AI Labeling Requirements: When They Apply & When Not

No, not every piece of AI-generated content requires labelling. The AI labelling obligation – the legal requirement under Art. 50 of the EU AI Act to disclose synthetic content as machine-generated where it carries a potential to deceive – takes effect on 2 August 2026 and applies exclusively to four defined categories. The widespread assumption that every AI-assisted blog post, product description and social media visual will need a label going forward is wrong. Art. 50 differentiates by deception potential, public-facing nature and degree of human control. This article identifies the categories actually affected, the concrete exemptions, and what the distinction means for marketing teams in euros and hours.

Why the fear of blanket labelling is unfounded

Since the EU AI Act was adopted, a narrative has circulated in trade media and at conferences that goes roughly like this: from August 2026, everything an AI has touched must carry a label. Anyone who believes this is either planning a compliance apparatus that devours half the content budget, or ignoring the obligation entirely because it seems unachievable. Both are wrong. The AI Act draws a precise distinction between content with deception potential and content where a human retains editorial control.

Wer KI im Unternehmen einsetzt, verlässt sich schnell auf Tools, die niemand freigegeben hat – und genau da entsteht das eigentliche Risiko. Shadow AI ist kein Randthema, sondern eine offene Flanke bei Datenschutz und Haftung. Wie sich KI-Nutzung mit Audits, verbindlichen Richtlinien und sicheren Integrationen auf einen DSGVO-konformen Boden für KI-Governance und Compliance stellen lässt, wird dort systematisch aufgeschlüsselt. Die Frage ist nicht, ob KI im Einsatz ist, sondern ob jemand die Kontrolle darüber behalten hat.

For marketing decision-makers in globally operating companies, the distinction between content that requires labelling and content that does not is budget-relevant. Knowing the exemptions avoids unnecessary process overhead and concentrates compliance resources on the genuinely critical cases – instead of routing 120 content pieces per month through the same review when only 15 of them are actually affected.

What the AI labelling obligation means – definition and legal framework

The AI labelling obligation is a transparency requirement under Art. 50 of the EU AI Act (Regulation (EU) 2024/1689). It obliges providers and deployers of AI systems to disclose synthetic content as machine-generated – but only where a potential to deceive the audience exists. The obligation takes effect on 2 August 2026. The penalty framework for violations reaches up to €15 million or 3% of global annual turnover, whichever is higher. In addition, the European Commission published the final version of the Code of Practice on Transparency of AI-Generated Content in June 2026 – voluntary, but usable as compliance evidence vis-à-vis supervisory authorities.

Provider vs. deployer – who is obligated when?

Art. 50 distinguishes two addressees with different obligations. Providers develop an AI system or place it on the market under their own name – they bear the obligations under paragraphs 1 and 2 (interaction systems, technical labelling of synthetic content). Deployers use an AI system under their own responsibility, for example a marketing team using ChatGPT or Midjourney – they bear the obligations under paragraphs 3 and 4 (deepfakes, texts of public interest). Most marketing departments are deployers, not providers. This reduces the catalogue of obligations considerably.

The four categories that actually fall under the labelling obligation

Art. 50 names four categories where transparency is mandatory. Anything that does not fall into one of these categories is not covered – and that is precisely the point most summaries omit. The four categories can be cleanly separated, and for each it is clearly regulated whether the provider or the deployer bears the obligation.

Interactive AI systems – chatbots and voicebots

Users must be able to recognise that they are communicating with a machine, not a human. The obligation falls on the provider of the system. A customer chatbot on the company website needs a notice – such as "You are communicating with an AI system". An internal research tool used only by employees also falls under this rule, provided the interaction is not obviously machine-driven.

Synthetically generated content – image, audio, video, text

This obligation applies to providers that make AI systems available for generating synthetic content. They must ensure that outputs are machine-readably identifiable as AI-generated – through watermarks, metadata or other technical methods. Anyone who merely uses such a system (deployer) is not affected by this specific obligation. The technical labelling happens in the background, not as a visible label for the end user.

Deepfakes – manipulated depictions of real persons or events

A deepfake within the meaning of the AI Act is AI-generated or -manipulated image, audio or video content that resembles real persons, places or events and could falsely appear authentic. Here the obligation falls on the deployer: whoever publishes a deepfake must disclose that the content was artificially generated or manipulated. In marketing, this applies for example to an AI-generated video in which a real person makes statements they never actually made.

Texts on matters of public interest

AI-generated or -manipulated texts that inform on political, societal, economic or cultural topics and are publicly accessible must be labelled as AI-generated. Here too, the deployer bears the obligation. An AI-generated expert article on the economic outlook on the company blog can fall under this category – a product description in the online shop does not.

AI labelling exemptions – when the obligation does not apply

The AI Act defines clear exemptions. Anyone who uses AI as a tool and retains editorial control is, in most cases, not subject to the labelling obligation. The exemptions are not loopholes but deliberate boundaries drawn by the legislator – they protect freedom of expression, artistic freedom and proportionality.

• Assistive use: AI corrects spelling, suggests phrasing, translates or summarises – the human substantially shapes the content. No labelling required.
• Non-public content: Internal documents, team communications, drafts without publication. No labelling required.
• Editorial responsibility: An employee reviews the AI text for substance and assumes responsibility – the obligation lapses for texts of public interest when a natural or legal person bears editorial responsibility.
• Art/satire exemption for deepfakes: Obviously artistic, creative, satirical or fictional content need only be labelled in a way that does not impair enjoyment of the work.
• Obviousness: Where it is immediately apparent to the average user that content is AI-generated – for example in comic style or as a recognisable illustration – the obligation does not apply.

Good to know: The decisive control question is: "Can the average user be deceived by the AI output?" If the answer is no, there is no labelling obligation.

Wer KI im Unternehmen einsetzt, verlässt sich schnell auf Tools, die niemand freigegeben hat – und genau da entsteht das eigentliche Risiko. Shadow AI ist kein Randthema, sondern eine offene Flanke bei Datenschutz und Haftung. Wie sich KI-Nutzung mit Audits, verbindlichen Richtlinien und sicheren Integrationen auf einen DSGVO-konformen Boden für KI-Governance und Compliance stellen lässt, wird dort systematisch aufgeschlüsselt. Die Frage ist nicht, ob KI im Einsatz ist, sondern ob jemand die Kontrolle darüber behalten hat.

Worked example – compliance effort under blanket vs. differentiated labelling

Blanket labelling of all AI-assisted content creates process overhead that is disproportionate to the legal risk. A worked example makes the difference tangible – and shows why differentiation is not legal hair-splitting but a commercial decision.

The assumption: 30 minutes of review per content piece – assessing AI involvement, documentation, label placement, sign-off. At an internal hourly rate of €85, the monthly saving comes to approximately €4,080. Annualised, that is close to €49,000 – budget that could flow into production rather than administration. The figures are conservative estimates. Anyone producing more than 120 content pieces per month multiplies accordingly.

AI content and GDPR, copyright, liability – demarcation from the labelling obligation

The labelling obligation under Art. 50 of the AI Act is not the only legal requirement for AI content. GDPR, copyright and liability questions exist in parallel – but concern different aspects than the transparency obligation. Anyone who conflates the three legal domains either builds the wrong process or operates under a false sense of security.

• GDPR: Relevant when personal data flows into AI systems or AI outputs contain personal data. No direct connection to the labelling obligation, but overlap in the case of deepfakes of real persons – where personality rights and data protection apply additionally.
• Copyright: AI-generated content does not enjoy independent copyright protection under current law, as no human creator is involved. The liability risk lies in the potential adoption of protected elements from training data – a problem that labelling does not solve.
• Liability: The deployer is liable for published content regardless of whether AI was involved. Labelling does not protect against substantive liability for false statements, personality rights violations or anti-competitive claims.

AI content in journalism and content marketing – where the line runs

The category "texts on matters of public interest" affects not only traditional newsrooms but also corporate blogs, thought-leadership formats and specialist publications on company websites. The demarcation is less clear-cut than it appears at first glance – but it can be operationalised with three criteria.

A product text in the online shop is not a text of public interest. An expert article on the economic outlook on the company blog can be. What matters is the combination of societal, political, economic or cultural relevance and public accessibility. Anyone who documents editorial responsibility – i.e. demonstrably reviews, edits and signs off – is exempt even when AI delivered the first draft. Documentation is the key: it is not the tool that decides, but who takes responsibility for the content.

A documented content strategy makes priorities and compliance requirements plannable. Those who do not want to build this internally can develop it with a specialist content marketing agency such as Crispy Content®.

Trends – how AI labelling will evolve through 2027

Regulation is not static. Between the entry into force in August 2026 and the Commission's first evaluation, at least four developments will affect practice – anyone building their process now should factor in this direction of travel.

• Technical standards: The EU Code of Practice (June 2026) establishes watermark and metadata standards based on C2PA. Providers such as OpenAI, Google and Adobe are already implementing these in their systems. For deployers this means: technical labelling will increasingly be delivered automatically.
• Extension to additional media formats: The Commission is examining whether interactive formats such as AR and VR content should receive additional transparency obligations. Anyone planning immersive campaigns should monitor this development.
• National enforcement: Germany has designated the Bundesnetzagentur as its AI supervisory authority. The concrete enforcement practice – depth of scrutiny, priorities, willingness to sanction – will only become apparent from autumn 2026 onwards.
• Convergence with platform regulation: The Digital Services Act (DSA) already obliges platforms to provide transparency on algorithmically curated content. Overlaps with Art. 50 of the AI Act will increase – particularly regarding the question of who on platforms is responsible for labelling: the creator or the distributor.

Implementing the labelling obligation operationally – process, not panic

Implementation does not require over-regulation but a clear internal process. Four elements are sufficient to both fulfil the legal obligation and limit effort to what is necessary.

• AI registry: Central documentation of all AI tools in use, their use cases and the responsible persons for each. Without this registry, no decision matrix can be applied – you simply do not know where AI is in play.
• Decision matrix: For each content type, answer four questions: Is the content public? Is it of public interest? Is the AI involvement substantial? Can the user be deceived? The obligation only applies when all four questions are answered with yes.
• Standard wording: Pre-formulated labels for text, image, video and audio – such as "This content was created with the assistance of AI" or the official EU icons the Commission made available in 2026.
• Review process: Document editorial sign-off before publication for potentially labelling-obligated content. The documentation itself is the compliance evidence.

Differentiation over blanket rules protects brand trust and budget

The AI labelling obligation from August 2026 applies to defined categories – not to a company's entire AI-assisted content output. Marketing teams that know the exemptions and establish a documented assessment process avoid both fines and unnecessary effort. The control question remains: can the average user be deceived? Where the answer is no, there is no obligation – and voluntary labelling is a strategic decision, not a legal one. Methods provide guarantees. Anyone who sets up the process properly once does not need to decide anew with every blog post but works with a system that already contains the answer.

Sources

HÄRTING Rechtsanwälte (2025): Transparenzpflichten in der KI-Verordnung (Art. 50): KI-Inhalte richtig kennzeichnen. URL: https://haerting.de/wissen/transparenzpflichten-in-der-ki-verordnung/ (accessed 20.07.2026).

Gesellschaft für Datenschutz (2025): KI Kennzeichnungspflicht nach Art. 50 KI-VO. URL: https://gesellschaft-datenschutz.de/ki-kennzeichnungspflicht/ (accessed 20.07.2026).

ECOVIS Unternehmensberatung (2026): EU AI Act: Kennzeichnungspflicht für KI-generierte Inhalte. URL: https://de.ecovis.com/unternehmensberatung/eu-ai-act-kennzeichnungspflicht-unternehmen/ (accessed 20.07.2026).

European Commission (2026): EU icons for labelling AI-generated content. URL: https://digital-strategy.ec.europa.eu/de/policies/eu-icons-labelling-ai-generated-content (accessed 20.07.2026).

European Commission (2026): Code of Practice on Transparency of AI-Generated Content. URL: https://digital-strategy.ec.europa.eu/de/policies/code-practice-ai-generated-content (accessed 20.07.2026).

Wirtschaftskammer Österreich (2026): Kennzeichnungspflicht für bestimmte KI-Inhalte ab August 2026. URL: https://www.wko.at/gewerbe-handwerk/kennzeichnungspflicht-fuer-ki-inhalte (accessed 20.07.2026).

ai-act-law.eu (2025): Art. 50 KI-VO – Transparenzpflichten für Anbieter und Betreiber bestimmter KI-Systeme. URL: https://ai-act-law.eu/de/artikel/50/ (accessed 20.07.2026).

AndersAI (2026): Urheberrecht bei KI-generierten Inhalten. URL: https://www.andersai.de/blog/urheberrecht-ki-generierte-inhalte-unternehmen (accessed 20.07.2026).