EU AI Act — Transparency Obligations for AI-Generated Content (Article 50)

Providers must ensure that AI systems intended to interact directly with natural persons are designed and developed so that natural persons are informed that they are interacting with an AI system, unless this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use.

The purpose is to allow individuals to make informed decisions on outputs, to avoid overreliance on systems and to calibrate trust in the system outputs and interactions.

Guidance inverses the order of the EU AI Act.

First, consider the individual, then “obviousness”.

Provider has burden. Standard draws on EU consumer protection concept of “average consumer” and a two-step analysis:

(1) The intended audience

(2) How well informed a member of that audience is

Relevant factors: §44

Should be interpreted restrictively.

Examples: §45

“Authorised by law” refers to EU or EU Member State law. (§47)

The purpose of the AI system must be to detect, prevent, investigate, or prosecute criminal offences (e.g., AI-undercover agent).

Entails bidirectional exchange of information or actions with a genuine conversational or responsive character.

Individuals should be able to provide input to the system, and the AI system should be capable of responding with outputs.

It does not need to be initiated by a human. Passive collection or one-time collection of feedback is not covered. (§30)

Real-time or near real-time interaction.

Excludes indirect or mediated interaction.

Products and services that blend AI-generated responses with human-curated content should be disclosed as being AI-generated unless the outputs have been “properly” reviewed and sent by a human as the recipient’s main interlocutor. (§30)

In scope of intended to interact with natural persons that instruct them or in the course of executing a task.

AI agents must be designed and developed to disclose their artificial nature and the identity of the person on whose behalf they are acting (including when it is reasonably likely to interact with a human).

Disclosures to be made for key steps such as authorization, reporting, validation, and at every new interaction. (§31)

Providers of AI systems (including GPAI systems) generating synthetic audio, image, video or text¹ must ensure outputs are:

• Marked in a machine-readable format
• Detectable as artificially generated or manipulated, using technical solutions that are effective, interoperable, robust and reliable. (Article 50(2))

Definitions:

• Audio: time-varying signal encoding sound perceivable by humans.
• Image: static spatial representations.
• Video: time-based sequence of images (including 3D, VR/AR/MR, digital twins).
• Manipulation in scope if it goes beyond “standard editing” (§§60+, 90)

Limitations

(Exemptions are covered in Part 2C)

Costs may be disproportionate to marginal gains. (§85)

There may be narrowly defined cases where a technical solution based on less robust metadata markings is sufficient for compliance, e.g. when system is embedded in a controlled and closed environment. (§86)

B2B applications where proportionality principle and limited benefits preclude application of the obligation. (§87)

Commitment 3: Signatories commit to strive to achieve the best possible balance between effectiveness, interoperability, robustness and reliability of the marking and detection solutions as described in Measures 3.1 to 3.4.

Operational constraints may impose limitations on computational time and resources, costs of implementation, as well as on scalability to very large or very small content.

Providers may rely on a single technique or combination, provided the overall solution is machine-readable and meets the four quality requirements. (§68)

Techniques include watermarks, metadata, cryptographic methods, logging, fingerprints (Recital 133, §73).

Not prescribed by the Guidelines.

Providers choose the technique(s) provided the overall solution meets Article 50(2) requirements. (§72)

Under the current state of the art, at least two layers (metadata + watermark) mandatory for content disseminated online.

Alternative techniques may be permitted in the future if proven equivalent (Measure 1.1).

Detection solution available free of charge (limited exceptions for high-volume users of small providers).

Results must be clear, comprehensible, indicating whether based on metadata, watermark, forensic detection or other (Measure 2.1, Measure 2.3).

Sub-measure 2.1.3 requires detection solutions involving uploads to address privacy, confidentiality, data minimisation, security, limited retention or zero retention where applicable, and EU data protection compliance.

Sub-measure 2.1.4 permits retirement of a detection solution only where replaced by an alternative with the same or better capabilities and backward compatibility.

Refers to the capability of different technical solutions for marking and detection to operate seamlessly across multiple systems, actors, contexts and technical implementations. (§79)

Providers must rely on publicly-available industry standard detection solutions to ensure full interoperability. (§76)

Until such standards exist, providers may rely on their own, a third-party or a shared solution. Should be a time-limited situation. A provider that uses a third-party solution is responsible for ensuring its compliance with the Act.

Effective: Capability of solution to detect their marks and enable natural persons to identify AI-generated or manipulated content.

Reliable: Capability of the solution to accurately identify and distinguish AI-generated or manipulated content from other content across a variety of content.

Robust: Capability of solution to accurately identify and distinguish AI-generated or manipulated content from other content under varying conditions. (§79)

Effective: –Considered effective when natural persons can access and understand the meaning of the detection results (Measure 3.1).

Reliable: Assessment consists of two components: (i) how accurate the detection solution is regarding detecting the marking in nominal conditions where there is no alteration to the content, and (ii) how the accuracy of the marking and detection solutions varies with respect to the length, size, diversity, and semantics of the content (Measure 3.2).

Robust: Solutions to maintain performance levels under varying conditions. Does not apply to AI systems subject to only to one lawyer of marking (Measure 3.3).

Encouraged: Fingerprinting/logging (supplement only — Sub-measure 1.1.3).

Encouraged: Richer provenance metadata (Measure 1.3).

Encouraged: Perceptible marking functionality for deployers (Measure 1.4).

Marking: Not prescribed. Providers may rely on a single technique or combination provided the four quality requirements are met. (§76)

Detection: Providers must make means of detection available and ensure the detection result can be provided to natural persons in a clear and distinguishable manner.

Marking: Under the current state of the art, two-layer approach (digitally-signed metadata and imperceptible watermark), consistent with audio/image/video. Contingent on state of the art. Alternatives permitted if proven equivalent (Measure 1.1).

Detection: Solution requirements apply identically. Results must be clear, distinguishable and accessible (Measure 2.1, Measure 2.3).

Marking: Single layer only, imperceptible watermark (metadata not possible). For text greater than 200 tokens, watermarking is mandatory even if reliability is lower (Sub-measure 1.1.2).

Detection: Access to detection mechanisms may be restricted to verified expert users (authorities, researchers, media, fact checkers) given lower reliability. Restriction limited in time until more reliable mechanisms emerge (Sub-measure 2.1.2).

Examples may qualify as either editing or alternations. Guidelines do not clarify. (§92)

Out of scope:

Grammar correction and spellchecking, linguistic and minor polishing, formatting, format conversions, technical compression, noise reduction or removal for enhanced clarity, without changing meaning.

Minor cropping, minor colour adjustments or corrections, lightening or darkening, sharpening for enhanced clarity or other standard editing and technical corrections, removal of dust spots, removal of red eye, deleting and obscuring backgrounds, pixelation or blurring of facts, rescaling a video clip, dynamic range compression and equalisation, limited video stabilization; standard adjustments to colour and contrast, minor adjustments to playback speed, noir corrections to level the horizon of an image, applying pixel filters to amplify parts of an image, applying colour maps, edge image completion, converting a black-and-white photo to colour and vice versa.

Pixel filling for clarity, automatic transition clips.

Medical image technical processing or reconstruction within medical devices.

Transcription of conversations.

In scope:

AI-generated summaries of text; paraphrasing or rewriting text that changes style, structure and meaning beyond mere grammatical and minor stylistic correction.

Removal, replacement or insertion of objects or persons in existing images and videos; facial replacement, modification.

Synthesis of realistic speech in a specific person’s voice, generating realistic video depicting events that did not occur; altering body shape or skin colour, extreme lightening or darkening or contrast changes that change the meaning, intent and messaging.

Creation of composite images or videos.

Case-specific assessment.

AI-assisted Augmentative and Alternative Communication (AAC)/customised neural voices that do not alter meaning are exempt. (§85-86)

Not exempt: Translations, summaries, adding/removing objects, pixelation/blurring faces, altering body shape/skin colour, extreme colour adjustments, black-and-white conversion, composite images. (§91)

Deployers of AI systems that generate or manipulate image, audio or video content constituting a deep fake must disclose that the content has been artificially generated or manipulated (Article 50(4), subparagraph 1).

Exception: Use authorised by law for criminal offense detection/prevention/investigation/prosecution

Four cumulative conditions:

(i) AI-generated/manipulated image, audio or video (not text)

(ii) Appreciable resemblance to existing subjects

(iii) 'Existing' means realistic, (unrealistic content excluded)

(iv) Would falsely appear authentic or truthful — assessed considering diverse audience composition (children, elderly, lower digital literacy), not an 'average person' test. The assessment is holistic and context driven.

AI-supported manipulation of insignificant substantive or technical aspects of pre-existing content may be of minor relevance for a person’s assessment of the authenticity or truthfulness of the content.

(§§113-116)

Article 50(4), subparagraph 1, read with Article 50(5).

Deployers must disclose that image or video deep fake content has been artificially generated or manipulated, in a clear and distinguishable manner at the latest at the time of first exposure.

Must be understandable/perceivable without technical tools or dedicated actions.

At the start of a video, when encountering image while scrolling, etc.

End credits alone do not comply with Article 50(5).

For live/varied-timing content, also at later stages or persistently. (§§110, 132)

Mandatory:

• EU icon ('AI' acronym) or equivalent label. Placed with no intervening overlay, ensuring immediate recognition. Visible for sufficient duration.
• For video: at beginning, after interruptions, at regular intervals. Directly embedded or equivalent overlay.
• Accessible disclosure required (audio descriptions, tactile/haptic cues, high-contrast, screen reader compatible, detectable by assistive technologies) (Sub-measures 1.2.1, 1.2.2, Measure 1.1).

Article 50(4), subparagraph 1, read with Article 50(5).

Same disclosure standard applies to audio deep fakes. ²

Mandatory: Short audible disclaimer in plain language at the beginning.

For long-form/live audio: Reminders at regular intervals and after interruptions. Where screen is available: additional visual disclosure (Sub-measure 1.2.3).

Encouraged: Interactive second layer ('generated' vs. 'modified').

Alternative audible disclosures (earcons) pending EU-wide audio solution, with awareness-raising measures.

Article 50(4), subparagraph 1.

The transparency obligation is not removed but limited to disclosure of the existence of such generated or manipulated content in an appropriate manner that does not hamper the display or enjoyment of the work.

Content must evidently fall within one of five categories — ambiguous content does not qualify. Will be subject to strict interpretation; relevant factors considered are listed. (§122)

Disclosure is still required even if the deep fake is evidently part of an artistic, creative, etc. work (§123) and Recital 134. What is appropriate disclosure is a case-by-base assessment.

Deployers must ensure appropriate safeguards for third-party rights (IP, data protection, personality rights, freedom of expression). (§124), Recital 134.

Categories:

• Artistic (art, music, cinema)
• Creative (deployer's creative choices, not functional/technical)
• Satirical (criticism through humour);
• Fictional (imaginary setting)
• Analogous (similar expressive/functional purposes)

Examples of deep fakes that do and do not qualify as creative/satirical/fictional or analogous works: §124

Icon/label placed appropriately to the type of work. At the latest at first exposure (e.g., accompanying notes, beginning/end credits — but end credits alone insufficient).

For digital/interactive: Icon may be adjacent to frame or in UI elements if perceivable without dedicated actions.

For non-digital (exhibitions, galleries, cinemas): At point of entry, accompanying materials, physical carriers (Commitment 3).

Article 50(4), subparagraph 1.

Full exemption where use authorised by law for criminal offence detection/prevention/investigation/prosecution.

Article 50(4), subparagraph 2.

Deployers of AI systems that generate or manipulate text published with the purpose of informing the public on matters of public interest must disclose the artificial origin.

Exceptions: (i) human review/editorial control + editorial responsibility (ii) law enforcement.

Accessible by an indeterminate, fairly large number of unrelated potential readers simultaneously/successively, whether paid or free.

Text that is not published: Private correspondence, organisation-internal documents, messages in small, closed groups. (§131)

Text must intend to communicate knowledge, opinions or facts.

Text with an entirely different objective (e.g. entertainment) is not in scope. (§131)

Matters relevant to society at large, meriting public debate or scrutiny: public administration, fundamental rights, public health, environmental protection, consumer safety, democratic processes, public security, financial developments, economic/political/scientific/cultural developments with important public implications. (§131)

In scope: AI news summaries on websites; AI-manipulated academic papers; AI-manipulated corporate investor reports; AI weather warnings on social media.

Out of scope: AI novels/poems; AI advertising text (provided the advertisement is not published with the purpose of informing the public on matters of public interest — e.g., containing claims related to health, consumer safety, sustainability or other matters of public interest; chatbot news summary available only to the individual user; text manipulated by a professional for client advice regarding required regulatory compliance measures.

(§123)

Article 50(4), subparagraph 2 requires deployers to disclose that qualifying published text has been artificially generated or manipulated.

Article 50(5) sets the clear, distinguishable, first-exposure and accessibility requirements.

Must be perceivable through visible or audible measures, without technical tools or dedicated actions.

At the start of the publication (not buried at the end).

Applicable accessibility requirements must be met. (§132)

EU icon or equivalent label above/at the top of the text, near the headline, or in the colophon — clear, consistent and distinguishable.

For short text where labelling would degrade readability: Contextual notice in UI e.g., adjacent indicator or disclosure at beginning of exposure session (Sub-measure 1.2.2(f)).

Deliberate substantive examination by competent natural person(s); or control by an editorial entity with authority to approve, alter or reject based on substance (including fact checking, a minimum requirement).

Superficial, purely formal or procedural checks (spell checking, grammar correction, automated review processes) do not suffice. (§§134-137)

Substantive AI intervention occurring after human review can cause the exception to be voided.

A natural or legal person must hold ultimate legal responsibility over the publication.

Identity of the natural person and thecontact information must be publicly available. (§138)

Article 50(4) disclosure obligation does not arise.

Article 50(5) requirements for deployer labelling do not apply.

The provider's Article 50(2)/(5) obligations continue to apply to the system.

(§§125-128)

Examples: §138

Media service providers under Article 2(2) of Regulation (EU) 2024/1083 may rely on existing procedures.

All other deployers must establish/maintain policies including identification of the person with editorial responsibility and organisational measures.

Contact details of the person/function with editorial responsibility must be published (Commitment 4).

Article 50(5).

Information referred to in paragraphs 1–4 shall be provided to natural persons concerned in a: (i) clear and (ii) distinguishable manner at (iii) the latest at the time of the first interaction or exposure. The information shall conform to applicable (iv) accessibility requirements.

Applies to each output with respect to any natural person exposed.

Providers/deployers may inform earlier.

For content encountered at different times (e.g. live broadcasts), disclosure should also occur at later stages or persistently. (§143)