EU AI Act Article 50: AI Content Labelling Rules for Marketers

Last updated: 9 November 2026

Who Article 50 applies to

The AI Act separates two roles:

Most marketers are deployers. Personal, non-professional use is outside the Act, but use that is part of a business or freelance activity counts as deployment. The rules also reach businesses outside the EU when their AI systems or outputs are used in the EU, which matters for Indian, US and other companies marketing to European audiences.

The four Article 50 duties

There are also duties for emotion recognition and biometric categorisation systems, which are less common in marketing.

Chatbots and AI agents

If you run an AI chatbot, voice agent or avatar that talks to customers, people must be able to tell it is AI, at the start of the first interaction, unless it is obvious from context. In practice:

• Add a clear line at the start: “You’re chatting with an AI assistant.”
• Do not give the bot a human name and photo that suggest a person.
• For voice agents, say it is an AI at the start of the call.
• Offer a route to a human for complex or sensitive issues.

See the best AI website chatbots and how to add an AI chatbot to your website.

Deepfakes in marketing

The Act defines a deepfake as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to be authentic or truthful. The European Commission’s guidance describes three combined tests: high resemblance, a subject that exists or could plausibly exist, and the capacity to falsely appear real.

Marketing examples that are likely to count:

• A realistic AI video of a real person, such as a founder or celebrity, saying something they did not say.
• A photorealistic AI image presented as a real customer, location or event.
• An AI voice clone of a real person in an ad.
• A realistic “news style” AI video about real events.

Examples that are less likely to count, depending on context:

• Obviously stylised illustrations or cartoons.
• Abstract or clearly fictional scenes.
• Product renders that do not claim to be photographs of real events.

When disclosure is needed, it must be clear and distinguishable, at the latest when people first see or hear the content. A machine-readable watermark added by the AI provider does not satisfy this on its own. For evidently artistic, creative, satirical or fictional work, disclosure only needs to be appropriate and must not spoil the work.

For AI video and avatar tools, see AI avatar generators and AI UGC video generators.

AI-generated text and the human review exemption

Deployers must label AI-generated text that is published, informs the public and concerns a matter of public interest. Examples of public interest topics include politics, public health, public security, environmental protection, consumer safety and relevant economic, scientific or cultural developments.

The key exemption: text that has undergone human review or editorial control does not need a label, when a person or organisation holds editorial responsibility for publishing it. The Commission’s guidance says:

• Human review means substantive examination by people with relevant expertise.
• Editorial control means a responsible entity can approve, change or reject the text on substantive grounds, including fact-checking.
• Spell-checking or grammar correction alone does not count.
• A person must hold ultimate legal responsibility for publication.

This is one more reason for real editorial review of AI-assisted content, which also helps quality and AI visibility. See how to write a blog post with AI and our AI usage policy template.

Watermarks and machine-readable marking

Providers of generative AI must mark outputs so they can be detected as AI-generated. That is why major AI companies have added watermarks and provenance data in 2026:

• Anthropic has added an invisible watermark to Claude text and C2PA provenance data to images and files, with a free checker to verify them.
• OpenAI added watermarking in October 2026. See ChatGPT watermark.
• Google uses SynthID and has expanded its SynthID Detector globally. See SynthID Detector guide.

Some outputs are not covered, such as short strings, source code and standard editing assistance. For marketers, the important point is that these marks are for detection. They do not replace your own visible disclosure for deepfakes or public interest text.

Dates and enforcement

The Commission has also approved a voluntary Code of Practice on transparency of AI-generated content, which signatories can use to show compliance with some duties.

A compliance checklist for marketing teams

1. Map your AI use. List every place you use AI: chatbots, images, video, voice, copy, personalisation.
2. Add chatbot disclosure at the start of every AI conversation.
3. Create a deepfake rule. Any realistic AI image, audio or video of real-seeming people, places or events gets a clear visible label.
4. Set editorial review standards for AI-assisted text, especially health, finance, safety and news-style topics. Name who approves.
5. Keep records of who reviewed and approved content.
6. Do not strip provenance data from AI files.
7. Brief agencies and freelancers so their work follows the same rules.
8. Update your AI policy. See our AI usage policy template.

What it means for businesses outside the EU

If you are an Indian, US or other non-EU business marketing to people in the EU, these rules can apply to you. Many global brands are adopting one standard everywhere rather than separate rules per market. Clear AI disclosure also tends to build trust with customers, wherever they are.

For related questions on ownership, see can you copyright AI content, and on data, AI data privacy for business.

Does labelling AI content hurt SEO or AI visibility?

There is no evidence that a disclosure label itself harms rankings. Google’s guidance focuses on whether content is helpful and original, not whether AI was involved. What hurts is low-quality, unreviewed content, which is exactly what the human review exemption discourages. Reviewed, expert-checked content tends to be better for readers, for compliance and for being cited by AI engines. See Google’s AI optimization guide.

Sample disclosure wording

Keep disclosures short, clear and close to the content. Adapt these to your brand and legal advice:

Roles and responsibilities in a marketing team

Common misunderstandings

• “All AI content must be labelled.” No. The text duty is narrower, and reviewed text on most topics is not covered.
• “The watermark covers us.” Not for deepfakes, which need visible disclosure.
• “We are not in the EU, so it does not apply.” It can apply when your AI outputs reach people in the EU.
• “A spell-check counts as human review.” No. Review must be substantive.

How Digital Tatva can help

Digital Tatva builds content processes that are useful, accurate and easy to defend. As part of our answer engine optimization work, we set up editorial review for AI-assisted content, make sure brand facts are consistent and help your content earn trust with readers and AI engines. We are not a law firm; for legal questions, consult a qualified adviser.

Frequently asked questions