructions, product feeds, and Model Context Protocol tools for live business data. The agent can answer questions, recommend products, schedule appointments, and qualify leads. The capability appears in ChatGPT Ads Manager for a limited group of advertisers, and the end-user experience has not been widely observed.
Two framework layers apply to that unit. Interface: Article 50(1) requires that people be informed when they are dealing with an AI system. Identity and Permissions: the governance question is what the agent is authorized to say on the brand’s behalf, which claims it may make about the product, and what audit trail exists when it makes a claim it should not have. An agent that qualifies leads is making commitments in a live commercial conversation, and configuration through custom instructions is a weak control surface for that.
Featured Insights
European Commission. Safer and more transparent AI. August 2, 2026. The primary announcement of the Article 50 transparency rules taking effect: marking and labelling of AI-generated content, disclosure when users interact with an AI system, and notification for emotion recognition and biometric categorisation. National market surveillance authorities, the AI Office, and the European Data Protection Supervisor enforce, with fines reaching €15 million or 3% of global annual turnover. Start with an inventory. You cannot mark what you cannot find. Build the register of AI-touched assets in market before you build the labeling workflow.
Davis+Gilbert LLP. EU AI Act Guidance Expands AI Disclosure Rules for Advertisers and PR Teams. July 31, 2026. The clearest practitioner read on scope. The Commission’s final guidelines broaden “deep fake” past digital replicas of real people, extend reach to US advertisers whose output lands in front of EU audiences, and preserve a text exception for content under meaningful human editorial control. Resolve deployer status in your agency contracts now. The obligation follows decision-making control over AI use, and both sides currently assume the other one holds it.
MarTech. OpenAI ad experiment could change what happens after the click. August 4, 2026. Constantine von Hoffman reports on a ChatGPT ad format that replaces the destination page with a configurable business agent, assembled from a crawled business profile plus product feeds and MCP tools. Treat the agent’s instruction set as regulated brand collateral. Version it, review it, and log what it says, the same way you already do for claims in a broadcast script.
Optimizely, via DecisionMarketing. Brits bemoan repetitive, poorly targeted and generic ads. August 4, 2026. Latest results from the Marketer’s Survival Guide: 75% of UK consumers regularly receive irrelevant marketing, 69% get duplicates, and 66% of marketers say tool sprawl creates unnecessary work. Use the compliance work as cover for a volume audit. The same inventory that finds your AI-touched assets will find how much of your output nobody wanted.
Key Takeaways
- Provenance has to survive the pipeline. Marking at generation satisfies nothing if assembly, trafficking, and platform ingestion strip it. Test the chain end to end on one live campaign before December 2.
- Deployer status is a contract term. Control over how AI gets used determines who carries the obligation. Brands and agencies should name that party explicitly in writing, before an enforcement inquiry settles the question for them.
- Human editorial responsibility is a compliance asset. A named person taking ultimate responsibility for AI-assisted text removes the labeling requirement for that text. Document the review, and document who owns it.
- Disclosure raises the cost of volume. With three-quarters of consumers already reporting irrelevance, every additional labeled asset carries a production cost and a trust cost. The economics favor fewer, better-targeted outputs.
Some parting thoughts
I have run diligence on enough acquisitions to know what a working control looks like. It is documented, traceable to a named owner, and it produces evidence without anyone reconstructing it after the fact. Most marketing content pipelines were never built to that standard, because nobody required it.
That requirement now exists. The useful part is that the work of answering a regulator is the same work as answering a CFO who wants to know what the content operation produces and whether any of it lands. Four months is enough time to instrument one workflow properly and use it as the pattern for the rest.
Original Source:
The Martech Futurist Blog – Greg Kihlström Marketing Technology & Digital Transformation
Original URL:
