mail@crispycontent.de
Last updated on September 8, 2026 at 12:08 PM.
The labelling requirement for AI-generated text is a transparency obligation under Article 50 of the EU AI Act (Regulation (EU) 2024/1689), applicable from 2 August 2026. Not every AI-generated text requires a label – the obligation only applies when two conditions are met simultaneously: the content informs the public about a matter of public interest, and no human has assumed editorial responsibility. This article uses a concrete yes/no logic to clarify which texts are affected, which exemptions the law provides, and what fines apply for non-compliance.
What does the labelling requirement for AI text mean under the EU AI Act?
The labelling requirement is the legal obligation to disclose to the audience that content has been generated or substantially modified by an AI system. The legal basis is Article 50(4) of the AI Act. It addresses deployers – i.e. organisations that use AI systems under their own responsibility, for example to produce text with large language models.
Anyone deploying AI in their organisation without governing its use is making a promise they cannot keep – namely that everything is in order. Shadow AI is not a footnote; it is a real risk: tools that no one has approved are processing data that no one is accountable for. How to put AI usage on a GDPR-compliant footing through audits, clear policies and secure integrations is explained here.
The central legal term is deep synthetic content – defined in Article 3(60) as AI-generated content (text, image, audio, video) that resembles real content and would not be recognisable as synthetic without labelling. The regulation clearly distinguishes between the obligation of the provider (technical labelling through watermarks and metadata) and the obligation of the deployer (disclosure to the audience).
| Term | Definition | Legal source |
|---|---|---|
| Labelling requirement | Disclosure that content is AI-generated or AI-manipulated | Art. 50(4) AI Act |
| Deep synthetic content | AI-generated content that resembles real content | Art. 3(60) AI Act |
| Deployer | Organisation that uses an AI system under its own responsibility | Art. 3(4) AI Act |
Do all texts created with ChatGPT need to be labelled?
No. The labelling obligation only applies when two conditions are met simultaneously: first, the text is published to inform the public about a matter of public interest; second, no human has assumed editorial responsibility for the content. If either condition is absent, there is no statutory labelling requirement under Article 50.
When does "public interest" apply?
Public interest within the meaning of the AI Act covers topics from politics, society, business and culture – anything that goes beyond individual commercial relationships and addresses a broad audience. The channel is secondary: websites, social media profiles, blogs and news portals all qualify equally, provided they are publicly accessible. Not covered are internal emails, team chats, closed distribution lists or intranet pages. A product description in an online shop informs about an offering, not about a matter of public interest – it therefore does not fall under the obligation.
What does "human review" mean in the legal sense?
Human review means that a natural or legal person demonstrably assumes editorial responsibility for the published content. Simply reading it through is not sufficient. What is required is a substantive review prior to publication – i.e. checking for accuracy, completeness and appropriateness. The European Commission recommends in its draft guidelines that this review be documented: through review logs, approval workflows or comparable evidence. Anyone who publishes an AI-generated text on a political topic and assumes responsibility with a documented sign-off is exempt from the labelling requirement.
| Question | Yes → | No → |
|---|---|---|
| Does the text inform about a matter of public interest? | Continue checking | No labelling requirement |
| Has a human assumed editorial responsibility? | No labelling requirement | Labelling requirement |
| Is the text used exclusively internally? | No labelling requirement | – |
Which AI content additionally falls under the transparency obligation?
Article 50 is not limited to text. The transparency obligation extends to four categories: chatbots (paragraph 1 – users must know they are interacting with a machine), synthetic image, audio and video content (paragraph 2 – providers must embed machine-readable labels), emotion recognition systems and biometric categorisation (paragraph 3), and deepfakes and texts of public interest (paragraph 4 – deployers must disclose to the audience). The obligations are distributed across two roles: providers are responsible for the technical infrastructure, deployers for disclosure to the end audience.
Deepfakes vs. obviously fictional content
A deepfake is AI-generated content that resembles real persons, places or events and is capable of deceiving – for example, a synthetic video showing a politician making a statement they never made. Obviously fictional content such as comics, fantasy illustrations or satirical animations does not fall under the deepfake definition and does not trigger a labelling requirement. The grey area lies with photorealistic AI-generated images in e-commerce – for instance, synthetic models that are virtually indistinguishable from real people. The Commission's draft guidelines recommend: when in doubt, label.
Cost calculation: What does a breach of the labelling requirement cost?
The fine framework for breaches of Article 50 is, according to Article 99 of the AI Act, up to €15 million or 3% of global annual turnover – whichever is higher. For a company with €500 million in annual turnover, the turnover calculation yields €15 million; here the fixed cap applies because both values are identical. For a company with €2 billion in turnover, 3% already amounts to €60 million – in that case the turnover threshold applies. The amounts are comparable in magnitude to GDPR fines but lower than the sanctions for prohibited AI practices under Article 5.
| Breach category | Maximum fine | GDPR comparison |
|---|---|---|
| Transparency obligation (Art. 50) | €15m or 3% of turnover | €20m or 4% of turnover |
| Prohibited AI practices (Art. 5) | €35m or 7% of turnover | – |
| False statements to authorities | €7.5m or 1% of turnover | – |
Which exemptions apply to the labelling of AI text?
The law provides five exemption grounds that eliminate the labelling requirement. The most important one for marketing teams: anyone who reviews AI-generated content prior to publication and demonstrably assumes editorial responsibility is not subject to the labelling requirement – regardless of the topic.
- Human editorial control: A human reviews the content and demonstrably assumes responsibility. The obligation is fully waived.
- No public interest: Product descriptions, technical manuals, internal documents – these do not fall under Article 50(4).
- Assistive function without substantial content modification: If the AI exclusively corrects spelling, formats or makes stylistic adjustments, no generation or substantial manipulation has occurred.
- Law enforcement and national security: AI systems legally authorised for crime detection are subject to separate rules.
- Art and satire: For obviously artistic, satirical or fictional works, the label may be designed so as not to impair enjoyment – for example in the credits rather than in the image.
How exactly must AI-generated texts be labelled?
The label must be visible no later than the moment of first perception of the content – not hidden in the footer, not in 8pt type, not behind a click. The European Commission requires in its draft guidelines: clear, unambiguous, accessible. This means sufficient contrast, screen-reader compatibility and a placement the user cannot miss.
A practical wording example: "This text was created with the assistance of AI." On the technical level, providers are required to embed machine-readable metadata – for instance following the C2PA standard (Coalition for Content Provenance and Authenticity), which writes provenance information directly into the file. For deployers, the visible label is sufficient; the technical layer is the responsibility of the AI system provider.
A documented content strategy defines which content falls under the labelling requirement and what the approval process looks like. Organisations that prefer not to build this process internally can develop it with a specialist content marketing agency such as Crispy Content®.
AI labelling and brand perception: What the research shows
The fear that an AI disclosure automatically destroys trust is not supported by evidence. A university study published in 2025 shows: the effect of an AI label on brand perception depends on context – not on the label itself. Brands with high credibility and clear positioning do not lose from transparency; they gain. Brands without substance lose – but not because of the label; rather because the label makes the lack of value visible.
In parallel, platforms such as Meta, Google and TikTok are implementing their own AI labels that apply independently of the statutory obligation. In September 2025, the EU Commission published a voluntary code of practice on transparency of AI-generated content – as a bridge until the binding deadline in August 2026. The trend is clear: machine-readable labelling through watermarks and metadata is becoming the technical standard. Manual labels alone – a sentence below the article – will not suffice in the long run because they are lost during automated redistribution.
For B2B marketing, this means: organisations that establish transparent processes now create a trust signal for stakeholders while simultaneously reducing their compliance risk. Methods provide guarantees – here, too.
Fact check: The five most common misconceptions about AI labelling
The debate around AI labelling reliably produces misunderstandings. Most arise because the legal text is not read but retold – and with each retelling, a layer of nuance is lost.
| Misconception | Fact |
|---|---|
| "Every AI text must be labelled" | Only texts on matters of public interest without human editorial review |
| "The obligation already applies now" | Art. 50 becomes applicable on 2 August 2026 |
| "AI spell-checking requires labelling" | Assistive functions without substantial content modification are exempt |
| "Internal documents must be labelled" | Only publicly accessible content is affected |
| "Labelling damages the brand" | Studies show: transparency strengthens trust when the context is right |
| Use case | Labelling required? | Rationale |
|---|---|---|
| Blog post on energy policy, fully AI-generated, no editorial review | Yes | Public interest + no human control |
| Product description in an online shop, AI-generated | No | Not a matter of public interest |
| Social media post on the federal election, AI-generated, reviewed and approved by an editor | No | Human editorial responsibility documented |
| Internal briefing for the marketing team, AI-assisted | No | Not publicly accessible |
| AI-generated news article without author attribution | Yes | Public interest + no editorial responsibility identifiable |
| Date | Obligation | Relevance |
|---|---|---|
| 2 February 2025 | AI literacy requirement (Art. 4) | Employees must develop AI competence |
| 2 August 2025 | Prohibition of certain AI practices (Art. 5) | Manipulative AI systems banned |
| 2 August 2026 | Transparency obligations (Art. 50) | Labelling requirement for AI content |
What marketing teams should implement by August 2026
Thirteen days. That is how much time remains until the deadline. The good news: the requirements are clear, and implementation is not rocket science – it demands structure, not budget.
Anyone deploying AI in their organisation without governing its use is making a promise they cannot keep – namely that everything is in order. Shadow AI is not a footnote; it is a real risk: tools that no one has approved are processing data that no one is accountable for. How to put AI usage on a GDPR-compliant footing through audits, clear policies and secure integrations is explained here.
- Create an AI registry: Which AI tools are used for which content? Without this overview, any compliance statement is speculation.
- Document the editorial process: Who reviews AI-generated content before publication? Log the sign-off – with date, name and a substantive review note.
- Define standard wording: Establish uniform labelling texts for web, social media and newsletters so that individual editors do not have to improvise.
- Assign roles: Appoint an AI lead within the content team who maintains oversight of obligations and exemptions.
- Document exemptions: For every unlabelled piece of content, record why no obligation applies – this is the safeguard in the event of an audit.
Position first, then publish.
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).
- Future of Life Institute (2024): Article 50 – Transparency obligations for providers and deployers of certain AI systems. Regulation (EU) 2024/1689. URL: https://artificialintelligenceact.eu/de/article/50/ (accessed 20.07.2026).
- ECOVIS (2025): EU AI Act: Labelling requirement for AI-generated content. URL: https://de.ecovis.com/unternehmensberatung/eu-ai-act-kennzeichnungspflicht-unternehmen/ (accessed 20.07.2026).
- Gesellschaft für Datenschutz (2025): AI labelling requirement under Art. 50 AI Act. URL: https://gesellschaft-datenschutz.de/ki-kennzeichnungspflicht/ (accessed 20.07.2026).
- European Commission (2025): Draft guidelines on the implementation of transparency obligations for certain AI systems under Article 50 of the AI Act. URL: https://digital-strategy.ec.europa.eu/de/library/draft-guidelines-implementation-transparency-obligations-certain-ai-systems-under-article-50-ai-act (accessed 20.07.2026).
- idw – Informationsdienst Wissenschaft (2025): AI labelling as brand management (university study). URL: https://idw-online.de/de/news874567 (accessed 20.07.2026).
- European Commission (2025): 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).
- IHK Köln (2025): AI Act: New transparency obligations from August 2026. URL: https://www.ihk.de/koeln/hauptnavigation/digitalisierung-und-innovation/digitalisierung/transparenzpflichten-nach-der-ki-verordnung-7100068 (accessed 20.07.2026).
Gerrit Grunert
Gerrit Grunert is the founder and CEO of Crispy Content®. In 2019, he published his book "Methodical Content Marketing" published by Springer Gabler, as well as the series of online courses "Making Content." In his free time, Gerrit is a passionate guitar collector, likes reading books by Stefan Zweig, and listening to music from the day before yesterday.
.png)
.jpg)
-1.jpg)
-1.jpg)
.jpg)
.jpg)
.jpg)
