On August 2, 2026, Beijing time, Article 50 of the EU's Artificial Intelligence Act on transparency obligations came into effect, covering direct interaction between humans and AI, machine-readable labels for synthetic content, and several public disclosure scenarios. This is not the first time this law has taken effect: it entered legal force on August 1, 2024, with prohibitive practices, AI literacy, and general AI model obligations being phased in subsequently. Official EU materials did not list any cases filed, fines imposed, or market access bans on the first day of Article 50's application during this round of verification.
Interaction Notification and Machine Marking
For AI systems that interact directly with natural persons, providers must ensure that relevant individuals are aware they are interacting with an AI system in the design and development stages. If a person with reasonable information, observation skills, and caution can already clearly identify this based on the usage environment, repeated notifications may be waived under the regulation; specific systems used for criminal investigation, prevention, investigation, or prosecution also have exceptions subject to safeguard measures. Therefore, the law stipulates identifiable outcomes and application boundaries without mandating that all products use the same type of pop-up windows, buttons, or notification frequencies.
For systems generating synthetic audio, images, videos, or text, including general AI systems, providers must ensure outputs are marked with machine-readable labels and can be identified as artificially generated or manipulated. The regulation requires technical solutions to remain effective, interoperable, robust, and reliable within the bounds of technological feasibility, taking into account the characteristics of different content types, technical limitations, implementation costs, and recognized technology levels. Official materials do not specify that all services must use the same watermarking protocol nor have they published detection success rates for each in-use product.
Deployer's Disclosure Boundaries
Parties deploying emotion recognition or biometric classification systems must inform natural persons in contact with the system that it is operational and that personal data is being processed according to applicable EU data protection rules. This obligation targets actual users of the system, not developers who cannot be assumed to have fulfilled notification on behalf of all using entities. Article 50 also sets exceptions and safeguards for specific uses legally aimed at detecting, preventing, or investigating crimes; existing official summaries do not provide a generic disclosure statement that can be applied directly without considering the context.
Deployers must also disclose artificial intelligence-generated or manipulated images, audio, and video constituting deepfakes. For AI-generated or manipulated text published for public interest matters, official guidelines list the absence of human review or editorial control as relevant conditions; if reviewed by individuals or institutions bearing editorial responsibility, different judgments apply. Custom disclosure rules exist for art, creation, satire, fiction, and similar works, so it cannot be generalized that all synthetic content must display identical visible watermarks.
Four Types of Practical Scenarios and Supervisory Bodies
In four practical scenarios, different entities undertake distinct actions: direct interaction notification and machine labeling for synthesized content primarily fall on providers, while public disclosure of emotion recognition, biometric categorization, deep fakes, and specific public interest texts mainly rest with deployers. Machine-readable labels serve technical detection, visible disclosures serve those who come into contact with the content, and although both may coexist, they are not interchangeable obligations. Product forms, content types, release methods, and legal exceptions all influence specific compliance paths.
Supervision is also tiered. The European Commission lists enforcement bodies as including market surveillance authorities of member states, AI offices responsible for oversight within their purview, and the European Data Protection Supervisor when EU institutions act as providers or deployers. The commission's public overview sets the maximum fine for related companies at €15 million or 3% of global annual turnover, with a cap of €750,000 for EU institutions; these figures represent statutory upper limits and are not automatic fines issued on August 2.
Transition Period and Subsequent Timeline
The timeline must be understood in conjunction with the scope of obligations. EU Regulation No. 1744/2026 sets a four-month transition period for synthetic content systems already on the market before August 2, 2026, requiring providers to meet the machine-readable labeling requirements under Article 50(2) by December 2; this is not an across-the-board extension of all transparency obligations. Content generated and made public before August 2 also does not require retroactive labeling. Following adjustments through the Digital Operational Law, the standalone high-risk use rules in Annex III are deferred to December 2, 2027, while the high-risk rules for embedded regulated products are extended until August 2, 2028.
The European Commission's final guidelines explain the scope and examples of Article 50. The voluntary AI-generated content transparency code of conduct provides a framework for demonstrating compliance. Not signing the code itself does not constitute a violation; non-signatories must still demonstrate how they fulfill their legal obligations in other appropriate ways. Signing is also not conclusive proof of compliance. Subsequently verifiable public records include regulatory inquiries and decisions, labeling schemes adopted by companies, performance assessments, lists of signatories to the code, and court interpretations of exception clauses; this round's sources have yet to provide these initial results.