AI Labeling Requirements 2026: What Applies and What Doesn't

The AI labelling obligation is a transparency requirement under Article 50 of the EU AI Act (Regulation (EU) 2024/1689) that compels providers and deployers of AI systems to disclose certain AI-generated content to users. The deadline is 2 August 2026. Not every piece of AI content requires a label — what matters is the potential to deceive, the type of content and the context of publication. Drafting a product description with ChatGPT does not trigger the obligation. Deploying a photorealistic AI avatar as a brand ambassador does. This article translates the regulation into concrete decision criteria for marketing teams — with case examples, thresholds and a checklist for operational implementation.

What Article 50 of the AI Act means for marketing teams

Article 50 defines four situations in which transparency towards users becomes mandatory — regardless of whether the AI system is classified as high-risk or not. The four obligations concern: chatbot interactions where humans communicate with AI; synthetic audio, image and video content (deepfakes); emotion recognition systems; and AI-generated text that informs the public about matters of public interest. For marketing departments this means: every company that uses AI tools — from sole traders to corporations — qualifies as a deployer under the regulation and carries its own disclosure obligations.

Most companies discover their AI exposure only after an employee has already pasted client data into a public chatbot. Shadow AI is not a footnote in the risk register — it is the register. The uncomfortable truth is that usage runs ahead of governance in almost every organization, and the gap is where the liability sits. Whoever wants to know how AI governance and compliance turn unmanaged tool usage into GDPR-compliant ground — through audits, binding policies and secure integrations will find the practical groundwork laid out there.

Provider vs. deployer — who is liable for what?

The AI Act separates two levels of responsibility. Providers — the companies that develop AI systems and place them on the market — must supply the technical infrastructure for labelling: machine-readable watermarks, metadata, interoperable markers. Deployers — every company that uses these systems — must ensure the label actually reaches the end user. The duty of visible disclosure rests with the deployer, not the tool provider.

A practical example illustrates the boundary: a company purchases a chatbot, rebrands it, customises its personality and integrates it into its own customer service. From that point on, the company itself becomes a provider under Art. 25 of the AI Act — with all the technical obligations that entails. The question "Who is liable?" is therefore not answered generically, but by the degree of customisation.

AI images in campaigns — when labelling is mandatory

Not every AI-generated image falls under the labelling obligation. Article 50(4) of the AI Act covers exclusively deepfakes — defined as AI-generated content that resembles real persons, objects or places and would falsely appear to be authentic. An abstract pattern, an obviously stylised advertising motif or a comic illustration does not trigger the obligation, even if it was generated entirely by AI. The decisive question is not "Did AI make this?" but "Could someone believe this is real?"

Three decision criteria for marketing leads

The Wettbewerbszentrale outlines three assessment steps in its guidance that marketing teams can apply to every AI image:

• Resemblance to real objects or persons: Does the image show something that could exist in the physical world — a specific person, a recognisable location, a realistic product? If so, the likelihood of a labelling requirement increases.
• Deception potential for the target audience: Would an average viewer in the specific publication context mistake the image for a photograph? The assessment is based on the recipient's perspective, not the technical method of creation.
• Context of publication: A photorealistic AI image on a news portal carries a different deception potential than the same image in a gaming community whose members expect synthetic content.

Case examples from the Wettbewerbszentrale guidance

Good to know: Virtual influencers and AI-generated models require labelling because they are deliberately designed to appear human-like and the deception potential is part of their business model.

ChatGPT texts for social media — does the labelling obligation apply?

Promotional texts — product descriptions, ad copy, social media captions — do not fall under the labelling obligation pursuant to Art. 50(4). The obligation applies exclusively to AI-generated texts that inform the public about matters of public interest. An Instagram post about a new sneaker model is not a matter of public interest. An AI-generated opinion piece on energy policy is. An additional exception applies: if a human substantively reviews the text, edits it editorially and assumes responsibility for its content, the obligation is waived — regardless of the topic.

When a social media post falls under the obligation

The line between "AI draft with human revision" and "AI text with cosmetic correction" is fluid. The AI Act requires substantive human review — meaning genuine assumption of editorial responsibility, not merely proofreading for typos. Anyone who only checks an AI-generated text on a political topic for grammar and then publishes it does not meet the criterion of editorial control.

AI avatars and synthetic voices — transparency obligations in customer interactions

Chatbots, AI avatars and synthetic voices must identify themselves as AI at the beginning of the interaction. Art. 50(1) of the AI Act is unambiguous: persons interacting with an AI system must be informed that they are communicating with an AI — unless the AI nature is obvious from the circumstances. A stylised robot avatar without human features may be obvious. A photorealistic AI avatar with a human voice is not.

For synthetic voices in telephone hotlines or podcast advertising: an audible disclaimer at the beginning of the interaction is required. The wording must be unambiguous. "You are chatting here with an AI assistant" meets the requirement. "This service is supported by AI" does not.

Why "supported by AI" is not sufficient

The Wettbewerbszentrale considers the phrase "supported by AI" to be misleading because it suggests a human agent is merely augmented by AI. If the interaction is conducted entirely by an AI system, the label must say so. The recommended wording is, in substance: "You are communicating here exclusively with an artificial intelligence." Anything less opens a grey area that is vulnerable to challenge under competition law.

Not every AI output needs a label — three key exceptions

The AI labelling obligation is not a blanket labelling requirement for everything AI has touched. Three exceptions significantly limit the scope:

• Assistive functions: Grammar correction, style suggestions, auto-completion — these functions do not produce standalone content and do not trigger a labelling obligation.
• Artistic, satirical or fictional works: A reduced obligation applies here. Labelling must be provided but may be discreet and must not impair artistic enjoyment — for example as a note in the credits or in the metadata.
• Content under editorial responsibility: If a human bears editorial responsibility and substantive review is demonstrable, the obligation under Art. 50(4) is waived.

Good to know: The EU Code of Practice on Transparency of AI-Generated Content distinguishes between "fully AI-generated" and "AI-assisted" — with different disclosure requirements. A text drafted by AI and substantively revised by an editor falls into the "AI-assisted" category and is subject to lower requirements than fully machine-generated content.

What the labelling must look like in practice

The AI Act does not prescribe an exact visual format but requires "clear and distinguishable" disclosure that is recognisable to the recipient before or at the latest upon consumption. The European Commission is developing a standardised label within the framework of the Code of Practice — "KI" for German-language markets, "AI" for English-language ones. Until final standardisation, companies orient themselves by the OLG case law on influencer disclosure: in the local language, at the beginning of the content, in legible size. Art. 50(5) additionally requires accessibility — the label must also be perceptible to people with disabilities.

AI use in marketing 2026 — figures and regulatory context

The regulation meets an industry that has long integrated AI into its core processes. 84 % of marketing decision-makers in Germany view AI as the most important factor influencing their work; 67 % believe marketing will no longer succeed without AI. In parallel, the BCG study "AI at Work 2025" shows that 67 % of employees in Germany regularly use AI tools — a figure that has risen by 15 percentage points since 2024.

The consequence is straightforward: compliance capability becomes a competitive factor. Ignoring the labelling obligation risks not only fines but cease-and-desist actions from competitors. Over-labelling — plastering every AI-assisted piece of content with warnings — creates friction in communication without legal necessity. The task is differentiation — and differentiation requires knowledge of the actual thresholds.

Companies that cannot map the translation of regulation into operational processes internally can develop the implementation with a specialised communications agency such as Crispy Content®.

Sanctions and competition-law risks for non-compliance

Violations of the transparency obligations under Art. 50 can be sanctioned on two levels. The AI Act itself provides for fines of up to €15 million or 3 % of global annual turnover — whichever is higher. In parallel, German competition law applies: a breach of the labelling obligation can be classified as misleading by omission under § 5a UWG (Unfair Competition Act). The Wettbewerbszentrale and competitors can assert injunctive claims — without involving authorities, by way of cease-and-desist letter.

An additional risk that gets lost in the debate: AI washing. Anyone who claims AI capabilities in their communications that the product does not possess violates the prohibition on misleading practices — independently of the AI Act. The labelling obligation and the prohibition of false AI claims are two sides of the same coin: transparency about what is AI and what is not.

Next steps for marketing decision-makers before August 2026

The deadline is set. Twelve days remain. Anyone who has not started does not need a strategy — they need an inventory. Five steps that can be completed by 2 August:

• Inventory AI systems: Which tools are in use — officially and unofficially? Chatbots, image generators, text tools, speech synthesis, avatars. List everything.
• Assess Art. 50 relevance: For each system, run through the four obligations. Do customers interact directly with the AI? Are deepfakes being produced? Are texts on matters of public interest published without editorial control?
• Integrate labelling processes into content workflows: Not as a downstream check but as a step in the production process. Anyone who only looks for the label at the point of publishing has already forgotten it.
• Use the EU Code of Practice as a benchmark: The Code of Practice provides a taxonomy, suggested wording and technical standards. It is voluntary, but it defines the state of the art — and thus the standard against which courts will measure.
• Impose contractual compliance requirements on AI providers: Does the provider deliver machine-readable metadata? Does the tool support C2PA watermarks? If not, the technical foundation for meeting your own obligations is missing.

Method provides assurance. Anyone who works through these five steps in a documented manner can demonstrate on 2 August that their AI usage meets the transparency requirements. Anyone who does not is hoping nobody looks. That is not a strategy — it is a promise you cannot keep.

Sources

Wettbewerbszentrale e.V. (2026): Kennzeichnung KI-generierter Inhalte – Wenn KI die Werbung macht: Was Sie als Unternehmen wissen müssen. URL: https://www.wettbewerbszentrale.de/wp-content/uploads/2026/02/2026_2_Leitfaden_KI_generierte_inhalte_1-1.pdf (accessed 20 July 2026).

European Commission, Representation in Germany (2026): Kennzeichnung von KI-generierten Inhalten: EU-Kommission veröffentlicht Verhaltenskodex. URL: https://germany.representation.ec.europa.eu/nachrichten-und-veranstaltungen/pressemitteilungen/kennzeichnung-von-ki-generierten-inhalten-eu-kommission-veroffentlicht-verhaltenskodex-2026-06-10_de (accessed 20 July 2026).

Future of Life Institute (2026): The EU AI Act's Transparency Rules: A Practical Guide to Article 50. URL: https://artificialintelligenceact.eu/transparency-rules-article-50/ (accessed 20 July 2026).

Bitkom e.V. (2026): Marketing im digitalen Wandel: Zwischen Effizienz, Automatisierung und Wettbewerb 2026. URL: https://www.bitkom.org/Bitkom/Publikationen/Marketing-im-digitalen-Wandel-2026 (accessed 20 July 2026).

Boston Consulting Group (2025): AI at Work 2025. URL: https://web-assets.bcg.com/fd/0d/bcc5dfae4cbaa08c718b95b16cf5/ai-at-work-2025-slideshow-june-2025-edit-02.pdf (accessed 20 July 2026).

European Commission (2026): Code of Practice on Transparency of AI-Generated Content. URL: https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content (accessed 20 July 2026).