The European Commission has issued guidelines clarifying how the transparency obligations under Article 50 of the AI Act should be implemented. The obligations themselves enter into application on 2 August 2026.
Background
Regulation (EU) 2024/1689, the AI Act, entered into force on 1 August 2024, establishing harmonised rules for the use of artificial intelligence across the Union. These systems are subject to the obligations set out in Article 50, which apply from 2 August 2026.
On 20 July 2026, the European Commission published new guidelines developed through broad stakeholder consultation. The guidelines are non-binding and are intended to give providers, deployers, and competent authorities practical assistance in applying Article 50 consistently across Member States.
Rationale and objectives
The purpose of transparency under the AI Act is to preserve trust and integrity in the information ecosystem. As generative and interactive AI becomes harder to distinguish from human made content, new risks arise around misinformation, manipulation at scale, fraud, and impersonation. The rules are built around three objectives.
The first is enabling informed decisions. Transparency allows individuals to recognise when they are interacting with or exposed to AI, so they can calibrate their level of trust and avoid overreliance on the system. The second is protecting fundamental rights, including personal autonomy, human dignity, freedom of expression, privacy, and non-discrimination. The third is safeguarding democracy, by reducing the risk of deception and limiting the potential impact of synthetic content on democratic processes and public trust.
Horizontal requirements applicable to all disclosures
Article 50(5) of the AI Act mandates horizontal requirements that apply to all transparency information provided under the article’s various paragraphs. These requirements stipulate that information must be:
- Clear and distinguishable: The information must be easy to understand, noticeable, and separate from other information or the environment in which it is presented. It should not be "hidden" within long terms of use or complex menu layers.
- Provided at the first Interaction or exposure: Notifications must be delivered at the latest when a natural person first interacts with the system or is first exposed to its output.
- Accessible: Information must conform to applicable accessibility requirements, particularly to meet the needs of natural persons with disabilities.
- Tailored to vulnerable groups: If a system is likely to interact with children or the elderly, notifications must be age-appropriate, child-friendly, and easy to understand for those specific audiences.
Obligations for interactive AI systems (Article 50(1))
Under the EU AI Act, specific transparency obligations for interactive AI systems are primarily governed by Article 50(1). These rules ensure that individuals can recognize when they are communicating with an artificial entity, helping them avoid overreliance and make informed decisions regarding the system's output.
Core obligation: the duty to inform
The primary obligation falls on providers of AI systems intended to interact directly with natural persons. They must design and develop these systems so that the persons concerned are explicitly informed of the artificial, non-human nature of their interacting counterpart.
For a system to fall under this obligation, it must meet four criteria:
- It is an AI system (not a simple rule-based mechanism).
- It is intended to interact in a bidirectional exchange of information.
- The interaction is direct (typically real-time or near real-time).
- The interaction is with natural persons.
Implementation & timing
- Embedded design: The notification mechanism must be embedded into the system's design and operation.
- Timeline: Information must be provided at the latest at the time of the first interaction.
- AI agents: Specialized AI agents must disclose their artificial nature and the identity of the person on whose behalf they are acting. If the provider cannot foresee every interaction, the agent must be designed to disclose its identity whenever it is reasonably likely to interact with a human.
Providers can choose their disclosure method, but the guidelines recommend certain best practices:
- Effective techniques: These include textual labels (e.g., "You are interacting with an AI system"), auditory disclosures at the start of a session, visual cues like persistent icons or "AI" symbols, and multi-modal combinations.
- Insufficient techniques: Compliance is not achieved by placing disclosures only in URLs or documentation, using non-perceivable machine-readable marks, or using ambiguous terms like "assistant" without clarifying the system's artificial origin.
Exceptions to the obligation
Two exceptions apply.
The obligation to inform does not apply where the artificial origin of the interaction is obvious to a reasonably well informed, observant, and circumspect person, for instance specialised AI tools used by professional developers.
It also does not apply to systems authorised by law to detect, prevent, investigate, or prosecute criminal offences, although this exception does not extend to systems made available to the public for reporting crimes.
Obligations for AI systems generating or manipulating synthetic content (Article 50(2))
Article 50(2) of the AI Act mandates that providers of AI systems generating or manipulating synthetic content implement technical solutions to ensure their outputs are marked and detectable. The goal is to allow individuals to distinguish AI-generated material from human-created content, thereby safeguarding the integrity of the information ecosystem.
1. Scope of the obligation
This requirement applies to AI systems (including general-purpose AI systems and AI agents) capable of generating or manipulating content in the following modalities:
- Audio: Time-varying signals encoding sound, such as speech or instrumental music.
- Images: Static representations encoding visual information, including 3-D images.
- Video: Time-based sequences of images, including virtual reality (VR) and augmented reality (AR).
- Text: Discrete symbolic content composed of characters or numbers.
2. The technical solution: marking & detection
Compliance requires a two-fold approach; fulfilling only one element is insufficient.
- Marking: Providers must ensure outputs are marked in a machine-readable format. This allows software to identify and extract the marks without human intervention. Techniques include watermarks, metadata, cryptographic methods for provenance, or fingerprints.
- Detection: Providers must ensure that means of detection are available to those exposed to the content. These tools should produce human-readable results at the time a person wishes to verify the origin of the content.
3. Quality requirements for technical solutions
The Act stipulates that these solutions must meet four specific quality standards "insofar as this is technically feasible" and aligned with the state of the art:
- Effectiveness: The capability to detect marks and distinguish AI-generated content.
- Reliability: The ability to accurately identify AI origin under normal conditions across a variety of outputs.
- Robustness: The ability to maintain accurate detection under varying conditions, including common alterations or adversarial attacks.
- Interoperability: The capability for solutions to operate seamlessly across different systems and actors, regardless of which marking technique was originally used.
4. Exceptions & exclusions
The marking and detection obligations do not apply in the following scenarios:
- Standard editing: AI functions used for minor tasks like grammar correction, spellchecking, noise reduction, or formatting that do not substantively change the content's meaning.
- Non-substantial alterations: Cases where the AI does not significantly manipulate the input data or its semantics (e.g., cropping or minor colour adjustments).
- Law enforcement: Systems authorised by law for detecting, preventing, or investigating criminal offences.
- Specific technical outputs: Short sequences of symbols (like alt-text or UI labels), source code, and outputs intended exclusively for machine-to-machine communication.
- Industrial/B2B applications: In limited cases where content is strictly technical, only perceived by professional staff, and not shared externally.
These obligations generally apply from 2 August 2026. However, a "grandfathering" rule exists for generative AI systems placed on the market before that date, giving providers until 2 December 2026 to bring those specific systems into conformity.
Obligations for labelling deep fakes and certain text publications (Article 50(4))
Under Article 50(4) of the AI Act, deployers of generative AI systems have specific transparency obligations regarding deep fakes and text publications intended to inform the public on matters of public interest. These obligations are distinct from the machine-marking requirements for providers and focus on ensuring the end-user is clearly informed.
1. Deep fakes
A deep fake is defined as AI-generated or manipulated image, audio, or video content that appreciably resembles existing persons, objects, places, or events and would falsely appear to a person to be authentic or truthful.
- Standard obligation: Deployers must clearly and distinguishably disclose that the content has been artificially generated or manipulated. This disclosure must be perceivable (e.g., visible or audible labels) so that a person does not need technical tools to recognize it.
- Artistic and satirical exception: A "lighter" disclosure regime applies to deep fakes that are part of evidently artistic, creative, satirical, or fictional works. In these cases, transparency must be provided in an appropriate manner that does not hamper the enjoyment or display of the work (e.g., credits at the end of a movie). However, this must still be subject to safeguards for the rights and freedoms of third parties.
- Examples:
- In-scope: Voice cloning of a podcast presenter, AI-generated videos of politicians, or realistic synthetic avatars of CEOs.
- Out-of-scope: Content that is clearly unrealistic (e.g., a sphinx flying over the Eiffel Tower) or standard technical adjustments like noise reduction that do not change the substance of the content.
2. Text publications
This obligation applies to AI-generated or manipulated text that is published with the purpose of informing the public on matters of public interest.
- Scope: "Published" means the text is accessible to a large, indeterminate number of people. "Matters of public interest" include topics relevant to society at large, such as politics, public health, environmental protection, and public security.
- Disclosure obligation: Deployers must clearly and distinguishably disclose that the text has been artificially generated or manipulated.
- The "Human Review" exception: Transparency is not required if the AI-generated content has undergone substantive human review or editorial control, and a natural or legal person holds editorial responsibility for the publication. Fact-checking is considered a minimum requirement for this review.
- Examples:
- Requires labeling: AI-generated summaries of town council decisions or AI-generated public safety warnings published without human review.
- Does not require labeling: AI-generated fantasy novels (not informing on public interest) or text that has been thoroughly fact-checked and edited by a human editor-in-chief.
Enforcement & timeline
Providers and deployers who fail to meet the transparency obligations under Article 50 are subject to significant financial penalties, enforced by national market surveillance authorities, the AI Office, or the European Data Protection Supervisor.
General providers and deployers face administrative fines of up to EUR 15,000,000 or, for undertakings, up to 3% of total worldwide annual turnover for the preceding financial year, whichever is higher. EU institutions, bodies, and agencies face fines of up to EUR 750,000. For SMEs and start-ups, the lower of the two figures applies, to protect their economic viability.
When setting the amount of a fine, competent authorities consider the nature, gravity, and duration of the infringement and its consequences, whether it was intentional or negligent, the degree of cooperation with market surveillance authorities, and any other aggravating or mitigating circumstances. Adherence to a code of practice that the Commission has assessed as adequate may be treated as a mitigating factor when the fine is calculated.
These transparency rules, and the penalties attached to them, enter into full application on 2 August 2026.