Guidelines on transparency obligations for providers and deployers of AI systems

The Guidelines and the summary provided aim to serve as practical guidance for all actors in the AI ecosystem – from competent authorities to providers and deployers of AI systems – as well as for interested consumers. They aim to assist in implementing the transparency obligations under Article 50 of the AI Act in a consistent, effective and uniform manner.

1. Background and objectives

The European Commission drew up the Guidelines together with a Code of Practice on the marking and labelling of AI-generated content. While the Code of Practice focuses on the labelling of content, the Guidelines cover the whole scope of the legal transparency obligations.

The main content of the Guidelines is summarised below. Reference is made to the full version of the Guidelines for the exact details.

Although the Guidelines are non-binding, they provide appropriate guidance in cases of legal uncertainty.

2. Overview and horizontal topics

Article 50 of the AI Act lays down comprehensive transparency obligations for AI systems (see above). Implementation of these obligations requires an understanding of the division of roles between providers and deployers, as different obligations apply.

Detailed explanations of these roles can be found here.

Providers must embed transparency into design and operation, while deployers must take proportionate measures to ensure the labelling of content that they implement.

These obligations may apply cumulatively, so a single AI system may have to meet several different requirements and responsibility for compliance with the obligations may be divided between different operators.

Certain activities are exempt from the transparency obligations, including the personal, non-professional use of AI systems and use for research and development. Providers of open-source AI systems nevertheless still need to comply with the obligations.

Compliance with the transparency obligations under Article 50 does not conflict with other provisions of the AI Act, in particular the prohibitions under Article 5 and the requirements for high-risk AI systems.

3. Transparency for interactive AI systems

Interactive AI systems – such as chatbots and virtual assistants – require particular attention with respect to transparency because they interact directly with users. Article 50(1) requires providers to inform users that they are interacting with an AI system in order to enable users to make informed decisions and avoid overreliance.

Article 50(1) – information obligation for AI systems interacting directly with natural persons (summary)

Obligation
Providers of AI systems directly interacting with natural persons must clearly inform users that they are communicating with an AI system.
Conditions (all four must be met)
1. The system is an AI system within the meaning of Article 3(1).
2. The system is intended for (bidirectional) interaction with natural persons.
3. Interaction is direct (no purely technical interface).
4. Interaction is with a natural person.
Exceptions (outside scope)
- Backend or machine-to-machine solutions
- Closed industrial/production settings
- Pure recommender, filter or decision-support systems without direct user contact
Information requirements
- Clear, understandable and by the time of the first interaction (according to Article50(5))
- Needs of vulnerable groups to be taken into account
- No prescribed format: text, icon/watermark, audio or multimodal cues
- Insufficient: only in terms and conditions, only technical descriptions not referring to the interaction, imperceivable machine-readable markings
Exceptions (to be interpreted strictly)
1. Obvious interaction with an AI system – if the artificial nature is clearly obvious to an average natural person who is reasonably informed, observant and circumspect (depending on the context and the audience)
2. Authorisation by law for law enforcement purposes, subject to appropriate safeguards being in place

Interplay with other legislation
The transparency obligation supplements existing information requirements in consumer protection, data protection, digital and non-discrimination legislation.

[ENDE]

3.1 Main components and transparency obligations

The transparency obligation under Article 50(1) applies to providers of AI systems directly interacting with natural persons, unless this interaction falls within certain exceptions.

3.1.1 AI systems for direct interaction with natural persons

Four conditions must be met for a system to fall within the scope of this rule:

The system is an AI system: The system must fulfil the definition in Article 3(1) (see the exact definition here). Simple, non-AI automated mechanisms are excluded.

The system is intended to interact: The system must be designed for a bidirectional exchange with a natural person – irrespective of the form, initiator or duration of the interaction.

The interaction is direct: The interaction must be in real time or near real time; the mere possibility for humans to review or intervene in the outputs does not provide exemption from the disclosure requirement.

The interaction is with natural persons: The system must interact with natural persons – whether professional deployers, end-users or other persons using the system on their behalf.

Examples of AI systems falling within the scope of Article 50(1)
- AI-enabled voice assistants, chatbots, AI avatars and bots in social media
- AI agents capable of interacting with the persons instructing them or with other natural persons

Examples of AI systems falling outside the scope of Article 50(1)
- Industrial robots operating in a closed setting
- Algorithms for content recommendation, spam filters and pure decision-support systems without direct interaction
- AI systems operating in closed physical environments, backend machine-to-machine calls between AI systems whose outputs are not intended for natural persons, and virtual environments without any direct contact with natural persons

3.1.2 Information obligations under Article 50(1)

Providers of interactive AI systems must inform users in a clear and understandable way that they are communicating with an AI system before the interaction. The information to be disclosed includes the limitations of AI, possible errors, and information about data usage and human intervention.

The AI Act does not specify any particular disclosure format. Providers can use different techniques as long as they fulfil the requirements of Article 50(5) (clear and distinguishable information by the time of the first interaction) and take into account the needs of vulnerable groups. Possible techniques include:

textual disclosure: clear, prominently positioned labels or first-turn greetings using simplified wording (particularly for children and users with lower digital literacy);

visual cues: icons or visible watermarks indicating the AI nature of an interaction, ideally standardised;

auditory disclosure: spoken statements in voice assistants at the beginning and periodically during longer interactions. Support through distinct audio cues (e.g., “earcons”) is possible, but not sufficient on its own for recognition.;
multimodal combinations: text, audio and visual cues combined for clarity and better accessibility.

Wording contained only in terms and conditions or technical descriptions not directly referring to the interaction are not sufficient. Imperceivable machine-readable markings do not fulfil this transparency obligation either. Unclear or ambiguous signals, human-like representations and generalised disclosures are likewise insufficient.

3.2 Exceptions

There are two exceptions to the information obligation:

Obvious interaction: The artificial nature of the interaction is obvious to an average natural person who is reasonably informed, observant and circumspect. This requires an assessment of the audience and the context.
Systems authorised by law: AI systems that are authorised by law to detect, prevent, investigate or prosecute criminal offences are exempt from the information obligation, subject to appropriate safeguards for the rights and freedoms of third parties being in place.

3.3 Interplay with other Union legal acts

The transparency obligation under Article 50(1) applies without prejudice to existing EU consumer protection, data protection, digital or non-discrimination legislation that may require other information disclosures. The information obligations under the AI Act are contextual and aim to ensure the awareness of users about interaction with an AI system.

4. Marking and detection of AI-generated or manipulated content

Article 50(2) requires providers of AI systems to implement technical solutions for the detection and marking of synthetic content. The objective is to enable users to distinguish between AI-generated content and content created by humans and to increase the integrity of the information ecosystem.

The beneficiaries of the transparency obligation under Article 50(2) are all natural persons exposed to AI-generated or manipulated content, as well as key actors who play an important role for trust in the information ecosystem, such as market surveillance authorities, researchers, media, online platforms, search engines and content creators.

Article 50(2) – marking obligation for AI-generated content (summary)

Obligation
Providers of AI systems (including general-purpose AI (GPAI) systems) generating or manipulating synthetic audio, image, video or text content must implement technical solutions to clearly mark this content.
Scope (all conditions must be met)
1. The system is an AI system within the meaning of Article 3(1).
2. The system is capable of generating and/or manipulating synthetic content.
3. The output is in one of four modalities (or multimodal).
4. The system does not fall within an exception.
Objective: ability to distinguish between AI-generated or manipulated content and content created by humans; protection of the integrity of the information ecosystem
Exceptions (outside scope)
- Only selection/ranking of existing content (playlists, recommendations)
- Mere industrial sensor/measurement data, source code
- Machine-to-machine outputs without exposure to humans
- Intermediate outputs in closed development/production workflows (unless they are the final output)

Permissible marking techniques: watermarks, metadata identifications, cryptographic methods for proving provenance, logging, fingerprints – implementable at any stage of the value chain; responsibility remains with the provider
Technical requirements
- Human-readable results
- Ideally open industry standards, locally executable
- Interoperable, effective, reliable, robust and in line with the state of the art; costs must be proportionate
Three exceptions
1. Assistive function for standard editing (for example spellchecking/grammar correction, translations, format conversions, minor image/video adjustments)
2. No substantial changes to the input data or their semantics (case-specific assessment; semantic changes such as summaries, face swaps and voice cloning require marking)
3. Authorised by law for law enforcement purposes

[ENDE]

4.1 Main components and concepts

The provisions of Article 50(2) apply if the following cumulative conditions are met:

The system is an AI system: The system must fulfil the definition of an AI system within the meaning of Article 3(1) (see the exact definition here).

The AI system is capable of generating or manipulating synthetic content: This covers both the creation of new content and the modification of existing content.

The content is in one of the following modalities: Audio, image, video or text. This also includes multimodal content, that is, a combination of these modalities.

The system does not fall within an exception
Certain applications such as assistive functions for standard editing or use by law enforcement authorities.

4.1.1 AI systems generating or manipulating synthetic content

Article 50(2) applies to all AI systems, including GPAI systems, generating synthetic audio, image, video or text content. Content generation by AI systems is generally not a distinguishing characteristic because all AI systems generate output. Article 50(2) also covers AI systems used to manipulate content.

Examples of AI systems falling within the scope of Article 50(2)
- Systems creating deep fakes
- AI-assisted image editing programs
- Chatbots generating synthetic text
- Systems manipulating audio recordings
- AI agents taking actions whose output is perceptible as synthetic content

4.1.2 Modalities of synthetic content

The modalities of synthetic content covered by the transparency obligation include the following:

audio: time-varying signals presenting sound that are capable of being heard by humans, such as speech, instrumental music and other audio signals;
image: static spatial representations encoding visual information that are capable of being seen by humans;
video: time-based sequences of images, which may be synchronised with audio, that are capable of being seen by humans;
text: discrete symbolic content composed of characters or numbers that are capable of being read and interpreted semantically by humans.

This also includes content made up of a mix of these modalities (multimodal content).

4.1.3 Content falling outside the scope of Article 50(2)

Certain outputs of AI systems are exempt from the marking obligation because they either merely reproduce existing content, they are not perceptible by humans or the risk of misinformation and manipulation is low.

Examples of content falling outside the scope of Article 50(2)
- Music playlists, recommender systems that only select or rank existing content without creating anything new or manipulating the content itself
- Mere data recordings and transmissions in industrial settings, such as measured data from sensors and GPS data from vehicles
- Source code written in a programming language and interpreted, compiled or executed by a computing system, including natural-language comments and contextual information that form an integral part of the source code
- Outputs that are exclusively communicated and automatically processed by machines without being perceived by humans, such as agent-to-agent communication and anti-spam signals
- Outputs that are only used in closed loop environments in industrial and product development workflows, unless they are the final output of the AI system in the form of AI-generated or manipulated text, audio, image or video content

4.2 The marking and detection obligation

Article 50(2) requires providers of generative AI systems to implement technical solutions for the marking and detection of their output. Each of these two elements must be fulfilled to effectively achieve the objectives of the transparency obligations:

Marking in a machine-readable format: The output of the AI systems must be marked in a machine-readable format.
Detectability: It must be ensured that this marking is also detectable to enable users and other relevant actors to distinguish between AI-generated content and content created by humans.

Technical solutions complying with Article 50(2) are a combination of both marking and detection techniques. A technique for marking without a means for detection is not sufficient.

4.2.1 The marking element

Providers must implement machine-readable marking. Perceptible marks and labels are allowed as a complementary measure. Providers can use a single technique or a combination of techniques, as long as the solution is machine-readable and meets the requirements for effectiveness, reliability, robustness and interoperability (see below).

Examples of marking techniques
- Watermarks
- Metadata identifications
- Cryptographic methods for verifying provenance and authenticity of content
- Logging methods
- Fingerprints
- Combinations of these techniques

Providers can implement the marking solution at different stages of the value chain, for example after content generation, at the level of the AI model or in the inference process. They can also use solutions from upstream providers or third parties, provided that the solutions comply with Article 50(2). Responsibility for compliance with this provision nevertheless remains with the provider of the AI system.

4.2.2 The detection element

Providers of AI systems must ensure that AI-generated or manipulated content is detectable as such for users. Providers must ensure that means of detection are available that are capable of producing human-readable results indicating whether content has been AI-generated or manipulated.

These detection solutions should ideally be based on publicly-available industry standards to ensure full interoperability. This enables third parties to implement detection and allows seamless integration into self-contained networks. The detection solutions should ideally be locally executable on digital devices.

Where there are no publicly-available standards, in particular at the initial stage of the implementation of Article 50(2) for watermarking technologies, providers can temporarily use their own, third-party or shared detection solutions. These solutions must, however, ensure interoperability with the systems used by other providers. This interim solution can be used until harmonised standards are in place and a standardised, provider-agnostic, interoperable detection solution is developed that is secure, privacy-preserving and locally executable.

If a provider of an AI system depends on another actor for the provision of or access to a detection solution, the provider must ensure that the solution complies with the requirements of Article 50(2) and (5). In addition, the actor must display the detection results in compliance with Article 50(5).

4.2.3 Requirements for technical solutions

The technical solution for marking and detection must be effective, interoperable, robust and reliable.

Effectiveness: The capability to detect marks and enable a distinction to be made between AI-generated content and content created by humans.

Reliability: The capability to accurately identify AI-generated or manipulated content in nominal conditions.

Robustness: The capability to remain effective and reliable under varying conditions and in the event of attacks.

Interoperability: The capability to operate across multiple systems and actors.

Providers are responsible for ensuring that the combination of technical solutions meets all the requirements and is technically feasible. The solutions must be capable of being implemented using currently available technologies and methods and must comply with the state of the art. Providers can take account of the costs of implementation to ensure an appropriate balance between costs and benefit.

4.3 Exceptions to the obligations

Article 50(2) provides three explicit exceptions to the marking and detection obligations for AI-generated or manipulated content:

Assistive functions for standard editing

The first exception concerns AI systems performing an assistive function for standard editing. This comprises the process of preparing existing content for publication or distribution, for example small edits to improve readability, grammar, quality and format. It does not involve generating new content.

Examples of standard editing that benefit from the exception and therefore do not require marking under Article 50(2)

Text editing
- Grammar correction and spellchecking
- Linguistic and stylistic polishing without changing the substance
- AI-generated translations of text
- Transcriptions of conversations
- Assistive technologies for persons with disabilities (for example augmentative and alternative communication (AAC) or customized neural voices (CNV)) that transform human input without semantic changes

Image editing
- Formatting and format conversions
- Technical compression and noise reduction for enhanced clarity
- Minor cropping and colour adjustments
- Removal of dust spots or red-eye
- Obscuring backgrounds visible in the original file
- Pixelation or blurring of faces
- Standard adjustments to colour and contrast
- Minor adjustments to playback speed
- Corrections to level the horizon and pixel filters to amplify certain parts of an image
- Applying colour maps to grayscale images and vice versa
- Edge image completion and aspect ratio adaptation
- Automatic transition clips and other non-substantive cosmetic/technical edits

Medical image processing
- Medical image technical processing or reconstruction within medical devices (for example annotations, boxes, contours, circles, heatmaps)

Video-editing
- Scaling of video clips
- Dynamic range compression and equalisation
- Limited video stabilisation
- Standard adjustments to colour and contrast

No substantial changes to the input data

The second exception applies where an AI system does not substantially alter the input data or their semantics. A case-specific assessment is always required.

Examples of semantic changes that require marking under Article 50(2)
Text content
- AI-generated summaries of text
- Paraphrasing or rewriting text that changes style, structure and meaning beyond mere grammatical and stylistic correction
- AI-generated content transforming authentic human input using assistive technologies and altering semantically the meaning of the content

Image and video content
- Removal, replacement or insertion of objects or persons in existing images and videos that changes the meaning and substance of the content
- Face replacement or substantial facial modification
- Creation of composite images or video clips modifying the representation of persons, objects, events or facts
- Substantial changes to the body shape or skin colour of a person
- Extreme lightening, darkening or colour changes that change the meaning, intent and messaging of the content

Audio content
- Synthesis of realistic speech in a specific person’s voice
- Generation of realistic videos depicting events that did not occur

Authorised by law for law enforcement purposes

The final exception to the marking and detection obligations applies to generative AI systems authorised by law to generate or manipulate synthetic content to detect, prevent, investigate or prosecute criminal offences.

4.4 Interplay with other Union legal acts

Article 50(2) relates to how content has been created and its artificial origin, not who created the content. Providers must use marking and detection solutions that comply with EU data protection law, such as data protection by design and data minimisation.

The AI Act highlights the relevance of the transparency obligations (Article 50(1), (2) and (4)) for the implementation of the Digital Services Act (DSA), in particular for very large online platforms (VLOPs) and very large online search engines (VLOSEs) and very large online search engines (VLOSEs). Providers of VLOPs and VLOSEs must assess and mitigate systemic risks from AI-generated content, in particular negative effects on democratic processes and civic discourse. Machine-readable marks help to detect and label such content and thus diminish the risk of disinformation. Providers of VLOPs and VLOSEs may also qualify as providers of AI systems if they integrate AI systems into their services.

The marking obligations for AI-generated content (Article 50(2) and (4)) are without prejudice to the processing of notices on illegal content by hosting service providers (Article 16(6) DSA). The two sets of obligations apply independently of each other. The legality of content is assessed on the basis of the applicable law.

Example
If a labelled deep fake image is flagged as infringing on trademark or copyright law, the application of the label does not affect the potential illegality of the content. The assessment is conducted solely on the basis of the applicable law.

Providers of online platforms and online search engines and other actors are encouraged to preserve and enable the marking of AI-generated or manipulated content (Article 50(2)) in order to disclose the origin of content to users.

5. Emotion recognition systems and biometric categorisation systems

An emotion recognition system infers an inner state (emotion/intention) of a natural person, while a biometric categorisation system sorts the natural person into a group/category.

These two types of system were first explained in more detail in the Commission Guidelines on prohibited artficial intelligence practices. They can be distinguished as follows:

**Emotion recognition systems and biometric categorisation systems**
Characteristic	Emotion recognition system	Biometric categorisation system
Key question	“How is the person feeling at the moment?” or “What is the person intending to do?”	“Which group does the person belong to?”
Objective/purpose	Identifying or inferring emotions or intentions	Assigning a person to a certain category
Examples of cues	Happiness, sadness, anger, surprise	Facial features, fingerprints, voice recognition; gait, facial expressions
What it is NOT	Not identifying physical states (such as fatigue and pain)	Not checking a person’s identity (it is not a question of who the person is but which group the person belongs to)
Special feature	Distinguishes between recognising (direct comparison) and inferring (generating new information through learning processes)	Only applies to categorising individual persons (not analysing whole groups)

Both use biometric data. Article 50(3) requires deployers of emotion recognition systems and biometric categorisation systems to inform natural persons concerned about the operation of the systems, unless an exception applies. This serves to protect privacy and guarantee transparency.

Article 50(3) – information obligation for emotion recognition and biometric categorisation systems (summary)

Obligation
Deployers of systems using biometric data must inform natural persons concerned that an emotion recognition system or a biometric categorisation system is being operated.
Relation to classification as high-risk
- Emotion recognition systems are classified as high-risk (Annex III(1)(c)) – unless they are prohibited under Article5(1)(f) (workplace/education areas).
- The transparency obligation under Article 50(3) applies in addition to the high-risk requirements.
- The information obligation applies to biometric categorisation systems, regardless of whether or not they are classified as high-risk.
Scope of information
- Information that the natural person concerned is being exposed to the operation of such a system (exposure/operation)
- No details of purposes or legal basis (this is regulated by the General Data Protection Regulation (GDPR))
Format and timing
- Clear, understandable and accessible (in particular for children and persons with disabilities)
- At the latest at the time of the first exposure
- No prescribed format: in writing, by standardised icons, orally or using combinations of these – depending on the individual context (place of deployment, audience, type of exposure)
Exceptions
- Systems permitted by law to detect, prevent or investigate criminal offences, subject to appropriate safeguards (such as access restrictions and logging) being in place
Legal classification
- The information obligation does not automatically make use lawful and does not replace the prohibition under Article 5 or under non-discrimination legislation.
- It supplements the information requirements under data protection law (such as the GDPR) and can, for example, be integrated into the privacy notice or consent.

[ENDE]

5.1 Main components, concepts and transparency obligations

This section explains the terms “emotion recognition system” and “biometric categorisation system” and the related transparency obligations under Article 50(3).

5.1.1 The notion of an emotion recognition system

An “emotion recognition system” is defined in Article 3(39) as an AI system for the purpose of identifying or inferring emotions or intentions of natural persons on the basis of their biometric data. The detailed explanation of this notion provided in the Commission guidelines on the classification of high-risk systems, which are still being finalised, is equally relevant in the context of Article 50(3).

Since all emotion recognition systems are classified as high-risk under Article 6(2) and Annex III(1)(c) (unless prohibited under Article 5(1)(f) in the areas of workplace and education), the transparency obligations of Article 50(3) apply in conjunction with the other requirements applicable to high-risk systems.

5.1.2 The notion of a biometric categorisation system

A “biometric categorisation system” is defined in Article 3(40) as an AI system for the purpose of assigning natural persons to specific categories on the basis of their biometric data, unless it is ancillary to another commercial service and strictly necessary for objective technical reasons. The detailed explanation of this notion provided in the Commission guidelines on the classification of high-risk sysstems, which are still being finalised, is equally relevant in the context of Article 50(3).

Article 50(3) applies to all biometric categorisation systems (for example age or gender classification based on biometric data), regardless of whether or not they fall within the scope of high-risk AI systems under the AI Act.

5.1.3 The information obligation under Article 50(3)

Deployers are required to inform natural persons concerned that they are exposed to an emotion recognition system or a biometric categorisation system. The AI Act does not require the reasons for the system’s operation or other purposes to be disclosed, as this is regulated by the GDPR.

The information must be clear, understandable and accessible for all natural persons concerned, in particular children and persons with disabilities. The information can be provided in various ways depending on the context, including:

place of deployment, for example:
- virtual gaming platform,
- bricks-and-mortar store,
- train station;
addressees, for example:
- children,
- the elderly,
- persons with disabilities,
- customers;
type of exposure, for example:
- long-term,
- one-off,
- continuous,
- intermittent

The information can be provided in writing, by standardised icons (also when presented electronically), orally or using combinations of such ways. The information must be provided at the latest at the time of the first exposure of the natural person concerned.

Examples of providing information
- A pop-up window when a computer game is launched indicating that the person’s face is recorded to capture their emotions.
- A clearly visible notice at the entrance to an exhibition room with information that visitors’ facial images are captured to assign them to a specific age group.

5.2 Exception for law enforcement purposes

The information obligation does not apply to emotion recognition and biometric categorisation systems that are permitted by law to detect, prevent or investigate criminal offences, subject to appropriate safeguards for the rights and freedoms of third parties being in place.

While the other exceptions in Article 50 require an explicit law authorising the non-transparent use of AI systems, Article 50(3) makes such use possible when permitted under the legal rules governing the powers of law enforcement authorities to use emotion recognition and biometric classification systems to detect, prevent or investigate criminal offences without informing the persons concerned.

5.3 Interplay with other Union legal acts

The information obligation under Article 50(3) does not automatically make the use of an emotion recognition system or biometric categorisation system lawful. Nor does it legitimise intrusive or discriminatory uses that might infringe on Article 5 or other EU law.

Deployers must comply with the information obligations in Article 50(3) in conjunction with the applicable information requirements under EU data protection law. The information can, for example, be integrated into the privacy notice or when obtaining consent.

6. Labelling of deep fakes and certain text publications

Article 50(4) lays down two obligations for deployers of generative AI systems:

deployers must clearly label deep fakes as such;
deployers must also label AI-generated or manipulated text that is published with the purpose of informing the public on matters of public interest, unless certain exceptions apply.

Article 50(4) – labelling obligations (summary)

Obligations
- First subparagraph – deep fakes: AI-generated/manipulated image/audio/video content that falsely appears to be authentic must immediately be visibly or audibly labelled as being of artificial origin.
- Second subparagraph – AI text: AI-generated or manipulated text that is published with the aim of informing the public on matters of public interest must be labelled as such.
Criteria for deep fakes (all three must be met)
- High level of similarity to a real subject
- Relation to an object/event that actually exists or can plausibly exist
- Creation of a false appearance of authenticity (assessment based on image/message, context, public)
Exceptions for deep fakes
- Artistically creative, satirical or fictional works: labelling only in a form that does not hamper the work, if classification is evident
- Law enforcement purposes: full exemption if system is authorised by law to detect, prevent, investigate or prosecute criminal offences (subject to safeguards)
Labelling of AI text – conditions
- Publication for the general public (not private)
- Purpose is information
- Matter of public interest (politics, public security, public health, environmental protection, economic, scientific or cultural development, etc)
  
  Obligation does not apply if:
- the content undergoes a process of human editorial control (fact/source checks) and a natural or legal person holds editorial responsibility;
- the exception for law enforcement purposes applies.
Format and timing
- Clear, understandable and accessible (in particular for children and persons with disabilities)
- Immediately by the time of the first publication or first public contact
- No prescribed format: text, symbol, orally or a combination
Relevant legal references
- DSA Art. 35(1)(k) (VLOP/VLOSE markings)
- GDPR (personal data)
- Trademark rights, copyrights and personality rights
- European Media Freedom Act (EMFA) (editorial responsibility)

[ENDE]

6.1 Main components, concepts and transparency obligations for deep fakes

Article 50(4) first subparagraph applies if several conditions are fulfilled:

the system must be an AI system;
the system is used by deployers for professional purposes;
the system is used to generate or manipulate image, audio or video content that constitutes a deep fake;
the AI system must not fall under the exception.

6.1.1 The term “deep fake”

The term “deep fake” is defined in Article 3(60) as:

“AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful”.

For the purpose of clarifying this definition, the following four elements are decisive:

Resemblance: AI-generated content must bear an appreciable resemblance to an existing subject – not necessarily perfect, but sufficiently distinguishable. The assessment is the responsibility of the deployer and must be based on an objective comparison.

Existing: Content must resemble something existing or something that could exist; unrealistic scenarios that defy the laws of nature are outside the scope of the definition.

Subjects of resemblance: Content can resemble persons, objects, places, entities or events.

False appearance to be authentic or truthful: Content must falsely appear to a person to be authentic or truthful, regardless of the actual origin. This should be assessed as a whole, taking into account the level of resemblance, the substantive message, the foreseeable deployment context and the intended audience composition and their expectations.

The composition of the audience is decisive in the assessment of whether or not content constitutes a deep fake. Due consideration is especially important if content may potentially be perceived by children, the elderly or persons with low media literacy, since they may be more easily deceived.

Minor technical changes to existing material (for example colour corrections, lighting adjustments) do not render the resulting content to become a deep fake. However, substantial changes that negatively affect the authenticity or truthfulness of content (such as journalistic images) may be relevant.

Examples of deep fakes under the AI Act
- AI-manipulated image of two professional footballers in front of a stadium
- AI-generated audio with voice cloning of podcast presenters
- AI-generated video of a politician holding a speech
- AI-generated promotional video with an AI-generated depiction of an influencer
- AI-generated video featuring a realistic synthetic avatar of a company CEO congratulating employees on their work and the corporate results of the past year
- AI-generated image of a product in advertisement or packaging that can affect the audience’s perception and mislead as to the actual product appearance, characteristics or use (for example making the product appear not identical to the real product, more appealing or with improved quality than in real life)

Examples not constituting deep fakes under the AI Act
- AI-generated image of a sphinx flying over the Eiffel Tower
- AI-generated video of mice discussing cheese
- AI voice replication for fictional characters (for example for audiobooks, games or animation) when there is no deception as to the identity of the narrators
- AI-manipulated radio broadcast adjusting technical audio parameters (for example normalising volume levels, noise reduction, audio compression) without altering the actual words spoken by speakers or their way of speaking
- AI-generated cartoon of a pre-existing image depicting a historical event
- Video of a presenter of a scientific TV programme walking across a mountainous region and addressing the retreat of glaciers which is illustrated by an AI-manipulated background animation that shows how the glaciers retreated from their historical levels to their current levels
- AI-generated fictitious environments (for example fictitious forests and castles) in video games
- Films where real existing actors (not manipulated by AI) are playing against an AI-generated background (for example depicting a fictitious prairie, historical buildings of ancient cities, etc)
- A real product (for example a car) shown in an advertisement against an AI-generated background and surrounding environment as long as the ad is not likely to mislead

6.1.2 Disclosure obligation

Article 50(4) requires deployers of AI systems that generate or manipulate image, audio or video content and thus potentially create deep fakes to disclose that the content has been artificially generated or manipulated.

The AI Act requires clear and distinguishable labelling of this content to disclose its artificial origin. This labelling must be understandable and perceivable by users – for example with visible or audible labels – without the need for any specific technical tools or additional actions.

In light of Article 50(7), deployers may adhere to a code of practice that has been assessed as adequate to demonstrate compliance with the disclosure obligation. Further information on the effects of adherence to such a code of practice and on the supervision of other deployers not adhering to a code can be found in section 8.1 below.

6.1.3 Exceptions for certain content

There is an attenuated transparency obligation for deep fakes that evidently form part of artistic, creative, satirical, fictional or analogous works or programmes. In these cases, deployers only need to disclose that the content has been artificially generated or manipulated in a manner that does not hamper the display or enjoyment of the work.

The following conditions must be met:

Content categories: The exception applies to content falling under at least one of the following categories: artistic, creative, satirical, fictional or analogous works. Whether a deep fake falls under one or more of these categories requires a case-by-case assessment by the deployer.
- Artistic works: works created for the purpose of art (music, cinematographic works, visual arts)
- Creative works: works reflecting the deployer’s creative choices
- Satirical works: works criticising society, politics or public figures using humoristic techniques
- Fictional works: works set in an imaginary setting
- Analogous works: works sharing characteristics with the above categories but not fitting neatly into one
Evident nature: The fact that a deep fake falls within one of the categories must be evident. Content that is potentially ambiguous falls outside the scope of the exception. Relevant factors for assessing the evident nature of artistic, creative, satirical or fictional content include:
- whether the content displays formats or styles that are characteristic of the content categories (for example irony or exaggeration for satirical works, certain art styles, distinctive themes, techniques or aesthetics);
- the context in which the content is presented (for example if the platform, medium or place where the content is presented, made available or distributed is associated with artistic, creative, satirical or fictional use);
- audience expectations (for example a film, gaming environment or virtual reality scene).
Appropriate disclosure: Disclosure of the AI origin or manipulation of the content must not hamper the enjoyment of the work. The manner of disclosure requires a case-by-case assessment, taking into account the work, audience and context. Deployers must ensure that disclosure does not hamper the normal use and enjoyment of the work, while maintaining the utility and quality of the work.
Safeguards for the rights of third parties: The exception must not lead to infringements of the rights of third parties (for example personality rights or copyrights). Compliance with the attenuated transparency obligation and the use of artistic, creative, satirical or fictional deep fakes must be subject to appropriate safeguards for the rights and freedoms of third parties.

Deployers may follow a recognised code of practice in this case as well (see above).

Examples of deep fakes falling within the scope of the exception
- Artistic/fictional work: films (or film trailers) featuring AI-generated or manipulated deep fakes of de-aged, existing actors or digital replicas of dead actors or other persons, available in a cinema or on a streaming platform
- Artistic/creative work: AI-generated music in any kind of genre resembling the individual style of existing artists
- Satirical/fictional work: AI-manipulated image of an existing politician placing him in a scene clearly meant to criticise in a humorous manner certain policy decisions taken or supported by that person
- Analogous creative/fictional work: AI-generated gaming imagery involving deep fake simulations of real, existing persons

Examples of deep fakes falling outside the scope of the exception
- AI-manipulated video involving deep fake simulation of humans advertising a product in an AI-generated scene depicting the functional use of the product by the simulated consumers with the aim of persuading viewers to buy the product
- AI-generated image of celebrities implying their involvement in activities, lacking any fictional, satirical or analogous purpose
- AI-manipulated video featuring a realistic synthetic influencer testing out a sponsored real product, focused solely on displaying the product functionalities
- AI-generated video depicting realistic Holocaust scenes being shared on publicly available social media platforms

6.1.4 Exceptions for law enforcement

If the use of a deep fake is authorised by law to detect, prevent, investigate or prosecute criminal offences, deployers are exempted from the transparency obligation.

6.1.5 Interplay with other Union legal acts

Digital Services Act
The transparency obligations under Article 50(4) first subparagraph of the AI Act are closely related to Article 35(1)(k) of the DSA. Both sets of rules aim to ensure that users are informed about AI-generated or manipulated content.

The DSA requires providers of VLOPs and VLOSEs to take risk mitigation measures such as the clear labelling of “generated or manipulated image, audio or video that appreciably resembles existing persons, objects, places or other entities or events and falsely appears to a person to be authentic or truthful” (Article 35(1)(k) DSA). Providers may also make available an easy-to-use functionality enabling users to indicate such information.

Although the rules appear similar, there are important distinctions:

Scope of application: Article 35(1)(k) of the DSA covers a broader range of “false” content than Article 50(4) first subparagraph of the AI Act, regardless of the technologies used to create the content.
Addressees: Article 50(4) first subparagraph of the AI Act requires deployers of AI systems generating or manipulating deep fakes to label such content. By contrast, Article 35(1)(k) of the DSA applies to providers of VLOPs and VLOSEs disseminating the content or making such content discoverable through search.

Deployers of AI systems integrated into a VLOP or VLOSE are generally considered “recipients of the service” within the meaning of the DSA if they disseminate such content. If providers of VLOPs or VLOSEs make labelling tools available that enable deployers of AI systems to label their deep fake content in compliance with Article 50(4) first subparagraph (that is, a clear and distinguishable disclosure of the AI origin), deployers can use the tools to fulfil their transparency obligations within the context of disseminating the content on the platform.

Furthermore, providers of VLOPs and VLOSEs can also be considered deployers of an AI system within the meaning of the AI Act if they use an AI system under their authority for their own professional purposes (for example to create visuals as part of marketing material).

Data protection
If a deep fake includes information relating to a directly or indirectly identifiable natural person, deployers as controllers within the meaning of EU data protection law must comply with the respective data protection obligations.

These include complying with an appropriate legal basis for the processing of the data and guaranteeing transparency in relation to the data subjects.

Further information about this can be found in the joint guidelines on the interplay between the AI Act and EU data protection law, which are currently being drawn up by the European Commission and the European Data Protection Board (EDPB).

Intellectual property
Deployers of AI systems generating deep fakes must also comply with relevant EU intellectual property laws, such as EU trademark law and copyright law.

The application of a label or icon to a deep fake does not, as such, have any impact on its eligibility for copyright protection, which should be assessed in accordance with applicable Union copyright law.

Personality rights
If deep fakes depict existing persons, deployers must also comply with the relevant legal regimes protecting personality rights over images or voices.

The transparency obligation of Article 50(4) first subparagraph does not mean that AI-generated or manipulated deep fakes that are harmful and unlawful under the applicable Union or national law (for example misleading advertising or criminal law such as child sexual abuse material and non-consensual intimate images) may be generated and disseminated.

6.2 Main components, concepts and transparency obligations for AI-generated or manipulated text

Article 50(4) second subparagraph applies if several conditions are fulfilled:

the system must be an AI system;
the system is used to generate or manipulate text with the purpose of informing the public on matters of public interest;
the use of the AI system does not fall under any of the exceptions.

6.2.1 AI text in the public interest

If AI-generated or manipulated text is published to inform the public on matters of public interest, deployers are required to meet certain transparency obligations. For the purpose of clarifying the scope of Article 50(4) second subparagraph, the following three elements are decisive:

Published text
The AI-generated or manipulated text must be published. This means that it can be read by an indeterminate, sufficiently large group of persons – whether against payment or free of charge. By contrast, text is not considered “published” if access is restricted to specific individuals belonging to a closed, private group or if the group is too small or insignificant. Examples of such text include private correspondence (including for professional purposes) and organisation-internal documents.
Informing the public: The published AI text must aim to inform the public. This means that the text must intend to communicate knowledge, opinions or facts. Short texts that do not materially communicate knowledge, opinions or facts are not considered to inform the public.
Matters of public interest: The text must inform the public on matters that are relevant to society and that merit public debate. These include in particular:
- politics and democratic processes,
- public administration and services,
- administration of justice and law enforcement,
- protection of fundamental rights,
- public security,
- environmental protection and consumer safety,
- economic, financial, political, scientific or cultural development.

Matters that are considered to be of public interest can evolve over time and across contexts.

Examples of AI text in the public interest (requiring labelling unless subject to human review/editorial control, see section 6.2.3)
- AI-generated summary of a human-authored newspaper article discussing a recent decision by a town council
- AI-manipulated parts of a lifestyle-website article comparing the effects of various diets on a particular disease in middle-aged women
- AI-manipulated corporate reports published on a listed company’s website containing investor information
- AI-generated message on a meteorological institute’s social media profile warning citizens about stormy weather and related precautionary measures

Examples of AI text falling outside the scope of Article 50(4)
- AI-generated fantasy novels
- AI-manipulated text that is part of a company’s advertisement or product descriptions (not including any claims related to health, consumer safety or sustainability)
- News summary generated by a chatbot that is only available to the user that prompted the chatbot

6.2.2 Disclosure obligation

As for deep fakes, labelling and disclosure methods (including disclaimers) in accordance with Article 50(4) second subparagraph must be clear and perceivable by natural persons. This requires visible or audible measures that do not involve the need for the persons to rely on any specific technical tools or perform dedicated actions.

Deployers may follow a recognised code of practice in this case as well (see above).

6.2.3 Exceptions for text

There are exceptions to the labelling obligation for AI-generated or manipulated text if two cumulative conditions are met:

Human review or editorial control: The text must have been subject to careful examination by one or more competent persons. This examination comprises examination of the substance of the content and goes beyond a solely formal check (for example spellchecking). Editorial control refers to the authority to alter or reject the substance of the text. Fact-checking and ensuring the trustworthiness of the source are minimum requirements, without prejudice to existing review and editorial procedures and professional standards applicable to media service providers. Subsequent modifications render the review and control invalid.
Editorial responsibility: A natural or legal person must hold editorial responsibility for the publication. This means overall legal responsibility for the content, including the review or control. The identity and contact details of the person with editorial responsibility should be publicly available (for example in a website’s terms and conditions or other user-facing legal information).

This applies in particular to deployers that are not media service providers.

Examples of text meeting the conditions and for which disclosure is not required under Article 50(4) second subparagraph
- AI-manipulated article in a newspaper or AI-generated summary of an article that has been subject to editorial control of the respective editor-in-chief with editorial responsibility held by the legal person that publishes the newspaper
- AI-manipulated academic blog which has undergone internal peer review and where the entity managing the blog (for example a research centre) holds editorial responsibility
- AI-generated public safety warnings approved by a public official whose publication is under the responsibility of the relevant public agency for civil protection
- AI-generated sustainability reports published on a listed company’s website having undergone human review by professionals in relevant functions
- AI-supported translation of a human-written article whereby the translation has undergone human review

Examples of text not meeting the conditions and for which disclosure is required under Article 50(4) second subparagraph
- AI-generated articles posted on a website without any human review of the content
- AI-generated articles that are reviewed by another AI system and where a human editor performs a mere superficial, grammatical check
- AI-generated, self-published book made available on an e-commerce platform that has not undergone any review by a natural or legal person

6.2.4 Exception for law enforcement

If the use of AI systems for generation of text published with the purpose of informing the public on matters of public interest without human review (for example an AI-generated police warning regarding a terrorist threat published on social media) is authorised by law to detect, prevent, investigate or prosecute criminal offences, deployers are exempted from the transparency obligation under Article 50(4). The clarifications provided in Section 3.2.2 of the Guidelines are also relevant for this provision.

6.2.5 Interplay with other Union legal acts

Editorial responsibility is a crucial concept in EU media law. Article 2(8) of the European Media Freedom Act (Regulation (EU) 2024/1083) defines “editorial responsibility” as:

“the exercise of effective control both over the selection of programmes or press publications and over their organisation, for the purposes of the provision of a media service, regardless of the existence of liability under national law for the service provided”.

In certain situations, this definition may overlap with the notion of editorial responsibility used in Article 50(4) second subparagraph, in particular where the deployers of the AI system also qualify as media service providers under the EMFA. Editorial responsibility within the meaning of Article 50(4) should be interpreted in line with existing EU media law. However, it is a distinct concept and may also apply to other deployers assuming editorial responsibility for the publication.

Media service providers may rely on their existing editorial processes and standards to benefit from the exception under Article 50(4) second subparagraph.

7. Transparency at the first interaction and accessibility

Article 50(5) requires the information required under Article 50 (1) to (4) to be provided to the natural persons concerned in a clear and accessible manner by the time of the first interaction.

Article 50(5) – horizontal requirements (summary)

Timing
- Information must be provided at the latest at the time of the first interaction with the system or content
- The information can be provided earlier; if content may not be perceived from the beginning (for example live broadcasts), disclosure must be repeated
Audience
- Direct interaction partners (chatbots, voice assistants)
- Active and passive users seeing or hearing the AI-generated/manipulated content
- Persons subject to the use of emotion recognition or biometric categorisation systems
Scope of information
- Clear and distinguishable: easily noticeable and understandable, easy to identify as separate from other content
- Accessible: must be accessible to all audiences, in particular children (child-friendly, age-appropriate, succinct, in the official language, engaging, possibly given gradually over time)
Placement
- Directly during the interaction or as part of the content
- Hidden information in manuals or under menu layers not sufficient
Legal framework for accessibility
- Reference to existing EU legal acts:
  - Directive (EU)2016/2102 – accessibility of websites and apps of public sector bodies
  - Directive (EU)2019/882 – accessibility requirements for products and services
- Actors must assess if their offering falls within the scope of these Directives and provide information accordingly
- Article 50 does not impose distinct or additional accessibility requirements

[ENDE]

7.1 Information in a clear and easily understandable manner

Information is considered to be provided in a clear and distinguishable manner if it is easily noticeable and understandable and easy to identify as separate from other content. It must be accessible to all audiences, including children and persons with disabilities.

Information for children must be:

child-friendly, age-appropriate and easy-to-understand;
presented clearly and succinctly;
presented in a way that is easy to review and intuitively accessible;
presented in the official language of the country;
engaging for children (for example with graphics, videos or characters);
given to children gradually and over time.

Providers and deployers can present the information during interaction (in accordance with Article 50(1)) or as part of the relevant content (in accordance with Article 50(2) and (4)). Information is not considered to be provided in a clear and distinguishable manner if it is only included as part of a manual or hidden under layers of menu options.

7.2 First interaction or exposure

Article 50(5) requires the information to be provided to the natural persons concerned at the latest at the time of the first interaction with or exposure to the system or content. This means:

first interaction, for example:
- when launching a conversation with a chatbot,
- when starting physical interaction with an AI system;
first exposure, for example:
- at the start of a video featuring a deep fake,
- at the start of an AI-generated or manipulated text publication on matters of public interest,
- when encountering AI-manipulated deep fakes when scrolling on social media.

The information must always be provided when a person interacts with an AI system or is exposed to content for the first time. The information can also be provided earlier to increase effectiveness. If persons may not perceive the content from its beginning (for example live broadcasts), disclosure at the beginning only is not adequate. In such cases, the information should also be provided at later moments.

7.3 Compliance with applicable accessibility requirements

Providers and deployers of AI systems subject to transparency obligations under Article 50 must be aware that they may have to comply with accessibility requirements. These requirements are laid down in Directive (EU) 2016/2102 on the accessibility of the websites and mobile applications of public sector bodies and Directive (EU) 2019/882 on the accessibility requirements for products and services. Providers and deployers must assess whether their product, service or content falls within the scope of these Directives. If so, they must ensure the accessibility of the information provided in line with Article 50. No additional accessibility requirements apply.

8. Enforcement of the provisions

This section clarifies: (1) the effects of adhering to a recognised code of practice as regards compliance with the transparency obligations for AI-generated content, (2) the role of the competent supervisory authorities in the enforcement of Article 50, and (3) the applicable penalties for infringements of those obligations. It also addresses (4) the entry into application.

In brief

Proof: Conformity through recognised code of practice or comparable proof (for example gap analysis)
Market surveillance:
- National market surveillance authorities (Regulation (EU) 2019/1020)
- AI Office for GPAI models and more
- European Data Protection Supervisor (EDPS) for EU institutions
- Complaints at any time to the Bundesnetzagentur’s central point for complaints relating to the AI Act
Penalties: Administrative fines of up to €15mn or 3% of total worldwide annual turnover for the preceding year (whichever is higher), and for SMEs and start-ups whichever is lower; administrative fines of up to €750,000 for EU institutions
Timetable:
- Full application as from 2 August 2026
- Transitional period up to 2 December 2026 for the marking/detection of generative AI systems introduced before 2 August 2026
- Content from before this date does not need to be marked or labelled retroactively but voluntary labelling is encouraged

[ENDE]

8.1 Effects of adhering to a code of practice assessed as adequate

Providers and deployers of AI systems falling within the scope of Article 50(2) and (4) may demonstrate compliance with their transparency obligations by adhering to a recognised code of practice or through adequate alternative means. Such a Code complements the AI Act and the Guidelines as an EU-wide framework for providers and deployers to demonstrate compliance with the obligations, regardless of their place of establishment, operation and competent authorities. Adherence to the Code of Practice facilitates demonstrating compliance and can be taken into account as a mitigating factor when the amount of a fine is fixed.

For signatories to the Code of Practice, the European Commission and the competent market surveillance authorities will focus their supervisory activities on assessing adherence to the Code of Practice. Signatories will benefit from increased trust from the authorities and the general public through transparency. Deviations from the Code of Practice will, however, lead to a loss of these benefits.

Providers and deployers that are not signatories to the Code of Practice must demonstrate their compliance through other means, such as gap analyses. They may be subject to a larger number of requests for information and requests for access to assess the effectiveness and reliability of their technical solutions.

8.2 Market surveillance authorities

Market surveillance authorities can take enforcement actions on their own initiative or following a complaint.

Further information about competent market surveillance authorities within the meaning of the AI Act can be found here.

Any data subject or other natural or legal person having grounds to consider that there has been an infringement has the right to submit a complaint (Article 85).

The Bundesnetzagentur’s central point for complaints relating to the AI Act can be accessed here.

The responsibilities of the market surveillance authorities under the AI Act do not affect the powers and tasks of other supervisory authorities or fundamental rights authorities where both legal frameworks apply (for example data protection, consumer protection).

8.3. Penalties

Providers and deployers that do not comply with the transparency obligations laid down in Article 50 may be issued the following administrative fines:

up to €15mn or, if the offender is an undertaking, up to 3% of its total worldwide annual turnover for the preceding financial year, whichever is higher,
in the case of EU institutions, bodies and agencies, up to €750,000,
in the case of SMEs and start-ups, up to €15mn or up to 3% of the total worldwide annual turnover for the preceding financial year, whichever is lower.

The fines should be effective, proportionate and dissuasive and take into account economic viability. When deciding whether to impose an administrative fine and when deciding on the amount of the fine, the competent authorities assess the following circumstances:

the nature, severity and duration of the infringement,
the intentional or negligent character of the infringement,
other circumstances,
the degree of cooperation with market surveillance authorities.

8.4 Entry into application

Article 50 applies as from 2 August 2026. This means that all in-scope AI systems placed on the market or put into service in the EU must be compliant with the requirements, regardless of their date of placement on the market or putting into service.

Article 1(39)(b) of the Digital Omnibus on AI provides a transitional period for generative AI systems placed on the market or put into service before 2 August 2026. This only relates to the marking and detection obligations under Article 50(2).

AI-generated or manipulated outputs (including deep fakes) that have been generated or manipulated and published before 2 August 2026 do not need to be marked or labelled retroactively.

The labelling obligation does not apply to texts on matters of public interest that have been AI-generated or manipulated before 2 August 2026. However, if such texts are published on or after 2 August 2026, they need to be labelled.

Deployers of AI systems and other actors who are in the possession of or disseminate such content are encouraged to label the content voluntarily, but without any disproportionate efforts. This serves the objective of transparency and increased trust and integrity of the information ecosystem.

9. Review and update of the European Commission’s Guidelines

The Guidelines constitute a first interpretation with practical examples of the transparency obligations laid down in Article 50. The European Commission will review these on a regular basis in view of:

practical experience,
technological, societal and regulatory developments,
experience from market surveillance actions,
interpretations on Article 50 given by the Court of Justice of the European Union (CJEU).

During such a review, the European Commission may withdraw or amend the Guidelines. Providers, deployers and national market surveillance authorities are encouraged to contribute through various bodies and organisations, in particular by responding to future calls for public consultation.