Europe’s AI Disclosure Rules Begin on August 2
By Suad Seferi ·
The new requirements will not place a warning on everything created with AI. They will, however, make it harder for chatbots, deepfakes and automated public communication to quietly pass as human or authentic. From August 2, some of Europe’s most visible uses of artificial intelligence will have to become more transparent. People in the European Union must be informed when they are communicating directly with certain AI systems or encountering particular forms of AI-generated and manipulated content. The obligations come from Article 50 of the EU AI Act and enter into application on August 2, 2026. The European Commission published its final implementation guidelines on July 27, only days before the rules take effect. They clarify how the requirements apply to chatbots, synthetic images and audio, deepfakes and AI-generated text dealing with matters of public interest. The change does not mean that every email corrected by an AI tool, every edited photograph or every article drafted with some AI assistance must carry a large “AI-generated” label. The rules are more specific than that. They target situations in which people may reasonably believe that artificial content, communication or behaviour is authentic. Chatbots should no longer pretend to be people Companies and public institutions using AI systems that communicate directly with individuals will generally need to disclose that the person is interacting with a machine. The notice should be clear and provided early in the interaction, unless it is already obvious from the circumstances that the system is artificial. This affects customer-support chatbots, digital assistants and other interactive AI services. It also challenges a common design practice in which organisations give automated systems human names, photographs and conversational personalities while leaving users to discover for themselves that no person is present. The issue is not whether an AI assistant sounds friendly. It is whether the user understands who, or what, is answering. Deepfakes must be identified The rules also require disclosure when AI is used to create or manipulate images, audio or video that could falsely appear authentic. This includes deepfakes representing real or plausible people, places, objects, organisations and events. The disclosure must be understandable to the audience. Machine-readable metadata may help platforms and technical systems identify synthetic material, but hidden information inside a file will not always be enough for the person watching or listening. There are more flexible provisions for artistic, fictional, satirical and creative works. That does not create a broad exemption for synthetic political communication, fabricated institutional announcements or manipulated footage presented as evidence of a real event. This distinction matters far beyond Brussels. AI-generated campaign photographs, synthetic public messages and manipulated political content are already appearing across the Western Balkans, often without consistent labelling or even a basic explanation of how the material was produced. A creator may consider the fiction obvious. A citizen encountering a cropped image or reposted video days later may not. Public-interest text faces a different test One of the more sensitive requirements concerns AI-generated or manipulated text published to inform the public about matters of public interest. This can include communication relating to elections, public administration, health, justice, law enforcement, public safety and other issues that shape public debate. Such material may require disclosure when it is generated or substantially manipulated by AI and published without meaningful human review or editorial responsibility. This is particularly relevant to news organisations, government institutions, political parties and communications agencies experimenting with automated publishing. The regulation is not designed to punish a journalist for using transcription, translation, spelling correction or another standard editing tool. The Commission’s guidelines explicitly recognise exceptions such as standard editing. The harder question is what happens when AI produces most of a public statement, news article or institutional announcement and a person merely approves it without properly checking the facts, sources or meaning. A human click on the publish button does not automatically transform automated content into responsible communication. The real test is whether someone with the authority and competence to intervene has reviewed the content, can change or reject it and accepts responsibility for what is published. Technology companies and users have different duties Providers of generative AI systems are expected to make synthetic content detectable, including through machine-readable marking. The organisations and professionals deploying those systems may carry separate responsibilities to disclose deepfakes or certain AI-generated public-interest publications to their audiences. The Commission has also introduced a voluntary Code of Practice covering the marking and labelling of AI-generated content. Joining the code is optional, but the underlying Article 50 transparency obligations are not. Companies that choose another compliance method must be able to demonstrate that it is adequate. The EU has additionally prepared optional icons that organisations may use when labelling artificial content. Still, no icon can solve the entire problem. A label may tell the audience that AI was involved. It does not reveal whether the content is accurate, whether the sources were checked or whether anyone is prepared to take responsibility for mistakes. The Western Balkans will feel the pressure The AI Act applies within the European Union, but its influence will extend into neighbouring markets. Technology companies in the Western Balkans that provide AI products or services to EU users may fall within its reach. Regional busines…