AI Act transparency requirements: understanding Article 50

What Article 50 covers

Article 50 of the EU AI Act sets specific duties for certain systems and content. People should be able to tell when they are interacting with AI or encountering certain artificially generated or manipulated material. Article 50 transparency obligations have applied since 2 August 2026.

There is no blanket duty to label every use of AI. The system's function, the content, the organisation's role and the context determine what is required. Start with an AI inventory and identify the organisation's obligations.

The 2026 amendment introduced a narrow transition: providers of synthetic-content systems already placed on the market before 2 August 2026 have until 2 December 2026 to meet the Article 50(2) marking duty. Other Article 50 duties do not receive a general extension.

Providers and deployers

A provider develops an AI system, or has one developed, and places it on the market or puts it into service under its name or trademark. A deployer uses a system under its authority in a professional activity. The same organisation can hold different roles for different systems.

The deployer's configuration step in the chatbot row is practical good governance; Article 50(1) places the design duty on the provider. Technical marking and a notice people can perceive serve different purposes.

Direct interaction with people

Providers of systems intended to interact directly with natural persons must design them so that people know they are interacting with AI, unless this is obvious to a reasonably well-informed, observant and circumspect person in the circumstances. Customer-service chatbots, assistants and digital avatars can fall within this rule. Background automation does not necessarily amount to direct interaction.

An opening notice such as “You are speaking with an AI assistant” addresses the immediate need. There is no general requirement in this paragraph to explain the model's architecture. An organisation relying on obviousness should record why the interaction would clearly be understood as artificial.

Technical marking of synthetic content

Providers of AI systems, including general-purpose systems, generating synthetic audio, image, video or text must ensure that relevant outputs are machine-readable and detectable as artificially generated or manipulated under Article 50(2).

Possible methods include provenance metadata and other technical marks. The Act calls for effective, interoperable, robust and reliable solutions as far as technically feasible, accounting for content and the state of the art. It does not prescribe a single universal watermark or require that every technical mark be visible to viewers.

Standard editing and minor changes

Article 50(2) does not impose marking where a system merely performs an assistive function for standard editing, or does not substantially alter the input data or its semantics. Minor corrections and technical quality improvements may fit, depending on the actual output. Changing the meaning of a statement or creating an entirely new scene raises different questions.

Calling a feature “editing” is not enough. Compare the input, output and change in meaning before deciding.

Emotion recognition and biometric categorisation

Deployers using emotion recognition or biometric categorisation systems must inform the natural persons exposed to their operation. Applicable data protection rules also remain relevant.

Notice alone does not make any such use lawful. Some practices face prohibitions or other duties. Classify the system before designing the notice.

Deepfakes and visible disclosure

A deepfake under Article 3(60) is AI-generated or manipulated image, audio or video that resembles persons, objects, places, entities or events and falsely appears authentic or truthful. Not every AI image qualifies; resemblance and presentation matter.

Where it is a deepfake, the deployer must disclose its artificial or manipulated nature. Invisible provider metadata cannot replace a clear notice to the viewer or listener. Evidently artistic, creative, satirical or fictional works have a tailored disclosure rule, allowing an appropriate notice that does not unnecessarily spoil enjoyment. They are not wholly exempt.

Text on matters of public interest

Article 50(4) also covers AI-generated or manipulated text published to inform the public on matters of public interest. It does not automatically capture every internal draft, email or promotional slogan.

A company may prepare a public-health article with AI, review its substance, exercise genuine editorial control and take responsibility for publication. The statutory exception requires human review or editorial control, and editorial responsibility held by a natural or legal person. A spelling check alone does not qualify. Automatically publishing informative text without such a process calls for disclosure.

When and how to inform people

Information under paragraphs 1 to 4 must be clear and distinguishable, meet applicable accessibility requirements and reach people by their first interaction or exposure. Depending on the medium, this can mean an opening message, an adjacent label, an audible notice or an accessible interface indicator.

An obscure line buried in a long policy does little to inform someone encountering a deepfake video. A visible label at the outset is a more meaningful example.

Exceptions and their limits

The exceptions are specific: obvious AI interaction, standard editing or non-substantial changes for technical marking, certain legally authorised crime-related uses, tailored disclosure for creative works, and editorial safeguards for public-interest text. They should not be combined into a general exemption.

The Commission guidelines also clarify some outputs not intended for human exposure. Record which paragraph, function and context supports any exception. A broader voluntary notice may be a good practice, but should not be described as a universal legal requirement.

Business example: four AI uses

A company has a customer-service chatbot, an image-generation tool for campaigns, an assistant drafting articles and a synthetic video depicting its director announcing a new measure.

Chatbot

The provider designs the system to inform people unless the AI nature is obvious. The company checks that any opening notice actually appears and can be understood.

Image

The tool provider assesses the technical marking duty. Article 50(4) does not impose a blanket duty on the company to visibly label every AI image; it assesses whether a particular image is a deepfake.

Video

If it appears to show the real director making a statement she never made, it may qualify as a deepfake. The company must clearly disclose the manipulation even if technical marking is embedded in the file.

Text

A marketing slogan is not automatically public-interest reporting. A health-policy article may be: substantive review or editorial control together with editorial responsibility can engage the exception; automatic publication without that safeguard calls for an assessment of disclosure.

Evidence to retain

Article 50 imposes the relevant information and marking duties; it does not specify one universal document list for every organisation. Useful governance records include the system inventory, provider and deployer roles, intended use, Article 50 analysis, decision, notice wording, interface screenshots, technical configurations, editorial checks, responsible people, dates and periodic reviews.

Where a provider implements machine-readable marking, test representative outputs and record limitations. Evidence should show why the chosen control applies and how it was checked.

AI governance and transparency

A practical sequence is inventory → classification → role → applicable transparency duty → control → evidence → monitoring. An AI use policy can assign publication approval; vendor management can obtain marking information; and generative AI governance can make editorial review consistent.

ISO/IEC 42001 can structure responsibilities, controls and evidence. It does not replace or reproduce Article 50's legal duties.

Commission guidelines and voluntary code

On 20 July 2026, the Commission published Article 50 transparency guidelines, clarifying roles, exceptions, synthetic content and ways to inform people. They aid interpretation; the Regulation establishes the legal duty.

The Code of Practice on Transparency of AI-generated Content offers voluntary measures to demonstrate compliance with marking and labelling obligations. Voluntary adherence does not make Article 50 optional.

Where to start

For each use, identify the system and the provider or deployer. Determine the relevant Article 50 paragraph, its conditions and exceptions. Choose the appropriate marking or notice, test it from the perspective of someone encountering the content, and record the decision.

Transparency works when people receive the information they need at the point they need it. A generic label everywhere cannot replace a careful assessment of actual use.

Sources

• Consolidated Regulation (EU) 2024/1689, Articles 3, 50 and 113.
• Regulation (EU) 2026/1744, limited transition for Article 50(2).
• Commission guidelines, July 2026 and official FAQs.
• Code of Practice on Transparency of AI-generated Content.