Since August 2, 2026, binding transparency obligations for artificial intelligence have applied throughout the European Union. Anyone deploying chatbots, generating images, or publishing texts with AI assistance must comply with clear rules. The legal basis is Article 50 of the European AI Regulation (AI Act, Regulation (EU) 2024/1689).
Unlike what is often oversimplified, the article does not impose a blanket labelling requirement for “all AI content”. Instead, it differentiates by role – provider or deployer – and by use case: whether and how a disclosure must be made depends on whether a system interacts with humans, generates synthetic content, or involves deepfakes or texts of public interest.
This article explains what the labelling obligation concretely means, who it applies to, where the exemptions lie, and what consequences breaches entail.
What is the AI Labelling Obligation?
The AI labelling obligation is part of the European AI Regulation (EU AI Act) and is governed by Article 50. It requires providers and deployers of certain AI systems to disclose when content has been created or altered by artificial intelligence, or when users interact with an AI system instead of a human. In practice, this means: chatbots must be recognizable as such, synthetically generated audio, image, video, or text content must be marked in a machine-readable technical format, and in particular deepfakes as well as texts that inform the public on matters of public interest must be clearly labelled as AI-generated. This is not a blanket labelling requirement for every text or image created with AI, but rather graduated transparency measures depending on the situation and risk.
Why does the AI Labelling Obligation exist?
The regulatory background reflects a simple goal: people should be able to recognise when they are interacting with artificial intelligence or when content they encounter has been artificially generated or altered. The legislator is responding to a development that has become increasingly apparent in recent years: AI-generated texts, images, videos, and audio content are now barely distinguishable from human-made content.
This poses several risks that Article 50 is designed to specifically counter:
- Prevent deception: Users should not unknowingly believe they are speaking to a human when actually interacting with a chatbot.
- Curb disinformation: Deepfakes and AI-generated content can be deliberately used to fabricate persons, events, or statements, for example in political debates.
- Preserve trust in public communication: For content of public interest, it should remain transparent whether and to what extent AI was involved in its creation.
- Enable informed decisions: When people know that content is AI-generated, they can categorise it accordingly and assess it more critically.
Who is affected by the AI Labelling Obligation?
The AI Act fundamentally distinguishes between two roles, each with different obligations:
- Providers (Art. 3 No. 3): Companies that develop an AI system or have it developed and place it on the market under their own name. These include OpenAI, Google, Adobe, Midjourney, but also any company that offers its own AI product or resells a model under its own name.
- Deployers (Art. 3 No. 4): Companies that use an existing AI system under their own responsibility, for example to create content or in internal workflows. This affects the vast majority of companies that use third-party tools such as ChatGPT, Midjourney, or similar services.
For most companies using AI tools in everyday operations, the deployer role is primarily relevant. Purely private, non-professional use is explicitly excluded from the regulation.
Which AI content falls under the Labelling Obligation?
According to Art. 50 of the AI Regulation, the AI labelling obligation applies to:
- Deepfakes: AI-generated or manipulated image, audio, and video content that mimics persons, objects, places, or events so deceptively that the impression is created that they could exist in reality and therefore appear to a person to be truthful.
- Texts of Public Interest: AI-generated texts of public interest (e.g., newspaper articles, parts of scientific work, press releases) that are published without human review or editorial processing.
The 3 Transparency Obligations for Deployers
Article 50 of the EU AI Regulation distinguishes four transparency scenarios, two of which directly address deployers. For typical business operations, these are primarily the labelling of deepfakes and certain texts (both paragraph 4). Additionally, but rarely relevant, comes the disclosure obligation for biometric and emotion recognition (paragraph 3).
1. Deepfakes (Image, Video, Audio) – Art. 50 Para. 4
What is required: Deepfakes must be disclosed as AI-generated or AI-manipulated if they deceptively resemble real persons, objects, places, or events and the public could falsely believe them to be authentic. The labelling must be clear and recognizable to those affected in a timely manner.
Examples: Face-swaps or voice clones, photorealistic AI models or virtual influencers, AI-generated “customer voices” or testimonials, realistic-looking product photos with AI persons, virtually furnished apartments, manipulated recordings of real events.
Exception: Mere technical optimization such as cropping, exposure or colour correction, noise reduction, scaling, or compression, as long as the core message of the image is not altered. For obviously artistic, creative, satirical, fictional, or similar works, a subtle notice of AI use is sufficient, as long as it does not impair enjoyment of the work.
2. AI-Generated Texts on Matters of Public Interest (Art. 50 Para. 4)
What is required: Anyone who publishes AI-generated or AI-manipulated texts intended to inform the public on matters of public interest must disclose this, unless the text has undergone genuine human editorial review and a person or organisation assumes editorial responsibility for it. The labelling must be clearly recognizable at the latest at the time of publication and must not disappear into terms and conditions, imprint pages, or hidden menus.
Examples: Automatically published news articles without editorial review, AI-generated news summaries, unedited posts on politics, health, or economics.
Exception: If the text is reviewed for content and editorially vetted before publication, the labelling obligation is waived. Purely promotional or company-related texts such as product descriptions, landing pages, or classic marketing blog posts generally do not fall under this narrow obligation anyway, as long as they do not present themselves as independent reporting.
3. Emotion Recognition Systems and Biometric Categorisation Systems (Art. 50 Para. 3)
What is required: Anyone who deploys a system that recognizes emotions or categorizes persons based on biometric data (e.g., suspected origin, gender, or state) must inform the persons affected that they are exposed to such a system.
Examples: Video analysis in retail, AI-assisted recruitment procedures, emotion analysis in call centres.
Exception: Systems that are legally permitted to detect, prevent, or investigate criminal offences are not covered.
What does NOT require labelling
In the following cases, there is generally no labelling obligation:
- AI-generated images that are recognisably artistic, stylised, or cartoon-like, as well as icons, infographics, and abstract graphics.
- AI texts that have been editorially reviewed and are the responsibility of a human (e.g., AI draft plus editorial revision).
- Non-public, purely internal, or private use of AI content.
- Purely technical image editing such as cropping, colour correction, or standard retouching, as long as the content is not substantially altered.
- Internal presentations and documents that are not published accordingly.
- AI assistance in the normal editing of existing content, if the input or its substantive meaning is not substantially altered (e.g., real product photos in front of an AI-generated background only).
Labelling AI Images: Practical Examples for Companies
These AI-generated images require labelling
Image 1: Photorealistic photo of a photographer in a vineyard landscape. Both the person and the location are entirely AI-generated but appear realistic enough to seem real.
Image 2: AI-generated customer testimonial with a fabricated face, fabricated person, and fabricated quote. The representation suggests a genuine customer voice, even though neither the person nor the statement is real.
Image 3: AI-generated image of the flooded Marienplatz. The real location is combined with an event that never occurred, creating the impression of an actual occurrence.
Image 4: Entirely AI-generated image in which the armchair is depicted as significantly higher quality than it actually is. Purchase-relevant features such as quality, size, or equipment must not be falsified by AI. Here two risks arise simultaneously: the labelling obligation under Art. 50 of the AI Regulation and a warning letter for unfair competition due to misleading representation. An “AI-generated” label does not automatically make a factually incorrect product representation permissible.
These AI-generated images do NOT require labelling
Image 5: Illustrated photographer in landscape in comic style. The representation is recognisably drawn and not photorealistic, therefore there is no risk of confusion with a real photo.
Image 6: A real photo with AI-assisted exposure or colour correction, without substantive changes. This is purely standard technical editing; the core message of the image remains unchanged.
Image 7: Real cup with company logo in front of a blurred, AI-generated background. The product itself is unchanged and depicted authentically; the background was merely adjusted for aesthetic purposes. Such minor adjustments do not substantially impair the perception of authenticity, therefore as a rule there is no labelling obligation.
Image 8: Abstract tech and marketing icon collage. The representation is clearly recognisable as symbolic graphics and not as a realistic photo, therefore there is no risk of deception.
How must AI content be labelled?
The AI Act does not prescribe a specific method, but sets three requirements for any labelling: it must be clear, distinguishable, and recognizable at the latest upon first encounter with the content. Additionally, accessibility should be considered. Since June 2026, the EU Commission provides three standardised, free icons for this purpose:
- “AI GENERATED” for entirely AI-generated content,
- “AI MODIFIED” for hybrid forms of human and AI editing, and
- a general “AI” base icon.
Those who use these icons do not need to separately justify why their own labelling is “equivalent.”
Additional Risks Beyond the AI Labelling Obligation
1. Competition Law and Advertising
Product images must not be altered by AI in a way that misrepresents purchase-relevant properties. If an AI-generated image depicts a product as higher quality, larger, or more feature-rich than it actually is, two risks arise simultaneously: the labelling obligation under Art. 50 of the AI Regulation and a warning letter for unfair competition due to misleading representation. An “AI-generated” label does not automatically make a factually incorrect product representation permissible.
2. Copyright
Purely AI-generated texts, images, and logos generally do not enjoy copyright protection in Germany because § 2 para. 2 UrhG (German Copyright Act) requires a personal intellectual creation by a human. Practical consequence: competitors can often legally adopt or modify pure AI results. Protection can only arise if a human creatively reworks the result to a substantial degree. This human contribution should be documented accordingly.
3. Fines
Violations of the transparency obligations under Article 50 can be penalised with fines of up to 15 million euros or 3% of global annual turnover, whichever is higher.
Liability in Influencer Collaborations: Who Bears the AI Labelling Obligation?
In AI-generated collaboration content, the influencer themselves is usually the “deployer” within the meaning of the AI Regulation. Anyone who posts professionally and enters into collaborations is directly obligated to label AI content under Art. 50 of the AI Act. For traditional advertising labelling (UWG), the contracting company may also be liable, as it commissioned the campaign and benefits from the advertising. Those working with agencies or creators should therefore secure this contractually: a disclosure obligation of the service provider for every use of AI, clear specifications for the wording and placement of the label, a prohibition on removing technical AI markings, and an indemnification clause in case AI use is concealed.
Important: The AI labelling does not replace advertising labelling. A post with AI-generated advertising content requires both notices simultaneously.
Practical Checklist: Labelling AI Content Correctly
- Check whether content falls into one of the three mandatory categories: deepfake, unedited text on matters of public interest, biometric/emotion recognition.
- For AI text drafts: always conduct editorial review and clearly assign approval to a responsible person.
- For realistic-looking AI images depicting persons, places, or events: label in case of doubt.
- Embed the label permanently in the content, not only in the caption or accompanying text – otherwise it will be lost in screenshots or downloads.
- Use the official EU icons “AI” / “AI GENERATED” / “AI MODIFIED” where possible.
- Additionally check product images for competition law violations due to misleading representation.
- Regulate labelling requirements contractually with external partners and influencers.
Frequently Asked Questions
Questions and answers on the topic of AI Labelling Obligation:
No, this is a widespread misunderstanding. The AI Act aims to prevent people from being deceived about the origin or authenticity of content, which is why the obligations only apply to certain use cases such as deepfakes or published AI texts of public interest. For example, if someone uses AI only for internal documents or meeting minutes, they are generally not subject to a labelling obligation.
AI-generated texts that serve to inform the public must be identified as such, for example through a visible notice that is also readable by screen readers. The obligation only ceases if genuine human review with the possibility to amend or reject takes place and a clearly identifiable person assumes editorial responsibility. A mere spelling check is not sufficient for this.
Deepfakes are AI-generated or substantially manipulated image, video, or audio content that resemble real persons, places, objects, facilities, or events and appear so realistic that they could seem authentic. The term is interpreted broadly and covers elements that merely resemble or could plausibly appear to be real things. Whether content qualifies as a deepfake also depends on context: in fictional productions or advertising, a different standard may apply if there is no potential for deception.
Yes, as a provider of an AI system, you are subject to even stricter obligations under Art. 50 para. 2: all synthetically generated image, audio, video, or text outputs must be recognizable as AI-generated in a machine-readable format, for example via C2PA metadata, invisible watermarks, or IPTC headers. The only exception is mere standard editing where the core message of the input data remains unchanged.
