Realistic AI-made media now carries a clearer disclosure duty across the European Union. Transparency provisions in the EU AI Act began applying August 2, 2026, requiring labels for deepfake images, video and audio. The rules also add technical marking duties intended to help platforms and audiences detect content generated or altered by AI.
Deepfakes must disclose their artificial origin
The rules focus on media that could reasonably be mistaken for an authentic person, place, object or event. A convincing synthetic recording of a politician, executive or private citizen can travel as evidence even when no such recording occurred. A visible disclosure is meant to interrupt that deception at the point where an audience encounters it.
The European Commission’s July 31 enforcement announcement says deepfake images, video and audio must be labelled. It also says AI-generated or altered content must carry machine-readable marks so automated systems can identify its artificial origin more easily.
Visible and machine-readable signals solve different problems. A label gives a person immediate context, while embedded metadata or another technical marker can help a service screen large volumes of media. Neither method guarantees that a copied or edited file will retain the disclosure, making durable implementation a central challenge.
The obligations reach beyond deceptive video
Interactive AI systems such as chatbots generally must tell people when they are dealing with a machine rather than a human. The law also addresses text generated or manipulated by AI when it is published to inform the public about matters of public interest. Those provisions treat transparency as relevant to conversation and written information, not only spectacular visual fakes.
The Commission’s AI Act framework overview distinguishes those transparency risks from minimal-risk uses such as AI-enabled games or spam filters. It says providers of generative AI must make generated content identifiable and certain public-facing content must be clearly and visibly labelled.
Scope and exceptions matter. The law does not turn every use of image correction, automated translation or software assistance into the same kind of disclosure event. Providers and deployers have different obligations, and the treatment of artistic, satirical or editorial material can depend on how the content is presented.
That distinction prevents the rule from collapsing into a generic badge on everything touched by software. A disclosure must communicate useful information about artificial generation or manipulation, not merely signal that a modern tool was present somewhere in production. Detailed Commission guidance gives businesses a basis for deciding who must mark an output, who must display the label and how exceptions operate.
August 2 changed the legal status
The AI Act entered into force in August 2024, but its requirements have arrived in stages. Prohibited practices and AI-literacy duties began applying in February 2025, while rules for general-purpose AI models followed in August 2025. The broader transparency provisions and enforcement structure reached their scheduled application date on August 2, 2026.
From that date, the EU AI Office and national authorities became responsible for implementation, supervision and enforcement. That makes the labelling provisions more than an announced policy direction. Companies serving the European market must evaluate whether their systems or uses fall within the duties that are now applicable.
Labels depend on a chain of cooperation
A model provider can create a technical signal, but the file may pass through editing software, messaging apps and social networks before reaching an audience. Each step can preserve, display, hide or strip information. Effective transparency will require compatible standards and product choices throughout that chain.
Bad actors can crop out a visible label or remove metadata. The rules can still raise the cost of deception by making compliant services label ordinary output and by giving investigators a standard against which to judge omissions. Provenance signals work best as evidence within a wider verification system, not as an infallible stamp.
Publishers and platforms will also need to decide how a technical marker becomes understandable at the screen level. A machine-readable signal that remains invisible to the audience cannot perform the whole disclosure job. Clear icons, nearby text and accessible explanations can connect the underlying provenance data with a decision made by an ordinary reader or viewer.
Disclosure does not settle whether content is true
An AI label says something about production, not accuracy. A synthetic chart may faithfully summarize verified data, while an unaltered recording can be presented with a false caption. Audiences still need source information, context and independent reporting to decide whether a claim is reliable.
The rules nevertheless establish a useful baseline: realistic artificial media should not depend on every viewer spotting tiny visual errors or unusual speech patterns. As generation quality improves, human detection becomes less dependable. Requiring the producer or deployer to disclose the artificial origin places responsibility closer to the technology that created the ambiguity.
Europe’s implementation will test how well that principle survives routine publishing, cross-border distribution and deliberate removal. The labels will matter most when they remain clear to people and portable enough for machines to recognize after content leaves its original service.
This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.
More from Morning Overview
