In addition, Article 50(5) AI Act mandates that the information referred to in its paragraphs 1 to 4 shall be provided to the natural persons concerned in a clear and distinguishable manner at the latest at the time of the first interaction or exposure and conform to the applicable accessibility requirements.

The various transparency obligations laid down in Article 50 of the AI Act may apply cumulatively to (the output of) a single AI system, possibly engaging the responsibility of different actors (providers or deployers).

Responsible actors under Article 50 of the AI Act

1. Providers are natural or legal persons, public authorities, agencies or other bodies that develop AI systems, or have them developed, and place them on the Union market or put them into service under their own name or trademark, whether for payment or free of charge (Article 3(3) AI Act). Providers must comply with the relevant transparency obligations in Article 50(1) (interactive AI systems) and (2) (generative AI systems) applicable to them, irrespective of whether they are established or located within the Union or in a third country (Articles 2(1)(a), 3(3), (9) and (11) of the  AI Act). Providers established or located outside the Union are also subject to the AI Act if the output of their AI system is used in the Union (Article 2(1)(c) of the AI Act) or, as recital 22 explains, “to the extent the output produced by those systems is intended to be used in the Union“. Obviously, due to the need to specify further, the Guidelines elaborate that “incidental, unforeseen or unauthorized downstream use should not alone trigger the application of the obligations to such third country providers of interactive or generative AI systems”.

2. Deployers are natural or legal persons, public authorities, agencies or other bodies using AI systems under their authority, unless the use is for a personal non-professional activity (Article 3(4) of the AI Act). The transparency obligations from Article 50(3) and (4) of the AI Act apply to deployers if their place of establishment or location is within the Union, or if they are established or located in a third country where the output of the AI system is used in the Union (Article 2(1)(b) and (c) of the AI Act). Regarding entities located or established outside the Union, the Guidelines clarify further that „the transparency obligations apply where the deployer itself foresees dissemination and use of the AI outputs in the Union“ (i.e. by directing or authorising distribution within the Union, including by posting deep fakes on the globally accessible internet). However, third country deployers are not bound by the transparency obligations where the content of the AI system reaches audiences in the Union through channels that are unforeseeable and outside their control.

The Guidelines explain that the concept of ‘authority’ over an AI system, used to define “the deployer’’, should be understood as assuming responsibility over the decision to deploy the system and over the manner of the actual use of the system (including its outputs). It does not necessarily require technical control over the operation of the AI system, so long as the deployer takes the decision for what purposes and how to use the AI system (including in decentralised workflows and group corporate structures).

Deployers involved in complex content production and distribution value chains should take proportionate measures to ensure that the labelling of the content they have implemented pursuant to Article 50(4) AI Act is displayed in a clear and distinguishable manner for the targeted and foreseeable audience at the point of first exposure in accordance with Article 50(5) AI Act (e.g., via contractual conditions with distributing partners, user experience (UX) settings and interfaces to be displayed).

Where the deployer of an AI system is a legal person under whose authority the system is used (e.g. an advertising company), the individual employees that act under the instructions and under the control of that legal person (e.g. digital animators, web designers, content creators, journalists) should not be considered as separate deployers for that system. A legal person remains a deployer even if it involves third parties (e.g. contractors, freelancers) in the operation of the system on its behalf and under its responsibility and control.

3. By contrast, other actors (such as providers of hosting services, including online platforms, or broadcasters) whose role is limited to disseminating or transmitting AI-generated or manipulated content created by third parties, or who receive or are exposed to AI-generated or manipulated content without having authority over the use of the AI system for the purpose of the AI content generation or manipulation, are not deployers within the meaning of the AI Act. Without being responsible and liable under the AI Act, those actors may nevertheless play an important role in the value chain to preserve the trust and integrity of the information ecosystem. They are, therefore, strongly encouraged to preserve the marking and labelling implemented pursuant to Article 50 AI Act and implement proportionate and technically feasible measures to enable natural persons to detect the AI-generated or manipulated content. This status under the AI Act is without prejudice to the responsibility of providers and deployers of generative AI systems or any obligations or liability under Regulation (EU) 2022/2065 (the Digital Services Act or ‘DSA’) or other applicable Union or national law.

AI agents

Article 50(1) AI Act is addressed to providers of AI systems directly interacting with natural persons, who must design and develop their systems in such a way that the natural persons concerned are informed that they are interacting with an AI system. As explained in Recital 132 of the AI Act, the purpose of this information obligation is to enable those natural persons to take informed decisions regarding the system’s outputs, to avoid overreliance on such systems, and to allow those natural persons to calibrate their trust in the content and the interactions accordingly.

In addition to a comprehensive explanation of the main components, concepts and related transparency obligations of Article 50 (1), as well as examples that present (or not) the obviousness exceptions, the Guidance address the AI agent concept within the framework of this Article.

The AI Act does not mention AI agents. Under the Guidelines, AI agents are covered by Article 50(1) of the AI Act if they are capable of interacting with the persons instructing them or with other natural persons in the execution of the tasks (e.g. making bookings, managing correspondence, negotiating or concluding contracts, executing purchases, etc). AI agents must be designed and developed in such a way that they disclose both their artificial nature and the person on whose behalf they are acting, considering the need for transparency of the origin and the delegation of authority and accountability for the consequences of their actions. This includes cases where AI agents are capable of interacting in complex multi-agent architectures with other AI agents that directly interact with the natural persons in the execution of the task.

Nevertheless, it may not always be possible for the provider of the AI agent to identify in advance individual instances of direct interaction of an AI agent with other natural persons once the AI agent is deployed in different contexts. Where the provider cannot reliably determine whether the AI agent will directly interact with a natural person before placing the AI agent on the market or putting it into service, the agent should be designed at the architecture level, and instructed, to disclose itself as such in every situation where it is reasonably likely that the agent may interact with a natural person (including where a natural person may represent or act for a legal entity).

Furthermore, AI agents should also disclose themselves to the persons instructing them at key steps (e.g. at the point of authorisation, reporting, validation etc., including when the deployed AI agent receives, processes, or relies upon outputs generated by other AI systems rather than directly by a natural person) and at every new interaction.

At the same time, Article 50(2) of the AI Act, which prescribes the obligation of marking and detection of AI systems’ outputs, applies to AI systems that may serve multiple purposes, including agentic AI systems, so long as they generate synthetic audio, image, video or text content. However, Article 50(2) does not apply if outputs of an AI system are intended to be exclusively communicated machine-to-machine and processed automatically without any exposure to humans, which includes agent-to-agent communication.

Marking and detection

Article 50(2) AI Act requires providers of AI systems generating synthetic content to implement technical solutions that meet certain quality requirements for machine-readable marking and detection of their AI systems’ outputs. The objective is to enable natural persons to distinguish AI-generated or manipulated content from other content (for example, content created by humans) and to verify its origin, thus also contributing to increased integrity and trust in the information ecosystem.

Within the exhaustive explanation of the main components and concepts of Article 50(2) of the AI Act, the Guidelines clarify that the transparency obligations apply broadly to both AI-generated and AI-manipulated content, including content that combines human-created and AI-generated elements.

Content generation refers to the generation of synthetic material by the AI system (e.g., based on a human prompt), such as generating a synthetic image or a song. At the same time, Article 50(2) AI Act also applies to AI systems that may serve multiple purposes, including GPAI systems, and can produce a variety of types of content, as well as agentic AI systems, so long as they generate synthetic audio, image, video or text content.

By contrast, content manipulation refers to already existing content (whether synthetic or not) that is altered by an AI system (e.g., an existing image or a voice recording that is manipulated by an AI system in accordance with human instructions).

For content to fall under Article 50(2) of the AI Act, it is sufficient that it is AI-generated or manipulated beyond standard editing under the exceptions set out in this Article.

Article 50(2) of the AI Act specifies the modalities of synthetic content subject to the transparency obligation: audio, image, video, or text. This also includes multimodal content, that is, content made up of a mix of these modalities. The Guidelines further list examples of synthetic content falling within and falling outside the scope of Article 50(2) of the AI Act.

Also, Article 50(2) AI Act provides three explicit exceptions from the transparency obligations:

• to the extent the AI systems perform an assistive function for standard editing. Standard editing should be understood as the process of preparing existing content for publication or distribution (e.g., small edits to improve readability and grammar, quality and format) and does not involve generating new content;
• AI systems that do not substantially alter the input data provided by the deployer or the semantics thereof;
• generative AI system authorised by law to detect, prevent, investigate or prosecute criminal offence, 

while the Guidelines list examples of standard editing and minor alterations benefitting from these exceptions.

Labelling deep fake

Article 50(4) AI Act lays down two separate obligations for deployers of generative AI systems requiring disclosure of (i) deep fakes and (ii) AI-generated or manipulated text published with the purpose of informing the public on matters of public interest, except in defined cases. These labelling obligations apply in addition and without prejudice to the obligations for machine-readable marking and detection under Article 50(2) AI Act applicable to providers of generative AI systems

Within a comprehensive explanation of the main components, concepts, and related transparency obligations under Article 50(4) of the AI Act, the Guidelines clarify in detail the notion of ‘deep fake’. Article 3(60) AI Act defines ‘deep fakes’ 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 the notion of deep fake, four cumulative criteria from that definition are examined.

Resemblance: Article 3(60) AI Act specifies that AI-generated or manipulated image, audio or video content must ‘resemble’ existing subjects to constitute a deep fake. Recital 134 AI Act adds that such resemblance should be ‘appreciable’. AI-generated or manipulated image, audio or video content appreciably resembles a subject if there is a high level of similarity between the deep fake content and the subject (including any of its recognisable elements) being simulated by the deep fake. The content does not need to be identical to the subject. Whether the level of resemblance is appreciable is a case-by-case assessment to be done by the deployer based on an objective comparison between the simulated subject and the deep fake content, based on, among others, the extent to which characteristic or distinctive features are represented by the deep fake.

Existing: Article 3(60) AI Act specifies that AI-generated or manipulated image, audio or video content must resemble ‘existing’ subject(s) to constitute a deep fake. To effectively fulfil the purpose of Article 50(4) AI Act, the characteristic of existence implies that the AI-generated or manipulated image, audio or video content should resemble realistic subjects (e.g., persons, objects, places). Therefore, it is sufficient for simulated persons, objects, places, entities or events to resemble someone or something that exists, can plausibly exist or could have plausibly existed in reality to be considered a deep fake, since such content can be a major source of deception, manipulation and disinformation. By contrast, simulated persons, objects, places, entities or events that, for example, defy the laws of nature or physics or depict lifeforms that are not commonly accepted in biology (such as e.g. humans flying without mechanical aids, dragons, or elephants driving cars) and have no potential to mislead, are considered unrealistic and therefore fall outside the scope of the transparency obligation.

Persons, objects, places, entities or events: Article 3(60) AI Act specifies that AI-generated or manipulated image, audio or video content should resemble existing ‘persons, objects, places, entities or events’ to constitute a deep fake. ‘Persons’ is to be understood as realistic, human beings (including digital replicas of real persons, realistic AI-generated human avatars or personas, and personal characteristics or expressions, such as image, voice, behaviour, performances etc.). ‘Objects’ is to be understood as realistic, inanimate material items, including buildings, artworks, machinery, consumer goods etc. ‘Places’ is to be understood as realistic locations.‘Entities’ is to be understood as realistic, non-human but animate beings including animals or other biological lifeforms.‘Events’ is to be understood as realistic scenes or situations that can involve persons, objects, places and entities (e.g. evoking historical events or the depiction of professional or consumer services).

False appearance to a person to be authentic or truthful: Fourth, Article 3(60) AI Act specifies that the AI-generated or manipulated image, audio or video content resembling existing persons, objects, places, entities or events ’would falsely appear to a person to be authentic or truthful’. This criterion relates to the essential characteristic of deep fake content and its capacity to potentially deceive or mislead a person regarding the content’s authenticity or truthfulness. Content authenticity refers to whether the content is genuinely what it purports to be in terms of its source or creation process (including e.g. the involvement of real human beings or animals and their actual behaviour or actions, the actual appearance or use of objects or the delivery of services, the accurate unfolding or course of an event). Truthfulness pertains to the veracity of the content (e.g. factual correctness of the representations in the deep fake).

Whether content ‘falsely appears to a person to be authentic or truthful’ should be assessed as a whole (holistic), taking into account the level of resemblance, the potential substantive message of the content, the intended and foreseeable deployment contexts, the environment in which the content is presented, and the intended and reasonably foreseeable audience composition and their expectations. However, this assessment is objective and does not require the intention of the deployer to deceive or mislead the natural persons exposed to the content for it to constitute a deepfake. As regards the deployment context and audience expectations, this means that if the content is used in specific deployment contexts where the intended and reasonably foreseeable audience does not expect content to be authentic or truthful, the AI-generated or manipulated content may not falsely appear to be authentic or truthful even if it is, in fact, non-authentic or untruthful.

For example, in standard movie production, the AI-generation or manipulation of background scenes, special effects, or technical pre- and post-processing are not likely to make content falsely appear to the audience to be authentic or truthful.

Conversely, in advertisement or packaging, an AI-generated image of a product can affect the audience’s perception and mislead as to the actual product appearance, characteristics or use (e.g. making the product appear not identical to the real product, more appealing or with improved quality than in real life).

To help understand differences in praxis, the Guidelines list examples of deep fakes and examples that do not constitute deep fakes under Article 3(60) of the AI Act.

Transparency of artistic, creative, satirical, fictional or analogous deep fake content

Article 50(4), first subparagraph, AI Act foresees an attenuated transparency obligation for deep fakes forming part of evidently artistic, creative, satirical, fictional or analogous works or programmes. For such content, the transparency obligation is limited to the disclosure of the deep fake in an appropriate manner that does not hamper the display or enjoyment of the work.

The fact that a deep fake falls within one (or more) of the above content categories should be evident to the natural persons exposed to it. This implies that the categories should be interpreted strictly, justified by the fact that a lighter disclosure regime applies due to the potential impact on the freedom of expression and the freedom of the arts and sciences, while risks of misinformation, manipulation and deception still need to be addressed adequately (format or style, context in which the content is presented, audience expectations). In addition to requirements related to format or style, context in which the content is presented, and audience expectations, content is excluded from the scope of that category if its nature is exclusively informative or commercial and is recognisable as such (e.g. news reporting). In this respect, some kinds of content (e.g. advertisements or documentaries) containing deep fakes might be regarded as evidently creative or fictional works in certain, specific situations, but not in others, since the assessment is case-specific. When the deep fake combines multiple characters (e.g. informative and creative), the informative character should always prevail and require compliance with the standard labelling requirements.

Enforcement of Article 50 of the AI Act

Market surveillance authorities designated by the Member States, the AI Office[1] , and the European Data Protection Supervisor[2] are responsible for supervising and enforcing the rules for AI systems falling within their competence, including the transparency obligations laid down in Article 50 AI Act. Such enforcement takes place within the system of market surveillance and compliance of products established by Regulation (EU) 2019/1020 and the AI Act. The enforcement powers of market surveillance authorities in relation to AI systems are laid down in those regulations. Those authorities can take enforcement actions in relation to the obligations listed in Article 50 AI Act on their own initiative or following a complaint, which every affected person or any other natural or legal person having grounds to consider such violations has the right to lodge (Article 85 of the AI Act).

Penalties

Providers and deployers that do not comply with the applicable transparency obligations laid down in Article 50 AI Act may be fined up to EUR 15 000 000 or, if the offender is an undertaking, up to 3% of its total worldwide annual turnover for the preceding financial year, whichever is higher. EU institutions, bodies and agencies that violate the transparency obligations may be subject to administrative fines of up to EUR 750 000.

In the case of small and medium-sized enterprises (SMEs), including start-ups, each fine shall be up to the above percentages or amount, whichever is lower. The fines shall be effective, proportionate and dissuasive and take into account the interests of SMEs and small mid-cap enterprises (SMCs), including start-ups, and their economic viability.

[1] As the market surveillance authority for AI systems built on GPAI models provided by the same provider (Article 75(1) AI Act). This covers the obligation applicable to providers under Article 50(1) and (2) AI Act. Regarding Article 50(3) and (4) AI Act, the AI Office will only be competent if the respective GPAI model and system provider, is also a deployer of the system.
[2] As the market surveillance authority for the EU institutions, agencies and bodies.