When the EU AI Act Requires Disclosure
The EU AI Act does not ask one undifferentiated question—“Was AI involved?”—when it addresses transparency. It asks which actor is involved, what the system or output does, where the activity takes place, and which kind of interaction or content is at issue. This chapter turns those conditions into a practical way to identify the relevant Article 50 duty without treating every AI-assisted file as subject to the same label.
The starting point is responsibility. A file rarely tells a reader who supplied the system, who used it, or whether the output was used in the European Union. Those facts are part of the legal analysis, and they come before choosing a badge, a watermark, or a provenance format.
Start with the actor, not the file
Article 2 gives the regulation a wider reach than the location of an AI company might suggest. It applies to providers that place an AI system or a general-purpose AI model on the Union market or put it into service, whether they are established in the Union or in a third country. It also applies to deployers established or located in the Union. A provider or deployer in a third country can fall within the scope where the output produced by the system is used in the Union.
That last connection is important. An image generator hosted elsewhere does not become irrelevant to the European rule merely because its servers or corporate headquarters are outside Europe. The material fact is the use of the output. The regulation also lists other operators, including importers, distributors, certain product manufacturers, and authorised representatives. A public-facing explanation should not collapse all those roles into “the publisher,” because the duties can attach at different points in a supply chain.
Article 3 makes two distinctions especially useful. A provider is the natural or legal person, public authority, agency, or other body that develops an AI system or general-purpose AI model—or has one developed—and places it on the market or puts it into service under its own name or trademark. A deployer is the person or body using an AI system under its authority, except where the system is used in a purely personal, non-professional activity.
The distinction is functional. The provider controls how a system is designed and how its outputs are technically produced. The deployer controls a use of the system in a particular setting: a newsroom, public agency, company, platform, or other organization. One organization can occupy different roles in different arrangements. A company that develops its own generator and then uses it internally may have provider and deployer responsibilities, but the legal question for each action still depends on what it is doing.
The personal-use boundary also needs care. The Act excludes obligations of deployers who are natural persons using AI systems in a purely personal, non-professional activity. That is not a general statement that every consumer use is outside the regulation, and it does not erase every obligation that may fall on a provider or another operator. It is a reason to avoid describing the Act as a universal instruction to every individual who ever uses an AI tool.
Article 50 contains several different triggers
Article 50 is headed “Transparency obligations for providers and deployers of certain AI systems.” Its structure shows why a single “AI label” category is misleading.
Direct interaction with a person
Under paragraph 1, providers must ensure that AI systems intended to interact directly with natural persons are designed and developed so the people concerned are informed that they are interacting with an AI system. The exception is where that fact is obvious to a person who is reasonably well-informed, observant, and circumspect, taking account of the circumstances and context of use.
This is the conversational case. A person asking questions of a customer-service agent, tutor, or other interactive system may need a notice that the interlocutor is AI. The provision is about the interaction, not necessarily about a media file later generated by that system. A chat disclosure and a machine-readable image mark can both be transparency measures, but they answer different questions.
The paragraph contains a law-enforcement exception for systems authorised by law to detect, prevent, investigate, or prosecute criminal offences, subject to safeguards for third-party rights and freedoms. The details matter, but the broader lesson is simpler: the legal trigger includes a context and an actor, not merely a model’s capability.
Synthetic output and machine-readable marking
Paragraph 2 addresses providers of AI systems, including general-purpose AI systems, that generate synthetic audio, images, video, or text. The provider must ensure that the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated.
The required technical solution must be effective, interoperable, robust, and reliable as far as technically feasible. The Act says that implementation must take account of the specificities and limitations of different content types, the cost of implementation, and the generally acknowledged state of the art, including relevant technical standards.
Several words in this requirement do real work. Machine-readable means that the information is not only a sentence intended for a human viewer; it has a structured form that software can process. Detectable points to the ability to recognize the mark in the output. Interoperable matters because content moves between tools and platforms rather than remaining inside the system that created it. Robust and reliable signal that a mark which disappears under ordinary handling may not meet the practical objective, although the statute also acknowledges technical feasibility and cost.
The selected text does not prescribe C2PA, a particular watermark, or a particular metadata field. A provenance standard may be a candidate way to support this objective, but technical usefulness is not the same as automatic legal compliance. The legal requirement tells providers what capability the output marking must serve; it does not make every implementation of that capability equivalent.
Paragraph 2 also contains boundaries. The duty does not apply to the extent an AI system performs an assistive function for standard editing, does not substantially alter the input data provided by the deployer or the semantics of that input, or is lawfully used for criminal-offence detection, prevention, investigation, or prosecution. The wording is narrower than “a human touched the output.” It focuses on what the system did to the input and meaning.
Deployer disclosure for specific content
Paragraph 3 concerns deployers of emotion-recognition or biometric-categorisation systems. They must inform the natural persons exposed to the system of its operation, subject to the stated law-enforcement exception and data-protection requirements. This is not an AI-generated-content label, but it demonstrates that Article 50 covers transparency about a system’s operation as well as transparency about synthetic media.
Paragraph 4 shifts to content that a deployer makes or uses. A deployer of an AI system that generates or manipulates image, audio, or video content constituting a deep fake must disclose that the content has been artificially generated or manipulated. The Act defines a deep fake as AI-generated or manipulated image, audio, or video content that resembles existing persons, objects, places, entities, or events and would falsely appear to a person to be authentic or truthful.
The same paragraph also covers a deployer of an AI system that generates or manipulates text published with the purpose of informing the public about matters of public interest. That deployer must disclose that the text was artificially generated or manipulated, unless the use is lawfully authorised for criminal-offence work or the content underwent human review or editorial control and a natural or legal person holds editorial responsibility for publication.
This is not a rule that every internal draft, private note, or lightly edited sentence must carry the same notice. It is tied to publication, purpose, audience, and responsibility. It is also not a rule that a human edit automatically removes a duty. The public-interest-text exception combines human review or editorial control with a person or legal entity holding editorial responsibility.
Timing and form are part of the duty
Paragraph 5 says that the information referred to in paragraphs 1 to 4 must be provided to the people concerned in a clear and distinguishable manner no later than the first interaction or exposure. It also requires the information to conform to applicable accessibility requirements. This is the human-facing side of the rule. A mark that exists somewhere in a file but is not presented intelligibly at the relevant encounter may not serve the same purpose as a clear disclosure.
Paragraph 6 preserves other obligations in the Act and does not displace additional transparency duties under Union or national law. Paragraph 7 asks the AI Office to encourage and facilitate Union-level codes of practice for detecting and labeling artificially generated or manipulated content. The Commission may approve such codes through implementing acts, or specify common rules if a code is inadequate.
The timetable is also staged. Article 113 states that the regulation applies from 2 August 2026. It separately states that Chapters I and II apply from 2 February 2025; Chapter III Section 4, Chapter V, Chapter VII, Chapter XII, and Article 78 apply from 2 August 2025, with the exception of Article 101; and Article 6(1) and corresponding obligations apply from 2 August 2027.
For Article 50, the reproduced timetable places the transparency article outside the groups named for the earlier dates, so its general application point is associated with 2 August 2026 in this material. That is a useful operational orientation, not a substitute for checking the current Official Journal text, transitional provisions, codes, and later guidance before making a compliance decision. A regulation can have a general application date and still contain different dates for different chapters and obligations.
The path is deliberately conditional. A provider can face a technical marking question while a deployer faces a human disclosure question for the same output. A public-interest text workflow may require facts about editorial control that are irrelevant to a customer-service interaction. A third-country system may still matter if its output is used in the Union. And a personal non-professional use may change the deployer analysis without making all surrounding obligations disappear.
Classify before you choose a technology
Consider four simplified situations. A service chatbot designed to speak directly with customers raises the direct-interaction question for its provider. A general-purpose system that generates a new image raises the provider-side output-marking question. If an organization publishes a generated video that resembles a real person and would appear authentic, the deployer-side deepfake disclosure question becomes relevant. If an organization publishes AI-generated text to inform the public about a matter of public interest, the analysis turns on publication purpose and whether the required human review, editorial control, and editorial responsibility conditions are present.
These examples do not decide every real case. They show why choosing a technical mechanism first is backwards. Before asking whether a watermark, metadata field, or provenance manifest is suitable, identify the legal category and the audience that must receive the information. Then ask whether the chosen mechanism can preserve the relevant signal under the technical and social conditions of use.
The Act’s language also shows why “AI content” is too broad a category for operational decisions. An assistive editing function may fall within a stated boundary where a system does not substantially alter input data or semantics. A deepfake analysis depends on resemblance and the false appearance of authenticity or truth. A public-interest text analysis depends on purpose and editorial responsibility. The classification turns on facts about transformation, context, and control.
The important conclusion is therefore conditional: the EU AI Act requires different forms of disclosure from different actors in different situations. Its machine-readable output rule is not the same as a visible notice to a person, and neither one is a factual verdict about the content. Once those legal categories are clear, the next question is where the duties narrow or change—especially when the workflow involves standard editing, satire, creative work, editorial review, or content that only appears to be a deepfake.