{"id":4139,"date":"2026-08-05T05:30:00","date_gmt":"2026-08-05T05:30:00","guid":{"rendered":"https:\/\/naaia.ai\/?p=4139"},"modified":"2026-08-05T12:00:22","modified_gmt":"2026-08-05T12:00:22","slug":"ai-act-article-50-transparency-obligations-guidelines","status":"publish","type":"post","link":"https:\/\/naaia.ai\/en\/ai-act-article-50-transparency-obligations-guidelines\/","title":{"rendered":"Article 50 of the AI Act: the transparency obligations every organization deploying AI must anticipate"},"content":{"rendered":"\n

From 2 August 2026, most of the transparency obligations set out in Article 50 of the AI Act<\/a> will become applicable. Providers and deployers of AI systems will be required to implement appropriate transparency measures. In particular, they must clearly, prominently and accessibly inform natural persons when they are interacting with an AI system, when they are exposed to emotion recognition or biometric categorization systems, or when they access content that has been artificially generated or manipulated. They must also implement the technical solutions necessary to ensure that such synthetic content can be identified. <\/p>\n\n\n\n

One exception applies to Article 50(2), which concerns the marking and detection of AI-generated or AI-manipulated content. For generative AI systems placed on the market before 2 August 2026, the application of these requirements has been postponed until 2 December 2026. <\/p>\n\n\n\n

For organizations using chatbots, AI agents, generative AI tools or deepfake creation technologies, the priority is now to identify the relevant use cases and demonstrate compliance with the applicable requirements. <\/p>\n\n\n\n

In this regard, the European Commission’s recent Guidelines on Transparency Obligations of AI-Generated Content<\/a> and the Code of Practice on Transparency of AI-Generated Content<\/a> provide essential clarification on the scope of these obligations, the actors responsible for compliance, and the measures expected to meet regulatory requirements. <\/p>\n\n\n\n

In this article, we examine the transparency obligations established under Article 50, the AI systems concerned, the respective responsibilities of providers and deployers, and the practical measures organizations should implement to strengthen their AI compliance framework.  <\/p>\n\n\n\n

Article 50(1) – AI systems that interact directly with natural persons<\/strong> <\/h2>\n\n\n\n

1. Defining interaction: a\u00a0bidirectional<\/strong> exchange with an AI system<\/strong>\u00a0<\/h3>\n\n\n\n

Article 50(1) applies to AI systems intended to interact directly with natural persons. Direct interaction implies a\u00a0bidirectional exchange between a natural person and an AI system. This means that the user must be able to provide an input to the system, and the system must be capable of responding with a contextually relevant output. The interaction may take any form that is perceptible and understandable to a human and may occur either as a single exchange or over an extended\u00a0period of time.\u00a0<\/p>\n\n\n\n

Conversely, the mere passive collection of data, or indirect exposure to an AI-generated output without any direct exchange with the system, falls outside the scope of this obligation. The key criterion is therefore not the presence of AI operating in the background, but rather the existence of an interactive relationship between the natural person and the AI system. <\/p>\n\n\n\n

2. Defining direct interaction: real-time or near real-time exchanges without human intermediaries<\/strong> <\/h3>\n\n\n\n

Direct interaction presupposes engagement in real time or near real time, including situations where an AI system drafts and sends messages to natural persons or interacts with a physical or virtual environment that is perceptible to them. By contrast, indirect or mediated interactions, in which individuals are merely exposed to AI-generated outputs without directly engaging with the AI system itself, are excluded. <\/p>\n\n\n\n

This distinction is particularly important for chatbots, conversational agents, virtual assistants, AI avatars and autonomous AI agents. For these AI systems, the relevant question is not simply whether the AI generates an output, but whether the user could reasonably believe that they are interacting with a human rather than with an automated system. <\/p>\n\n\n\n

3. What information must be provided?<\/strong> <\/h3>\n\n\n\n

Providers must design and develop AI systems in such a way that affected persons are informed that they are interacting with an AI system. This information must be provided no later than the first interaction and, in some cases, may need to be maintained or reiterated on an ongoing basis. The Guidelines recommend various means of disclosure, including explicit messages, banners, visible badges, voice announcements, icons and graphical labels. <\/p>\n\n\n\n

By contrast, certain methods are considered insufficient when used in isolation, such as references contained solely in terms and conditions, metadata, ambiguous descriptions such as “assistant”, generic statements such as “this service uses AI”, or purely technical wording. <\/p>\n\n\n\n

4. The specific case of AI agents<\/strong> <\/h3>\n\n\n\n

AI agents must disclose both their artificial nature and the identity of the principal on whose behalf they act. This disclosure should be repeated at the beginning of each new interaction and at certain key stages, including validation, authorization and reporting processes. AI agents likely to interact with natural persons should therefore be designed to automatically identify themselves as AI systems. <\/p>\n\n\n\n

Article 50(2) – AI-generated and AI-manipulated synthetic content<\/strong> <\/h2>\n\n\n\n

1. The content concerned: text, images, audio, video, including multimodal content<\/strong> <\/h3>\n\n\n\n

Article 50(2) applies to AI systems that generate or manipulate synthetic content, whether in the form of text, images, audio, or video, including content that is wholly or partially generated or modified by AI. The obligation notably covers content generated by general-purpose AI systems and AI agents. <\/p>\n\n\n\n

The key criterion is whether the content is perceptible to natural persons. The obligation applies where AI-generated or AI-manipulated content is intended to be seen, read, or heard by users. <\/p>\n\n\n\n

By contrast, systems that merely recommend or display existing content fall outside the scope of the requirement, as do non-perceptible or purely technical outputs such as source code, APIs, JSON, or SQL. Machine-to-machine communications, unprocessed data, and certain internal uses that do not involve external exposure are likewise excluded. Only the final AI-generated or AI-manipulated content must be subject to marking and detection measures. <\/p>\n\n\n\n

2. A dual cumulative obligation: marking and detection<\/strong> <\/h3>\n\n\n\n

Providers must ensure that outputs are marked in a machine-readable format and can be detected as having been generated or manipulated by AI. This requirement obliges providers to implement two complementary technical measures: a machine-readable marker, for example through watermarking, metadata, or technical fingerprints, and a detection capability enabling the artificial origin of the content to be effectively identified. <\/p>\n\n\n\n

These measures must be effective, reliable, robust, and interoperable. Compliance must be assessed in light of the state of the art and technical feasibility. To date, no single solution fully satisfies all of these requirements, which means that organizations are generally encouraged to combine multiple marking and detection techniques. <\/p>\n\n\n\n

3. Responsibility remains with the provider<\/strong> <\/h3>\n\n\n\n

Marking measures may be implemented at different levels, including at the model level, the system level, or through a third-party solution. However, responsibility for compliance remains with the provider. This clarification is particularly important for value chains involving general-purpose AI models, system integrators, software vendors, and deploying organizations. <\/p>\n\n\n\n

Article 50(3) – Emotion recognition and biometric categorisation systems<\/strong><\/h2>\n\n\n\n

1. An obligation borne by the deployer<\/strong> <\/h3>\n\n\n\n

For emotion recognition and biometric categorisation systems, the transparency obligation rests with the deployer. Deployers must inform natural persons who are exposed to the operation of the system. This information must be provided to all exposed individuals, including children, and no later than the time of their first exposure to the system, although prior notification is encouraged wherever feasible. The means of providing such information may vary depending on the deployment environment, the intended audience, and whether a pre-existing relationship exists with the individuals concerned (for example, through an established communication channel). <\/p>\n\n\n\n

2. The need to align with data protection requirements<\/strong> <\/h3>\n\n\n\n

This transparency requirement must be considered alongside obligations arising under European Union data protection law. In particular, the notifications required under Article 50(3) should be incorporated into the privacy information provided to data subjects. <\/p>\n\n\n\n

In practice, this means that AI Act transparency obligations should not be treated as a standalone disclosure requirement. Rather, they should be integrated into a coherent information framework, including privacy notices, website disclosures, interface notifications, HR documentation, and other channels through which affected individuals are informed. <\/p>\n\n\n\n

The concepts of emotion recognition systems and biometric categorisation systems are further clarified in the European Commission’s guidelines on the classification of high-risk AI systems. The explanations and examples provided in those guidelines are also relevant for the interpretation and application of Article 50(3). <\/p>\n\n\n\n

Article 50(4) – Deepfakes and AI-generated text on matters of public interest<\/strong> <\/h2>\n\n\n\n

1. The concept of a deepfake: a false appearance of authenticity or truthfulness<\/strong> <\/h3>\n\n\n\n

The Guidelines further clarify the definition of a deepfake through several key elements. The content must bear a sufficiently convincing resemblance to real or plausible persons, objects, places, entities, or events, whether they currently exist or could reasonably exist in reality. In addition, it must be capable of appearing falsely authentic or truthful to a person, taking into account the reasonably foreseeable composition of the audience, including children, older persons, and individuals with limited digital or AI literacy. <\/p>\n\n\n\n

Minor or purely technical modifications, such as corrections or aesthetic enhancements, do not qualify as deepfakes where they do not affect the perceived authenticity or truthfulness of the content. <\/p>\n\n\n\n

2. A disclosure obligation borne by the deployer<\/strong> <\/h3>\n\n\n\n

Where an AI system generates or manipulates deepfakes, the deployer must disclose that the content has been artificially generated or manipulated. A lighter transparency regime applies to certain artistic, creative, satirical, fictional, or comparable forms of content. <\/p>\n\n\n\n

3. AI-generated text published to inform the public on matters of public interest<\/strong> <\/h3>\n\n\n\n

Article 50(4) also applies to certain AI-generated or AI-manipulated text published for the purpose of informing the public on matters of public interest. The relevant content includes text made accessible to an indeterminate and sufficiently broad audience, where the purpose is to communicate knowledge, opinions, or factual information. Matters of public interest encompass topics of societal relevance such as politics, public administration, justice, public security, public health, the environment, and the economy. <\/p>\n\n\n\n

An exemption applies where two cumulative conditions are met: first, the AI-generated or AI-manipulated text has been subject to human review or editorial oversight; and second, a natural or legal person assumes editorial responsibility for the publication, with their identity and contact details made publicly available.<\/p>\n\n\n\n

Article 50(5) – How information must be provided: clarity, visibility, and timing<\/strong> <\/h2>\n\n\n\n

1. Clear and distinct information<\/strong> <\/h3>\n\n\n\n

Article 50(5) establishes a horizontal requirement: information must be provided in a clear and distinct manner. Depending on the medium used, information is considered clear when it is perceptible, easy to understand, and accessible to the persons concerned. It is considered distinct when it can be readily identified as separate from other information and from the environment in which the content is presented. It must also be adapted to the intended audience, including children and persons with disabilities. <\/p>\n\n\n\n

2. What is not sufficient<\/strong> <\/h3>\n\n\n\n

The Guidelines specify that information will not be regarded as having been provided in a clear and distinct manner where it is included solely in a user manual or concealed behind multiple layers of menus within an online interface. <\/p>\n\n\n\n

3. First interaction, first exposure, and contextual reminders<\/strong> <\/h3>\n\n\n\n

The Guidelines also clarify when information must be communicated in order to ensure that it is effectively perceived by the persons concerned. <\/p>\n\n\n\n

The concept of first interaction covers the first natural person interacting with, or exposed to, the AI system or its output, as well as any subsequent first interaction or exposure involving another natural person. First exposure refers to any moment at which a person can reasonably be expected to be exposed to the system’s output and to perceive the disclosed information. It should be noted that providing a disclosure only at the beginning of the dissemination of content may be insufficient where it is reasonably foreseeable that some individuals will not be exposed to that content from the outset. <\/p>\n\n\n\n

For example, where a live video stream contains deepfakes, informing viewers only at the beginning of the broadcast is not sufficient. The disclosure should remain visible or be repeated at regular intervals for individuals who join the stream after it has started. <\/p>\n\n\n\n

Demonstrating compliance: Codes of practice, gap assessments, and documentation<\/strong> <\/h2>\n\n\n\n

Providers and deployers may demonstrate compliance with Articles 50(2), 50(4), and 50(5) by adhering to a Code of Practice on the Transparency of AI-Generated Content. Adherence to such a code constitutes a preferred means of demonstrating compliance, although it does not replace the requirements set out in the AI Act nor the guidance provided by the European Commission. <\/p>\n\n\n\n

Organizations that do not adhere to a recognized code of practice must demonstrate compliance through other appropriate means and be able to justify the effectiveness of the measures they have implemented. They are encouraged to conduct a gap assessment comparing their existing compliance measures against those set out in an appropriate code of practice in order to identify any shortcomings and define remediation actions. <\/p>\n\n\n\n

In practice, organizations should, at a minimum, maintain documentation covering: the AI systems concerned; their qualification under Article 50; the organization’s role (provider, deployer, or other actor); the measures used to inform affected persons; the marking and detection solutions implemented; any applicable exceptions; the tests and assessments carried out; key technical decisions and trade-offs; and the evidence demonstrating the effectiveness of the measures adopted. <\/p>\n\n\n\n

Conclusion: from declarative transparency to operational transparency<\/strong> <\/h2>\n\n\n\n

Article 50 of the AI Act requires organizations to move beyond declarative transparency and adopt an operational approach to transparency. It is no longer sufficient to simply state that a system uses AI. Organizations must determine whether users are interacting directly with an AI system, whether synthetic content is being generated or manipulated, whether a deepfake could mislead the public, or whether individuals are exposed to an emotion recognition or biometric categorisation system. <\/p>\n\n\n\n

Achieving compliance requires a structured approach: mapping AI systems, qualifying use cases, identifying the applicable obligations, designing appropriate disclosure mechanisms, selecting marking and detection solutions, and maintaining evidence of the decisions made throughout the process. <\/p>\n\n\n\n

\ud83d\udc49 With Naaia, transform the AI Act’s transparency obligations into an operational governance framework: classify AI systems, identify applicable obligations, manage compliance actions, and ensure traceability of key decisions. Discover our AI governance platform<\/a> and prepare with confidence for the application of Article 50 while demonstrating compliance in a structured and auditable manner. <\/p>\n","protected":false},"excerpt":{"rendered":"

From 2 August 2026, most of the transparency obligations set out in Article 50 of the AI Act will become applicable. Providers and deployers of AI systems will be required to implement appropriate transparency measures. In particular, they must… Lire la suite<\/a><\/p>\n","protected":false},"author":9,"featured_media":4140,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"naaia_last_modified":"","footnotes":""},"categories":[46],"tags":[],"class_list":["post-4139","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai-governance-blog"],"_links":{"self":[{"href":"https:\/\/naaia.ai\/en\/wp-json\/wp\/v2\/posts\/4139","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/naaia.ai\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/naaia.ai\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/naaia.ai\/en\/wp-json\/wp\/v2\/users\/9"}],"replies":[{"embeddable":true,"href":"https:\/\/naaia.ai\/en\/wp-json\/wp\/v2\/comments?post=4139"}],"version-history":[{"count":2,"href":"https:\/\/naaia.ai\/en\/wp-json\/wp\/v2\/posts\/4139\/revisions"}],"predecessor-version":[{"id":4145,"href":"https:\/\/naaia.ai\/en\/wp-json\/wp\/v2\/posts\/4139\/revisions\/4145"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/naaia.ai\/en\/wp-json\/wp\/v2\/media\/4140"}],"wp:attachment":[{"href":"https:\/\/naaia.ai\/en\/wp-json\/wp\/v2\/media?parent=4139"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/naaia.ai\/en\/wp-json\/wp\/v2\/categories?post=4139"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/naaia.ai\/en\/wp-json\/wp\/v2\/tags?post=4139"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}