EU AI Act Article 50 Labeling Rules for 2026
EU AI Act Article 50 divides chatbot disclosure and synthetic content labeling obligations between providers and deployers. This guide covers the rules applicable from August 2, 2026, the limited transition period, exceptions, and penalty structure.
- Article 50's transparency obligations apply from August 2, 2026.
- Providers of interactive AI must inform users that they are interacting with AI.
- Providers of synthetic content must apply machine-readable markings to outputs.
- Deployers of deepfakes and certain public-interest texts must visibly disclose the use of AI.
- The transition period until December 2, 2026, for certain existing systems is not a general grace period that postpones all obligations.
Article 50 of the EU AI Act requires chatbot disclosures, machine-readable markings for synthetic content, and visible disclosures for deepfakes and public-interest text from August 2, 2026. Only certain existing systems are granted a limited transition until December 2.
As of August 2026
What Changed on August 2, 2026
The transparency obligations under Article 50 apply from August 2, 2026. They cover AI providers and deployers that use AI professionally. The type of obligation varies depending on the system’s function and the purpose of the content.
The key changes fall into the following four categories.
- Interaction disclosures informing users that they are communicating with AI
- Machine-readable markings for synthetic images, audio, video, and text
- Visible disclosures for deepfakes and certain public-interest text
- Disclosures to people exposed to emotion recognition or biometric categorization systems
Article 50 is not a rule requiring the same wording to be attached to every AI output. Provider obligations and deployer obligations are distinct. Machine-readable markings and human-visible disclosures are also separate.
Comparison of Provider and Deployer Obligations
Providers are responsible for transparency at the system level. Deployers are responsible for the context in which content is actually used. Whether one company may hold multiple roles should be checked against the relevant definitions and obligations in Regulation (EU) 2024/1689.
| Category | Main obligation | Intended recipient of disclosure | Typical examples |
|---|---|---|---|
| Conversational AI provider | Disclose that the user is interacting with AI | User | Chatbot, AI customer service representative, AI agent |
| Generative AI provider | Apply machine-readable markings to outputs | Distribution systems and detection tools | Synthetic images, audio, video, and text |
| Deepfake deployer | Visibly disclose that content was artificially generated or manipulated | People viewing the content | Synthetic video that appears to depict a real person |
| Public-interest text deployer | Disclose that content was generated or manipulated by AI | Readers of the text | Text addressing public affairs or social issues |
| Emotion and biometric system deployer | Disclose that the system is operating | People exposed to the system | Users subject to emotion recognition or biometric categorization |
Here, a provider is an entity that develops an AI system and places it on the market. A deployer is an entity that uses a system under its authority. Purely personal, non-professional activities are excluded from the definition of deployer.
Disclosure Standards for Chatbots and AI Agents
As a rule, conversational AI must disclose that it is AI. The timing of the disclosure must be checked in the text of Article 50 of Regulation (EU) 2024/1689. The point at which users must be informed should be checked in the text of Article 50 of Regulation (EU) 2024/1689.
An exception may apply, however, where it is obvious that the system is AI. This is assessed from the perspective of a reasonably well-informed, observant, and circumspect user. It is not sufficient that the developer internally classifies it as AI.
An exception may also apply where the law authorizes criminal investigations or similar activities. This exception operates only within the scope of the relevant legal authority. It does not automatically apply to ordinary customer service or marketing.
Machine-Readable Markings for AI-Generated Content
Generative AI providers must place detectable markings on synthetic outputs. Covered content includes images, audio, video, and text. A marking is a technical mechanism distinct from a human-visible label.
Markings must take into account effectiveness, interoperability, robustness, and reliability. Technical feasibility and implementation costs must also be considered. The generally acknowledged state of the art at the time is another factor.
Exceptions may apply to functions that assist with standard editing. Functions that do not substantially alter the meaning of the input may also qualify. The boundary between simple correction and the generation of new meaning should be documented.
Labeling Deepfakes and Public-Interest Text
Deepfake deployers must disclose the manipulation in a way that people can see. A deepfake is synthetic content made to appear as though it depicts a real person, object, place, entity, or event. The definition of deepfake in Article 50 focuses on images, audio, and video.
The disclosure obligation does not disappear for artistic, creative, satirical, or fictional works. However, disclosure in a manner that does not hamper the display or enjoyment of the work is permitted. The location and format of the disclosure may be adapted to the context of the content.
AI-generated or manipulated text published to inform the public on matters of public interest is also subject to disclosure. An exception may apply where there has been human review or editorial control. The natural or legal person holding editorial responsibility for publication must also be clearly identified.
The mere fact that a person clicked the publish button is not sufficient. It must be possible to demonstrate an actual review process and editorial responsibility. Review records provide a basis for determining whether an exception applies.
Summary by Condition
Whether a disclosure is required must be assessed by considering both the content format and the context of use. The following table provides a quick distinction between the boundaries of Article 50. Individual cases should be assessed based on the actual functionality and distribution method.
| Condition | General treatment | Issues to verify |
|---|---|---|
| Chatbot that users cannot clearly recognize as AI | Provide disclosure at the first interaction | Placement and accessibility of the disclosure |
| AI generates a new image, audio, video, or text | Apply a machine-readable marking | Persistence and detectability of the marking |
| Performs only standard editing of color, noise, or similar elements | Possible exception under Article 50(2) | Whether the meaning of the input was substantially altered |
| Publicly releases synthetic video that appears to depict a real person | Apply a visible deepfake label | Whether appropriate disclosure is provided even for a creative work |
| Publishes unedited AI text concerning matters of public interest | Apply a visible AI-use label | Whether there was human review and editorial responsibility |
| Used by an individual for non-professional purposes | Possible exclusion from deployer obligations | Whether the use has shifted to business or professional activity |
| Created outside the EU with outputs used in the EU | May be covered | Placement on the EU market and where the outputs are used |
Companies outside the EU are not automatically excluded. The rules may apply if they place a system on the EU market. Outputs from operators in third countries may also be subject to review if they are used in the EU.
Comparison of Disclosure Methods
Machine-readable markings and visible disclosures serve different purposes. Applying one does not satisfy the other obligation. The same content may require both methods.
| Disclosure method | Main audience | Purpose | Implementation examples |
|---|---|---|---|
| Machine-readable marking | Platforms and detection systems | Automatically identify whether content is synthetic | Metadata, watermarks, provenance information |
| Visible disclosure | General users | Directly communicate that content has been manipulated | On-screen wording, audio notice, content description |
| Interaction disclosure | Chatbot users | Communicate that the other party is AI | Notice before the conversation begins, disclosure in the first response |
No single specific technology is fixed in the legal text. Markings must remain robust after transformation and redistribution. An appropriate combination should be designed for the characteristics of the service.
Example of Calculating the Transition Period for Existing Systems
The limited transition period is not a general grace period for all existing services. It may apply only until December 2, 2026 to certain existing systems covered by Article 50(2). The precise scope must be checked in the European Commission’s guidelines.
- Check whether the system was placed on the market before August 2, 2026.
- Determine whether the obligation concerns the machine-readable marking under Article 50(2).
- Compare the system against the eligibility conditions for the limited transition in the EU guidelines.
- If eligible, complete the technical measures by December 2, 2026.
- Apply chatbot disclosures and deployer disclosure obligations separately.
For example, an existing generative system may be combined with a chatbot. A limited transition may be considered for markings on generated content. The application date for chatbot disclosures must be checked in Article 50 of Regulation (EU) 2024/1689 and the provisions concerning application and transition.
Common Mistakes
The most common misunderstanding is assuming that adding a single statement to AI content is sufficient. In practice, provider and deployer obligations are separate. The output format and disclosure context must also be assessed separately.
- Mistaking a machine-readable marking for a visible deepfake label
- Assuming that all existing systems have a grace period until December 2
- Treating business promotion as personal use merely because it is conducted through a personal account
- Assuming editorial control merely because a person uploaded the content
- Treating every simple AI assistance function as covered by the standard-editing exception
- Assuming that deepfake disclosure is waived for artistic works
Accessibility requirements are also easy to overlook. Disclosures must be provided in a clear and distinguishable manner. Users with disabilities must also be able to perceive them.
Designing Evidence Across Distribution Stages
Disclosure obligations must be managed not only when content is generated but also throughout the distribution process. Technical markings may disappear through compression, conversion, or re-uploading. Visible wording may also be omitted from the distribution screen.
It is safer for operational records to link the following items.
- Model and system version used
- Type of content generated or manipulated
- Method used to apply the machine-readable marking
- Visible disclosure wording and timing of exposure
- Human reviewer and entity holding editorial responsibility
- Tests confirming that markings are retained during transformation and distribution
These records are not a separate certificate required by law. However, they provide evidence explaining how the obligations were fulfilled. Responsibility for retaining disclosures may also be allocated in supply contracts.
Supervisory Authorities and the Reporting and Penalty Structure
The authorities responsible for enforcing Article 50 should be identified by checking the enforcement provisions of Regulation (EU) 2024/1689. Member States’ market surveillance authorities and competent authorities conduct on-site supervision. Authority over providers of general-purpose AI models may lie with the European Commission and the AI Office.
Affected persons may lodge complaints with a market surveillance authority. The authority that receives a complaint depends on each Member State’s designated structure. The official AI supervisory contact point in the country where the service was provided should be checked.
Violations of Article 50 may fall within the penalty category for breaches of other obligations under the EU AI Act. The specific penalty ceiling must be checked in the penalty provisions of Regulation (EU) 2024/1689. The actual amount is determined by taking into account factors such as the nature of the violation and the size of the company.
Separate ceiling principles apply to small and medium-sized enterprises. Fines are not imposed as an immediate fixed amount upon a violation. Member State procedures and proportionality assessments also apply.
Where to Check the Legal Text and Detailed Guidelines
The original legal obligations are set out in Article 50 of Regulation (EU) 2024/1689. Detailed interpretations should be checked in the European Commission’s Article 50 transparency guidelines. The official FAQ supplements examples involving chatbots, deepfakes, and public-interest text.
Codes of practice are tools for aligning technical markings and labeling methods. Participation in a voluntary code does not eliminate the legal obligations themselves. The legal text and enforcement guidelines remain the primary criteria for assessment.
Frequently Asked Questions
Must AI chatbots always be labeled as AI?
As a rule, users must be informed that they are interacting with AI. However, an exception may apply if it is obvious from the circumstances that the system is AI. The timing of the disclosure must be checked in the text of Article 50 of Regulation (EU) 2024/1689.
Does a machine-readable marking alone also satisfy the deepfake disclosure requirement?
The two obligations are mechanisms intended for different recipients. A technical marking assists machine detection. A deepfake disclosure must inform people that the content has been manipulated.
Must every text written by AI be labeled?
Not all AI-written text is subject to deployer disclosure. The primary target is text published to inform the public on matters of public interest. An exception may apply where there is human review and editorial responsibility.
Must Korean companies also comply with Article 50?
The rules may apply if they provide AI systems in the EU market. Outputs created in a third country may also be covered if they are used in the EU. Applicability cannot be determined solely by the location of the legal entity.
Are all obligations postponed until December 2, 2026?
No. The limited transition applies only to certain existing systems. It is not a period that uniformly postpones chatbot disclosures and deployer disclosure obligations.
FAQ
When does Article 50 of the EU AI Act begin to apply?
The transparency obligations apply from August 2, 2026. A limited transition until December 2, 2026 may be allowed for machine-readable markings on certain existing systems.
Must AI chatbots always disclose that they are AI?
The point at which the fact that it is AI must be disclosed should be verified in the original text of Article 50 of Regulation (EU) 2024/1689. An exception may be possible where it is obvious to a reasonably well-informed, observant, and circumspect user that it is AI, but the service must not make this determination arbitrarily.
What markings are required for AI-generated content?
Providers of generative AI must apply machine-readable markings to synthetic images, audio, video, and text. Deployers that make deepfakes or certain public-interest texts publicly available may also be required to provide a separate disclosure that is recognizable by people.
Is it sufficient to add only a machine-readable marking to a deepfake?
No. Machine-readable marking is an obligation on providers for automated detection. Deployers of deepfakes must also disclose that the content was artificially generated or manipulated in a way that people can see.
Must deepfakes created for artistic or satirical purposes also be labeled?
The disclosure obligation does not disappear solely because the content is artistic, creative, satirical, or fictional. However, it may be labeled in an appropriate manner that does not hamper the display or enjoyment of the work.
Are all news articles or public-interest texts written by AI subject to labeling requirements?
The primary targets are AI-generated or manipulated texts published to inform the public about matters of public interest. An exception may be possible if there has been substantial human review or editorial control and the party holding editorial responsibility is clearly identified.
Is an AI-generated content marking also required for simple photo retouching?
Whether the function falls under an exception should be verified in the original text of Article 50 of Regulation (EU) 2024/1689. Work that creates new subjects or meaning is difficult to regard as simple editing.
Does this also apply to AI content created by individuals as a hobby?
Purely personal, non-professional activities may be excluded from the definition of a deployer. However, if the activity extends to business promotion, paid work, or professional posting, the nature of the use must be reassessed.
Does Article 50 also apply to AI companies outside the EU?
It may apply if they place systems on the EU market or provide services in the EU. The scope of application must also be reviewed where output generated outside the EU is used in the EU.
How much is the fine for violating Article 50?
The specific maximum penalty for the category of violations of other obligations should be verified in the penalty provisions of Regulation (EU) 2024/1689. The actual penalty is determined by taking into account factors such as the nature and duration of the violation, the size of the company, and proportionality.
Sources
Images
