The European Union is introducing some of the world’s most extensive transparency requirements for artificial intelligence, forcing companies and professional content publishers to disclose when people are interacting with AI or viewing certain synthetic media.
From August 2, 2026, Article 50 of the EU AI Act will apply to generative systems, chatbots, deepfakes and selected AI-produced publications. The rules are intended to make it easier for people to recognise content that could otherwise be mistaken for authentic photographs, recordings, videos or human-written information.
The change does not mean that every AI-assisted image, edited sentence or computer-generated graphic must carry a large visible warning. The obligations vary according to who provides the AI system, who publishes its output and how realistically the content represents actual people, places, objects or events.
The European Commission’s official transparency guidance explains that the requirements are designed to reduce misinformation, impersonation, fraud and consumer deception while allowing legitimate creative and commercial uses of generative AI to continue.
The Rules Target Content That Could Be Mistaken for Reality
The most visible requirement concerns deepfakes. Under the AI Act, a deepfake can be an AI-generated or manipulated image, video or audio recording that resembles an existing person, place, object, organisation or event and could falsely appear authentic or truthful.
A realistic video that appears to show a politician making a statement they never made would fall within this category. The same could apply to fabricated footage of a real event, an AI-generated recording that imitates a recognisable person’s voice or a realistic photograph placing someone in a situation that never occurred.
Professional users publishing this material must provide a clear and perceptible disclosure explaining that it was artificially generated or manipulated. The EU’s quick guide to Article 50 confirms that deepfakes require visible disclosure because an ordinary viewer may otherwise believe the material is real.
This distinction matters because not every AI image is a deepfake. A visibly fictional dragon, an abstract background or a highly stylised illustration may not create the same risk of confusion as a photorealistic image of a real public figure. The context, realism and way the content is presented all influence the applicable disclosure.
AI Providers Must Add Machine-Readable Markings
Companies developing generative AI systems have a separate technical responsibility. Their systems must be designed so that synthetic text, audio, images and video can be detected as artificially generated or manipulated.
This usually means adding a machine-readable marker, metadata signal, watermark or another form of provenance information that automated systems can identify. The mark does not necessarily need to appear as visible text across the image. Its purpose is to provide reliable technical evidence about the content’s origin.
The European Commission’s Article 50 overview states that providers must mark synthetic outputs unless the system performs only a standard editing function or does not substantially alter the meaning of the material supplied by the user.
A tool that corrects brightness, sharpens an image or removes minor background noise may therefore be treated differently from one that generates a new face, invents a speech or substantially changes what appears to have happened.
The distinction also separates the obligations of providers from those of deployers. The company creating an image generator may need to embed the technical marker, while a marketing agency, broadcaster or social-media creator publishing a realistic output may need to add the visible disclosure.
Public-Interest AI Text May Also Need a Label
The rules are not limited to visual and audio deepfakes. AI-generated or substantially manipulated text published to inform the public about matters of public interest must also be clearly labelled when it has not received meaningful human review or editorial control.
This could cover automatically generated reporting about elections, health policy, public safety, financial regulation or other issues that influence public understanding and decision-making.
The requirement is narrower than a general obligation to label every piece of AI-assisted writing. Text that has been reviewed, verified and published under the editorial responsibility of a person or organisation may qualify for an exception.
The Commission defines the covered category as public-interest text that has not undergone human review or editorial control. Its official transparency summary makes clear that both the subject matter and the absence of responsible human oversight are relevant.
A journalist using AI to help organise research would not automatically need to label the finished article when an editor has checked the facts and accepts responsibility for publication. A fully automated news page that generates and publishes politically significant stories without review would face a different compliance position.
Chatbots Must Identify Themselves as AI
Article 50 also covers direct interaction with AI systems. Providers must ensure that people are told when they are communicating with a chatbot, virtual assistant or another interactive AI system unless the artificial nature of the interaction is already obvious.
A customer-service bot should therefore not present itself as a human employee without disclosure. A realistic AI-generated sales representative, support agent or voice assistant must make its identity clear early enough for the person to understand the nature of the interaction.
The same framework requires disclosure when emotion-recognition or biometric-categorisation systems are being used, subject to the Act’s specific exceptions. The wider purpose is to ensure that individuals can decide how much confidence or personal information they want to place in an automated system.
Creative and Satirical Content Receives More Flexible Treatment
The EU rules recognise that an identical disclosure format would not be appropriate for every artistic or entertainment work.
Content that is evidently artistic, fictional, satirical or creative may use a disclosure method that does not interfere unnecessarily with the work’s display or enjoyment. A fictional film using AI visual effects, for example, may not need a permanent warning placed across every frame.
This should not be understood as permission to present deceptive synthetic media as real simply by describing it as creative. The presentation must still fit the nature and context of the work, particularly when real people or events are depicted convincingly.
Personal, non-professional activity is also treated differently from commercial or institutional deployment. The strongest obligations generally fall on AI providers and professional deployers that place systems or content into the European market.
Existing AI Systems Receive Additional Time
Although Article 50 applies from August 2, not every existing generative AI product must complete the technical marking transition on that exact day.
The Commission says generative AI systems placed on the market before August 2 receive a grace period for the machine-readable marking obligation until December 2026. Systems introduced after the deadline are expected to comply when they enter the EU market.
Deepfakes created before August 2 do not have to be labelled retroactively, although voluntary disclosure is encouraged. Material generated after the rules begin will need to be assessed under the new requirements when it is professionally published or distributed.
This transition period gives established platforms time to modify generation pipelines, introduce provenance technology and update the tools used by customers.
The EU Has Created Optional AI Icons
To make disclosures more consistent, the European Commission has produced three black-and-white icons that organisations can use to identify AI-generated or manipulated material.
Using these icons is optional. Companies may design their own labels, provided the disclosure remains clear, understandable and appropriately positioned. Displaying an EU icon also does not automatically prove that every legal obligation has been satisfied.
The Commission’s AI labelling icon resource states that deployers remain responsible for ensuring that the chosen disclosure meets Article 50.
The EU has also approved a voluntary Code of Practice on AI-generated content. It separates responsibilities between providers responsible for technical marking and deployers responsible for visible labels on deepfakes and selected text publications. Organisations that do not follow the code may use other methods, but they must still demonstrate equivalent compliance.
Non-Compliance Can Bring Significant Fines
Article 50 is backed by financial penalties rather than operating only as voluntary guidance.
Companies that violate the applicable AI Act obligations can face fines of up to €15 million or three percent of their total worldwide annual turnover, depending on which amount and enforcement calculation applies. EU institutions can face penalties of up to €750,000, while proportionality requirements are intended to account for the size of smaller businesses.
Enforcement will involve national market-surveillance authorities, the EU AI Office for systems under its supervision and the European Data Protection Supervisor when European institutions are involved.
The practical effect will extend beyond companies headquartered inside Europe. A technology provider, advertising agency, media company or online platform based elsewhere may still fall within scope when offering an AI system or publishing covered material in the EU market.
AI Labels Will Become Part of Everyday Digital Content
The August 2 requirements represent a major change in how synthetic media is presented. AI provenance will increasingly become part of the content-production process rather than a warning added only after a misleading post attracts attention.
Generative AI companies will need technical methods for marking outputs. Publishers and professional creators will need procedures for identifying deepfakes, reviewing public-interest text and placing visible disclosures where required. Platforms will need to preserve provenance information when content is uploaded, edited or redistributed.
The rules will not eliminate deceptive AI material, particularly when bad actors deliberately remove labels or operate outside compliant platforms. They can, however, establish a common expectation that realistic synthetic content should not be presented as authentic without explanation.
For ordinary users, the most noticeable result may be more AI notices on advertisements, videos, images, automated articles and customer-service interactions. For businesses, the more important change will occur behind the scenes, where content workflows must record how material was created, whether it received human review and which party is responsible for disclosure.
