Marketing agencies, publishers, and AI-generated public content

EU AI Act for Marketing Agencies and Publishers: Article 50 Rules for AI-Generated Content

Marketing teams and publishers often think they have a single “AI content” problem, but the real obligations under the EU AI Act depend on public exposure, content type, and editorial control. From 2 August 2026, Article 50 requires providers of generative AI systems to enable machine-readable marking of synthetic audio, image, video, or text outputs. Deployers must disclose deepfake-like videos, synthetic voices, or AI-generated text published to inform the public on matters of public interest—unless that text has undergone meaningful human review or editorial control with a named person or organisation holding editorial responsibility.[1][2]

Internal drafting followed by substantial human editing and clear sign-off usually avoids the disclosure trigger. The distinction matters for agencies producing client campaigns and publishers issuing news or analysis. Article 4 AI literacy requirements (already in force) mean your teams must understand these nuances in context. Focus on practical workflows, consistent documentation, and evidence of editorial responsibility rather than labelling everything.[3]

Law status (May 2026) Article 50 transparency obligations apply from 2 August 2026. Article 4 (AI literacy) has applied since 2 February 2025. The voluntary Code of Practice on marking and labelling of AI-generated content is in its second draft (published March 2026) and remains non-binding until potentially approved; its latest version offers greater flexibility on human review and removes rigid “generated vs assisted” taxonomies. No current amendments alter these core Article 50 rules. Check official EU sources for final developments.[4]

Article 50 creates four main situations relevant to marketing and publishing teams.[3]

• Interactive systems (chatbots): If your website, social campaign, or client service tool uses an AI chatbot that interacts with the public, users must be informed they are speaking with AI unless it is obvious from context. This is straightforward in most marketing deployments but requires clear upfront messaging.

• Deepfake-like content: Videos, synthetic voices, or images that realistically resemble real people, objects, places, or events and could appear authentic trigger a disclosure obligation for deployers. In marketing this appears in celebrity-style endorsements, realistic product demonstrations, or voiceovers that mimic known personalities. Disclosure must be clear and at first exposure. Artistic, satirical, or fictional works have a lighter obligation that must not spoil enjoyment.[1]

• Public-interest text publications: Publishers (and agencies producing thought-leadership or news-like content) must disclose when text informing the public on matters of public interest (e.g., health policy, elections, environmental regulation, or major societal issues) is AI-generated or manipulated. The obligation disappears if the content has undergone a process of human review or editorial control and a natural or legal person holds editorial responsibility. This is the provision most publishers rely on.[2]

• Machine-readable marking for providers: If your agency or publisher develops, fine-tunes, or supplies generative tools, you must ensure outputs can be marked in a machine-readable format (metadata, watermarks, fingerprints) that is detectable. Solutions must be effective, interoperable, robust, and reliable as far as technically feasible, taking cost and content type into account. The developing Code of Practice provides voluntary technical guidance here.[4]

These rules aim to reduce deception and maintain trust without banning useful AI tools.

The biggest operational distinction is between internal drafting and public-facing publication.

Using generative AI to produce a first draft, brainstorm headlines, or suggest structures is normal internal work. When a qualified editor then rewrites sections, verifies facts, adds original analysis, and the organisation takes public responsibility, this generally satisfies the “human review or editorial control” test for public-interest text. No disclosure is required under Article 50(4).[3]

Public-facing publication without that layer—publishing near-raw AI output as news, analysis, or authoritative commentary on public matters—triggers the disclosure duty. The same logic applies to marketing: a synthetic video that could mislead viewers about a real event or person needs clear labelling. A stylised campaign visual that does not impersonate real entities may need only technical marking.

Strong human editorial review therefore changes the analysis. Record the editor’s name, key changes made, fact-checking steps, and explicit assumption of responsibility. This evidence protects against later disputes and demonstrates responsible practice. AI literacy training (Article 4) helps editors and creators recognise when output needs heavier human intervention.[5]

Example: A publisher uses an AI tool to draft coverage of a new EU regulation on digital services. The journalist heavily revises the draft, adds interviews and context, and the editor-in-chief signs off with editorial responsibility. No disclosure label is required. By contrast, an agency publishes an AI-written “trend report” on climate policy under the brand name with minimal changes: disclosure is required.

Example: A brand team creates a synthetic voice advertisement that sounds like a well-known influencer. Even if the influencer consented, clear disclosure that the voice is AI-generated is needed unless it qualifies as artistic work.

Practical compliance requires embedding checks into existing pipelines rather than bolting on extra work.

1. Content pipeline mapping — Tag every brief or ticket with content type (marketing image, video, public-interest article, internal memo) and risk factors (deepfake risk, public-interest topic, synthetic elements).

1. Approval stages — Add mandatory gates: (a) Was AI used? Which tool? (b) Does this qualify as deepfake-like? (c) Is this public-interest text? (d) Has sufficient human review occurred and is editorial responsibility assigned? Only approved content proceeds to publication.

1. Metadata plan — For generative outputs, use available standards (such as those referenced in the Code of Practice) to embed machine-readable markers where the provider has enabled them. For platforms that do not support metadata, use consistent visible notices or alt-text.

1. Label placement — Make labels clear, distinguishable, and accessible. Place them at the start of videos, in article bylines or footnotes, in campaign metadata, or as on-screen text. The second draft Code of Practice suggests flexible but uniform approaches including a possible common EU icon.[4]

1. Archiving — Keep dated records: prompt logs (where practical and privacy-compliant), version history showing human edits, editor sign-off, responsibility statement, and screenshots of final labelled output. These form your evidence of compliance and AI literacy measures.

Link these steps to your existing editorial or creative review processes. Train teams on AI literacy so they can make informed judgements about when heavier review is needed. See the Article 50 transparency obligations explained for deeper workflow templates.

• Marketing agency creating AI visuals: The team generates product lifestyle images with generative tools. They embed available machine-readable markers and add a subtle “AI-enhanced creative” credit in campaign decks. No deepfake resemblance exists, so visible consumer labels are light. They archive tool outputs and human art-director sign-off.

• Publisher using AI to help draft public-interest coverage: A newsroom uses AI for initial research summaries on a new climate policy. Journalists rewrite with original reporting and named editors take responsibility. The final article carries the newsroom byline with no AI disclosure, supported by internal review records.

• Brand team using synthetic voice or video: A campaign features a synthetic spokesperson resembling a generic narrator (not a real celebrity). They add an audible and on-screen notice at the start (“Voice generated by AI”) and keep records of the model used and creative approval. This satisfies the deepfake-style obligation.

These errors create unnecessary regulatory risk or client friction:

• No content labels or metadata where technically required, leaving outputs undetectable.
• No editorial responsibility record for public-interest text, so the human-review exemption cannot be demonstrated.
• Unclear approval chain—content moves to publication without documented review of AI use or deepfake risk.
• Mixing “AI-assisted” and “AI-generated” categories without a written policy or consistent criteria, leading to inconsistent labelling and weak evidence.
• Assuming all internal drafting needs the same treatment as published client work, wasting effort on unnecessary labels.
• Failing to maintain archived evidence (screenshots, logs, sign-offs) that proves decisions were taken responsibly.

Does every AI-generated ad need a label? No. Only content meeting the specific criteria (deepfake resemblance or public-interest text without human review and editorial responsibility) triggers disclosure. Purely stylistic marketing images or internal drafts generally do not. Machine-readable technical marking may still apply where you act as provider.[3]

What counts as public-interest text? Text published to inform the public on matters such as health, politics, environment, regulation, elections, or significant societal issues. Operational test: would this be considered editorial or news-like content that informs public debate? Purely commercial marketing copy rarely qualifies.

Does strong human editorial review change the analysis? Yes. For public-interest text, meaningful human review plus clear editorial responsibility removes the disclosure obligation. Document the process (who reviewed, what changed, who takes responsibility) to evidence compliance. The latest Code of Practice draft supports this flexible, practice-oriented approach.[6]

How should agencies prove they handled this responsibly? Maintain consistent records: content-type tagging, approval checklists, editorial logs, responsibility statements, and archived labelled outputs. Combine with staff AI literacy training records. These artifacts show you followed a defensible process even if rules are challenged later. Tools like the Article 50 Disclosure Generator can help standardise outputs.

Action checklist

• Map all content types and flag Article 50 triggers in your briefs and approval systems.
• Update editorial guidelines to require AI-use declaration, review depth assessment, and explicit responsibility assignment.
• Implement technical marking where available and consistent visible labelling for deepfakes or non-exempt text.
• Train teams on AI literacy tailored to marketing and publishing contexts.
• Create an archiving routine for review logs, sign-offs, and final labelled versions.
• Review client contracts to set expectations on AI disclosures and shared responsibility.
• Test your workflow with a few live campaigns or articles before the August 2026 deadline.
• Monitor official updates via the AI Act Service Desk.

Ready to operationalise this? Use the Article 50 Disclosure Generator to create consistent labels and disclaimers, or download the Sample AI literacy plan for a marketing agency to see what a practical team readiness report looks like. Combine with the Article 50 transparency obligations explained and Article 4 AI literacy: what you actually need to do for deeper templates. Start building evidence-based workflows today.