# AI-Generated Content: Article 50 Obligations and Deepfakes > Article 50 of the AI Act will apply to all generative AI systems from August 2, 2026. It imposes transparency requirements on AI content, including deepfake labeling, technical marking, and user information, regardless of the risk level. - Tags: article 50 ai act, transparence ia générative, deepfakes ai act, étiquetage contenu ia, watermarking intelligence artificielle, conformité ai act, code de pratique ai act - Main keyword: transparence contenus IA Article 50 AI Act ## Key points - Article 50 of the AI Act imposes transparency obligations on all generative AI systems, without distinction of risk, from August 2, 2026: chatbots, image generators, deepfakes, texts, and multimedia content must be signaled and labeled according to precise modalities. ## Introduction Article 50 of the AI Act imposes, as of August 2, 2026, a series of transparency obligations on all generative AI systems, regardless of their risk level. This concerns chatbots as well as image generators, speech synthesis tools, writing assistants, or any device producing text, audio, video, or images synthetically. All organizations that design, provide, or use such systems must anticipate these new requirements, the scope of which is much broader than that reserved for high-risk systems alone. ## Content ## Scope and Reach of Article 50 Article 50 of Regulation (EU) 2024/1689 details four distinct transparency obligations applicable to any use of generative AI systems. Unlike the obligations in Chapter III, these do not depend on a risk classification: they apply as soon as one of the following cases arises. ### 1. Information Obligation During AI Interactions (Art. 50§1) Providers must ensure that any person interacting with an AI system is informed that they are communicating with a machine and not a human. This requirement specifically targets chatbots, virtual assistants, and other AI in direct contact with individuals. **Exception:** If the context makes the artificial nature of the interlocutor obvious, the obligation may not apply. However, this exception is interpreted restrictively in practice. ### 2. Technical Marking of Synthetic Content (Art. 50§2) Providers of generative AI systems, including GPAI models, must ensure that each generated content (text, image, audio, video): - Is marked in a machine-readable format (e.g., watermarking, metadata, provenance techniques) - Remains detectable as artificially generated or manipulated, via robust, interoperable, and reliable solutions Marking mechanisms must withstand common manipulations (compression, cropping, format changes). Providers must also offer free detection tools, accompanied by a confidence score. **Exceptions:** AI systems used for standard editing or that do not substantially modify the provided data, as well as those authorized for the detection or prosecution of criminal offenses, are exempt. ### 3. Transparency on Emotion Recognition and Biometric Categorization (Art. 50§3) Deployers of AI systems capable of recognizing emotions or categorizing biometric data must inform the affected persons, prior to any exposure, about the functioning of these devices. ### 4. Labeling of Deepfakes and AI-Generated Texts (Art. 50§4) Deployers who publish content generated or modified by AI that constitutes a deepfake have the obligation to clearly signal its artificial nature. This requirement extends to texts produced by AI and disseminated to inform the public on matters of general interest. A deepfake, according to Art. 3, point 60, refers to any image, audio, or video content generated or modified by AI that mimics real persons, objects, places, or events and is likely to mislead regarding its authenticity. **Exceptions:** - Works that are manifestly artistic, creative, satirical, or fictional: minimal and non-intrusive disclosure is sufficient - Content used in the legal context of detecting or prosecuting criminal offenses - Texts that have undergone human review and are assumed by an editorial manager ## Division of Responsibilities: Providers and Deployers Article 50 distinguishes between two levels of actors, each with specific obligations. ### Providers: Ensuring Technical Marking Providers, who develop and market AI systems, must integrate technical marking—watermarking, metadata, detection mechanisms—by design. They must also provide deployers with the tools necessary to implement labeling obligations. The Code of Practice currently under development specifies that foundation model providers must integrate these mechanisms "by design" before any market release, to allow downstream providers to meet their own obligations. ### Deployers: Ensuring Visible Labeling Deployers, i.e., entities that use AI systems under their responsibility, are responsible for visibly informing end-users. This includes affixing a common icon, inserting warnings, and ensuring that the artificial origin of the content is clear to the user. Strictly personal and non-professional use of an AI system does not make the user a deployer. However, the public dissemination of AI-generated content can, even for an individual, constitute an act of deployment. ## The Code of Practice for AI Content Transparency To facilitate the application of Article 50, the AI Office launched the drafting of a Code of Practice on the marking and labeling of AI-generated content at the end of 2025. Although non-binding, this Code will become the reference standard for compliance: regulators and judges will refer to it to evaluate corporate practices. ### Specialized Working Groups The drafting of the Code involves experts and stakeholders, divided into two groups: - **WG1: Marking and Detection Techniques** (watermarking, metadata, interoperability, robustness, technical governance) - **WG2: Disclosure of Deepfakes and AI Texts** (visible labeling, taxonomy, distinction between "fully AI" and "AI-assisted" content, platform responsibilities) During the January 2026 meetings, debates focused on the complementarity between technical marking and visible labeling, the prevention of information fatigue, and alignment with other European texts such as the Digital Services Act. ### Common "AI" Icon The Code provides for the creation of a harmonized icon at the European level to signal AI content. Pending the finalization of an interactive symbol, a visual label containing the acronym "AI" or its local equivalent ("IA" in French, "KI" in German) is recommended. ### Taxonomy: "Fully AI" vs. "AI-Assisted" The Code distinguishes between: - **Fully AI-generated**: content produced without authentic human intervention - **AI-assisted**: content resulting from significant collaboration between human and AI This distinction has an impact on legal protection: "fully AI" content might not benefit from European copyright, making it freely reusable. ### Code of Practice Timeline - December 2025: Publication of the first draft - March 2026: Publication of the second draft (with stakeholder feedback) - June 2026: Final version expected - August 2026: Entry into force of Article 50 obligations ## Concrete Modalities by Content Type The draft Code specifies disclosure measures adapted to each content format. ### Live Deepfake Video The AI icon must be displayed permanently and non-intrusively during the entire broadcast. A warning must also be presented at the beginning. ### Recorded Deepfake Video An initial warning or a visible icon continuously at a fixed location. For artistic or fictional works, the icon must appear for at least 5 seconds. ### Deepfake Images The common icon must be placed at each exposure, at a fixed location on the image. ### Deepfake Audio A short audio warning in plain language must be broadcast at the beginning. For content exceeding 30 seconds, the warning must be repeated. For artistic works, an initial non-intrusive warning is sufficient. ### AI Texts of Public Interest The icon must be displayed in a fixed, clear, and visible position (at the top, next to it, in the colophon, or after the conclusion). ## Articulation with Other European Regulations Article 50 does not apply in a legal vacuum: - **Digital Services Act (DSA)**: platforms deploying AI content must reconcile the requirements of Article 50 and those of the DSA (moderation, transparency, risk management) - **GDPR**: emotion recognition or biometric categorization systems (Art. 50§3) raise personal data protection issues - **GPAI Obligations**: for GPAI models (Art. 51-56), the transparency obligations of Article 50 are in addition to the specific obligations of providers - **High-Risk Systems**: systems that are both generative and high-risk combine the transparency requirements of Article 50 and those of Chapter III ## Key Steps for Companies by August 2026 With the entry into force scheduled for August 2026, organizations have little time to achieve compliance. Here are the priority actions: ### 1. Map Concerned AI Systems Inventory all AI systems in your structure that generate or modify content (text, image, audio, video) or interact with people. Include third-party tools used (writing assistants, image generators, chatbots, etc.). ### 2. Identify Your Role in the Value Chain Determine if you are a provider (development of the AI system) or a deployer (use of an existing system). Obligations differ depending on the role. Many companies combine both statuses depending on the case. ### 3. Evaluate Existing Marking Mechanisms For providers: verify that your systems integrate robust, interoperable, and tamper-resistant watermarking, metadata, and detection solutions. ### 4. Review Publication Processes For deployers: analyze how AI content is published in your organization. Are labeling procedures in place? Are the concerned teams trained? ### 5. Anticipate Copyright Implications The "fully AI"/"AI-assisted" distinction can have an impact on the legal protection of your content. Evaluate the consequences for your intellectual property. > To evaluate your transparency obligations according to your role (provider or deployer), the [free compaia diagnostic](https://compaia.eu/conformite-ai-act) guides you according to the AI Act regulation. ## Official source - [Source](https://www.aiacto.eu/fr/blog/article-50-ai-act-transparence-deepfakes-contenu-ia)