The obligations under Article 50 aim to mitigate risks of deception and manipulation and foster the integrity of the information ecosystem. They apply as from 2 August 2026 and complement other rules under the AI Act, such as those for high-risk AI systems and general-purpose AI (GPAI) models. Transparency is the core: AI-generated content, deep fakes and certain AI-based publications must be labelled and made recognisable.
The Code of Practice on Transparency of AI-Generated Content was drawn up by independent experts in a multi-stakeholder process facilitated by the AI Office. It helps providers and deployers of generative AI systems to comply with their obligations under Article 50(2), (4) and (5). Even though adherence to the Code is voluntary, the transparency requirements under Article 50 are legal obligations.
The Code of Practice on transparency of AI-generated content has two sections:
- Section 1: providers – rules for marking and detection of AI-generated and manipulated content,
- Section 2: deployers – rules for labelling deep fakes and AI-generated and manipulated text.
The Code frequently draws attention to the required proportionality of the measures, taking into account the size and resources of the signatories, in particular with regard to small and medium enterprises (SMEs) and small mid-cap enterprises (SMCs), to minimise the burden and simplify compliance.
A set of icons has been developed that deployers of generative AI systems can use to label their AI-generated content.
The Code was published on 10 June 2026 and was assessed by the European Commission as adequate on 9 July 2026. All providers and deployers signing the Code must use the specified measures to demonstrate compliance with the AI Act’s rules for the labelling and detection of AI-generated content, deep fakes and certain text publications. This aims to reduce the administrative burden and create predictability, legal certainty and trust across Member States.
By contrast, providers and deployers deciding to comply through other means must be able to demonstrate that their measures are adequate. This is assessed individually by different market surveillance authorities. Signatories to the Code will also collaborate in task forces that will be set up to share best practices and advance the implementation of the marking and labelling obligations.
Providers and deployers of generative AI systems can sign up to the Code. The signatory form and instructions and questions and answers about signing the Code are available on the European Commission’s website.
The content of the Code’s two sections is summarised below. The obligations and measures apply to the providers and deployers of
AI systems who have signed the Code (and are therefore referred to in some places as “signatories”).
Section 1: providers – rules for marking and detection of AI‑generated and manipulated content
As no single marking technique suffices to meet all the requirements in Article 50(2), a combination of marking techniques and detection mechanisms is necessary. Implementation of such solutions has to take account of the whole value chain.
Commitment 1: marking of AI-generated or manipulated content
To comply with the obligation under Article 50(2) – marking the outputs of generative AI systems (including GPAI systems) in a machine-readable format – marking solutions have to be implemented as applicable to the content generated or manipulated by the individual AI systems. This applies to audio, image, video and text content, and any combinations of these types of content.
In brief
At least one machine-readable marking technique that is effective, reliable, robust and interoperable must be implemented. If the data format supports metadata, information must be recorded in the metadata on whether the content is “AI-generated” or “AI-manipulated”. Imperceptible watermarks can be embedded during or after generation. Other techniques (such as fingerprinting) can be implemented and further information can be recorded in the marking (for example the name of the AI model and a timestamp) as an optional supplementary measure. Existing metadata markings must be retained and must not be intentionally altered or removed. The terms and conditions accompanying the generative AI systems must therefore include a relevant prohibition.
[ENDE]
Measure 1.1: machine-readable marking techniques
A marking solution that consists of at least one machine-readable marking technique must be used that, in conjunction with the respective detection mechanism, meets the required level of effectiveness, reliability, robustness and interoperability. As no single marking technique can currently ensure compliance with all four requirements, a multi-layered marking approach is permissible, where the outputs of generative AI systems are marked with at least two layers of machine-readable marking.
Sub-measure 1.1.1: digitally signed metadata
If AI-generated or manipulated content is in a data format that supports attaching metadata (especially audio, video, image or containerised text), information must be recorded in the metadata on whether the content is AI-generated or manipulated. All the recorded information must be digitally signed and time-stamped in a secure and tamper-evident manner.
Sub-measure 1.1.2: imperceptible watermarking
AI-generated or manipulated content must be marked with an imperceptible watermark, with the exception of very short text. The watermark should be embedded into the content in a manner that is difficult for it to be separated from the content. Two options are possible:
- “post-hoc watermarking”, that is, applying the watermarking technique once the content has been generated or manipulated by an AI system, and
- “model watermarking” during the inference process, that is, embedding the watermarking into the content while the content is generated.
Sub-measure 1.1.3: fingerprinting or logging (optional)
An optional supplementary measure is fingerprinting or logging by the AI system provider. This can serve to verify whether content has been generated or manipulated by the AI system.
Measure 1.2: non-removal of markings
Metadata markings on input data and AI-generated or manipulated content must be retained and must not be removed. Two measures must be implemented:
- Existing metadata markings must be retained and must not be intentionally altered or removed. Exceptions include specific legitimate processing such as for security audits and research purposes.
- The intentional removal of or tampering with metadata markings by third parties must be prohibited in the terms and conditions or the documentation accompanying a generative AI system, with the exception of the above-mentioned legitimate purposes.
No tools may be provided or promoted whose purpose is to circumvent these markings. Providers of platforms and search engines are encouraged to retain metadata markings for AI-generated or manipulated content.
Measure 1.3: transparency of the provenance information (optional)
Additional context information can optionally be added to the metadata, such as the name of the AI system, the company name of the AI provider, a timestamp indicating when the content was generated or manipulated, the identifier and version number of the AI model and information about the type of operations performed to manipulate content (for example removal of an object in an image).
Measure 1.4: functionality for perceptible markings (optional)
An optional functionality can be provided in the user interface to allow users to add a perceptible label upon generation of the content. These measures are without prejudice to the responsibility of deployers to disclose deep fakes and AI-generated or manipulated text in a clear and distinguishable manner in accordance with Article 50(4).
Commitment 2: detection of markings of AI-generated or manipulated content
In order to fulfil the transparency obligations under Article 50(2) and (5), measures must be taken to enable the detection of machine-readable markings in AI-generated or manipulated content such as audio, image, video or text content. The detection result must be provided to the natural persons concerned in a clear, distinguishable and accessible manner.
In brief
Providers of generative AI systems should make available a detection solution that enables deployers, users of generative AI systems, third-party integrators, end-users and other legitimate parties to verify the AI-based origin of content. The detection solution can consist of one or more detection mechanisms but must in any case include a suitable detection mechanism for each marking technique implemented. The detection solution must be implementable as widely as possible, easily accessible through a user-friendly interface and implemented in compliance with EU data protection law.
It must be possible to download the detection result (including a hash and information about the detection solution). The result itself must include the basis of the detection mechanism (for example metadata marking, watermark marking or other techniques) and may be supplemented with additional information (such as the proportion (in percentage) and localisation of AI changes in the total content).
[ENDE]
Measure 2.1: detection mechanisms for markings
Providers of generative AI systems must make available a detection solution composed of one or more detection mechanisms to enable deployers, users of generative AI systems, third-party integrators, end-users and other legitimate parties (such as competent authorities, independent researchers, civil society and media organisations) to verify whether content has been generated or manipulated by an AI system. The detection and marking solutions must meet the level of effectiveness, interoperability, robustness and reliability required by Article 50(2), Commitment 3 of the Code and the following sub-measures.
Sub-measure 2.1.1: making the detection solution available
The detection solution should be implementable as widely as possible and easily accessible and should be made available in one or more of the following forms:
Possible forms for making the detection solution available in the EU| Form | Description | Recommendation |
|---|
| (i) Public specification | An (ideally standardised) specification allowing any third party to implement a detection mechanism | Provision of a reference implementation |
| (ii) piece of software | Standalone executable or library | Version portable on most operating systems (including, where technically feasible, mobile phone systems) with minimal hardware requirements |
| (iii) Cloud-based service | Cloud-based service accessible to users in the EU through an API | |
It must be ensured that the detection solution includes a suitable detection mechanism for each marking technique implemented pursuant to Commitment 1. Shared or third-party detection mechanisms can be used, for example from providers that integrate markings at the model level or from providers specialised in transparency marking techniques. In order to facilitate compliance by downstream providers, providers of generative AI models should make available detection mechanisms for the content generated or manipulated by their models prior to a model’s placement on the market.
The detection solution must generally be made available free of charge, although there are certain exceptions for companies with fewer than 1mn monthly users.
Sub-measure 2.1.2: access to the detection solution
Access to a detection solution must be through a user-friendly interface appropriate for the audience of end-users. If access to these detection solutions is limited, it must be ensured that the detection solutions are accessible to those user groups that also have access to the AI-generated or manipulated content, such as the general public or only a few legitimate persons.
The detection result of supposed AI-generated or manipulated content must be able to be downloaded upon request in a digitally signed format including at least a hash of the content submitted for detection, information about the detection solution used (such as a URL or an identifier) and a timestamp.
Sub-measure 2.1.3: personal data protection, privacy and security
The detection solution must be implemented in compliance with EU data protection law. The following key points apply to any solution requiring content to be uploaded:
(a) sensitive processing in compliance with data protection law,
(b) confidentiality and integrity during transit and at rest,
(c) data minimisation,
(d) purpose limitation,
(e) immediate deletion following detection,
(f) relevant security safeguards guaranteeing the confidentiality, integrity, availability and resilience of the service,
(g) compliance with other requirements under data protection law,
(h) implementation of best practices in cybersecurity.
Sub-measure 2.1.4: retirement of a detection solution
A detection solution may be retired as long as it is replaced by an alternative solution that has the same or better capabilities and is backward compatible, enabling the detection of previously marked content.
Measure 2.2: forensic detection mechanism (optional)
The detection solution can be supplemented by an optional, forensic detection mechanism that detects AI-generated or manipulated content even though the marking has been removed. The above-mentioned requirements relating to data protection and security and to easy accessibility and clear presentation of the results also apply here.
Measure 2.3: clear and accessible disclosure
The results provided by the detection solutions must be presented in a way that is clear and easily comprehensible to natural persons wanting to verify the origin of content in accordance with Article 50(5).
The detection results must indicate whether they are based on a metadata marking, a watermark marking, forensic detection or other techniques, to the extent technically feasible. If the watermark or metadata contains additional information, this information must be incorporated into the detection results. Further information can be included, such as the proportion of the content that has been generated or manipulated by the AI system and where exactly changes have been made in the content.
Measure 2.4: support literacy on AI marking and detection solutions (optional)
It is recommended for deployers and other expert users to be provided with documentation and other relevant information (such as the form of access, application, interpretation) enabling them to make informed decisions on the marking and detection solutions to be used. In addition to deployer-focused materials, it is recommended for easily understandable resources to be developed to promote literacy and awareness among end-users.
Commitment 3: measures to meet the requirements for marking and detection solutions
The quality requirements described in the following measures must be complied with to ensure that the marking and detection mechanisms for AI-generated or manipulated content are effective, interoperable, robust and reliable in accordance with Article 50(2). Compliance is assessed holistically across all marking and detection techniques rather than for each technique individually. Practical constraints – such as limitations on computational time, resources, costs or scalability to very large or very small content – can be taken into account.
Conformity with the testing, verification and compliance processes specified in Commitment 4 must be assessed and demonstrated before a generative AI system is placed on the market or put into service and throughout the lifecycle of the system.
In brief
The marking and detection solutions deployed must primarily comply with the following four quality aspects:
- Effectiveness
Detection results must be accessible and their meaning must be understandable by users. - Reliability
Detection solutions must accurately distinguish the origin of AI-generated or manipulated content from other content. This must function across diverse contexts and use cases (= high level of accuracy). - Interoperability
Marking and detection solutions must operate across multiple systems, actors, context and technical implementations. Detection must be enabled regardless of the marking technique deployed. - Robustness
The marking and detection solutions must be robust to both typical content-distorting processing operations and localised changes and to malicious manipulation and attacks on the solutions.
[ENDE]
Measure 3.1: effectiveness
Technical marking and detection solutions must be implemented that, in conjunction, are capable of enabling natural persons to distinguish AI-generated or manipulated content.
A solution is considered effective if users can access and understand the meaning of the detection results; the effectiveness is assessed on the basis of user feedback. It is therefore especially recommended to publish specifications to guide downstream providers on how to display the detection results.
Measure 3.2: reliability
Marking and detection solutions must be implemented that, in conjunction, achieve a high level of reliability in all expected use cases. Reliability in this context means that the solutions can accurately distinguish the origin of AI-generated or manipulated content from other content and operate reliably across diverse contexts and use cases. This is assessed on the basis of two aspects:
- How accurate is the detection solution with regard to detecting the marking in nominal conditions (without any subsequent alternations)?
- How does the accuracy vary with respect to the length, size, diversity and semantics of the content?
The detection accuracy must be measured with relevant metrics (such as the error rate) for the type of markings and demonstrated on a variety of samples.
Measure 3.3: robustness
The marking and detection solutions implemented must, together, maintain intended performance levels under varying conditions, covering both common alterations and adversarial attacks. This requirement does not apply to AI systems that are exceptionally subject only to one layer of metadata marking as specified in Measure 1.1.
Firstly, solutions must be robust to typical processing operations that may distort content, for example:
- in-place modifications: noise addition/removal, filtering, (re)compression, screenshots, voice enhancement, lexical substitution, homoglyphs, change of file format, etc;
- desynchronisation mechanisms: mirroring, cropping, up-/downscaling, rotation, aspect ratio change, pitch shifting, time stretching, characters insertion/deletion, paraphrasing, translation cycles, etc;
- survival of the analogue hole: print-and-scan (and optical character recognition), audio playback and recording, screen camcording, etc.
In addition, detection must be robust to localised changes, for example blurring faces for privacy reasons.
Secondly, adversarial robustness must be assessed, that is, how well the solutions are protected against malicious behaviour such as copying, removal, regeneration and modification attacks on the marking. Robustness is measured by the ability of the marking and detection solutions to identify whether marking has been tampered with or modified. The implementation of standard cybersecurity practices (such as rate limits) and frequent updates of threat assessments are recommended to prevent and counteract malicious use and attacks against marking and detection solutions.
Measure 3.4: interoperability
Marking and detection solutions must be implemented that operate across multiple systems, actors, contexts and technical implementations to enable detection of AI-generated or manipulated content, regardless of the marking technique deployed, to the extent technically feasible.
At the time of publication of the Code, generally recognised interoperability standards were not yet available, except for digitally signed metadata. A staged implementation has therefore been adopted.
In the initial phase as of the entry into application of Article 50(2):
- established standards or best practices to implement metadata marking and detection must be adopted to enable open detection of such markings;
- public information on how to integrate and access detection solutions for all other implemented marking techniques must be provided, for example through websites or publicly available registries;
- a minimum level of interoperability among detection solutions for watermarking must be achieved by 2 February 2027 by implementing at least one of the options listed in the Code.
The solution chosen must not be detrimental to the robustness of other embedded watermarks and must prevent misuse.
Measure 3.5: advancing to the state of the art (optional)
Contingent upon capacity, investment in scientific research and development and collaboration with authorities, researchers, civil society organisations and other stakeholders are recommended to advance the state of the art in marking techniques and detection mechanisms for AI content.
Commitment 4: testing, verification and compliance
The following core tasks must be performed to fulfil and demonstrate compliance with the obligations under Article 50(2) and (5) as well as with the commitments and measures specified in the Code:
- setting up compliance, testing, verification and monitoring processes,
- updating these processes in line with the state of the art,
- implementing the defined measures in practice and demonstrating and documenting compliance,
- verifying compliance with the marking and detection solutions.
In brief
A process to ensure compliance with the obligations must be implemented and demonstrated, for example through documentation. Continued compliance of the marking and detection solutions with all the requirements specified in the Code must be regularly tested under conditions that are representative of real-world use. Personnel and external contractors who are responsible for implementation must be provided with training about the commitments specified in the Code. Cooperation with the competent market surveillance authorities is required to demonstrate compliance.
[ENDE]
Measure 4.1: compliance process
A compliance process that describes at a high level how the different measures in this section of the Code have been implemented to ensure compliance with Article 50(2) and (5) must be documented, implemented and kept up to date in line with the state of the art.
Compliance can be demonstrated through existing processes and documentation to the extent that they cover the required measures. If solutions provided by third parties are implemented, documentation provided by those third parties can be used. The use of open-source solutions or standards must be documented. However, ultimate responsibility for compliance with Article 50(2) and (5) remains with the provider.
Measure 4.2: testing, verification and monitoring
Before a generative AI system is placed on the market or put into service, and regularly afterwards, the compliance of the system’s marking and detection solutions with the requirements specified in the section of the Code must be tested under conditions that are representative of real-world use. Downstream providers may rely on the results of testing performed by an upstream model provider or a third-party provider as long as the marking and detection solutions comply with the applicable requirements; however, the ultimate responsibility under Article 50(2) and (5) remains with the provider of the AI system.
Until recognised performance evaluation methods and benchmarks for marking and detection techniques are in place, the techniques implemented must be tested and documented using internal benchmarks and industry best practices, if applicable with the involvement of external experts or in the context of AI regulatory sandboxes.
Any compliance shortcomings identified or reported must be monitored and remedied with appropriate corrective actions.
Measure 4.3: training
Proportionate efforts must be made to provide appropriate training to personnel and external contractors who are involved in the design and development of the generative AI systems and are responsible for implementing the commitments and measures specified in the Code.
Measure 4.4: cooperation with market surveillance authorities
Cooperation with the competent market surveillance authorities is required to demonstrate compliance with Article 50(2) and (5) and the obligations. The documentation specified in Measure 4.1 and access to the marking and detection solutions must be provided following a reasoned request.
Section 2: deployers – rules for labelling deep fakes and AI-generated and manipulated text
Signatories recognise the need for consistent and effective implementation to inform individuals about the artificial origin or the role of AI systems in manipulating content. An optional EU icon provides an easy, practical and uniform pathway for compliance with Article 50(4) and (5); implementation can be adapted to take into account sector-specific good practices. Disclosure should take place at the latest at the time of first exposure to generated or manipulated content and should be easily perceivable and understandable, especially for persons with disabilities, children and the elderly.
The goal of the measures is to empower individuals to make informed decisions by enabling them to understand the origin and nature of synthetic content.
Finally, it must be acknowledged that the transparency obligations under the AI Act apply in parallel to, and do not affect, other obligations or requirements under, for example, data protection, consumer protection, digital services, media and criminal law.
Commitment 1: disclosure of deep fakes and published text
To fulfil the obligations under Article 50(4) and (5), the artificial origin of deep fakes or published text must be clearly labelled. The EU icon in Annex 1 of the Code or an equivalent icon or label that complies with the design and placement specifications must be used. The icon can be integrated into existing disclosure practices under applicable EU or national legislation or professional codes of conduct, to the extent that they comply with the Code and Article 50(4) and (5). However, the act of labelling does not constitute exemption from other laws, such as those related to the protection of personality rights or regarding obtaining consent of the depicted persons or rightsholders.
In brief
Certain design specifications must be implemented for disclosure. For visual content, an icon with the capitalised “AI” acronym (in English or, if appropriate, in the national language), such as the standard EU icon, should be used and the style (such as size, scale, colour) must ensure that the icon is clearly recognisable and readable. For audio content, there are corresponding rules relating to audible disclaimers (“earcons”). There are also certain placement specifications that ensure that labelling is clear and distinguishable directly at the time of first exposure to AI-generated or manipulated content. These specifications apply in particular across online platforms and search engines, in both offline and online environments.
[ENDE]
Measure 1.1: design specifications
The following design specifications must be implemented, depending on the category:
Category
| Visual (image, video, text) | Non-visual (audio-only content) |
|---|
Labelling type
| Icon with the capitalised acronym “AI” in the English language (or, if appropriate, in the national language)
EU standard icon (Annex 1) available to use freely
| Short audible disclaimer at the beginning of a deep fake in natural language or in the form of an earcon
EU earcon currently under development
|
|---|
Key specifications
| - Letters have the same vertical dimension; proportions preserved when resized
- Size can vary depending on content but must be clear and distinguishable
- Style (such as colour, contrast, typography) may differ as long as the icon remains readable
| - Audio disclaimer in simple language (in the same language as the content or in English)
- Disclaimer must clearly disclose the artificial origin
|
|---|
Additional options
| - Supplement to AI icon with (interactive) second layer with:
- text indicating “generated” or “modified” (or local translation)
- text or pictogram indicating what exactly has been modified by AI (for example face, text)
| - Supplement to audio disclaimer indicating whether the audio was “generated” or “manipulated”
- Indication of which elements were modified by AI (for example voice, background noise)
|
|---|
When implementing the specifications, account must be taken of the diverse composition of the audience (such as knowledge, language proficiency, vulnerability) and the context (such as financial, medical or education sector).
The labelling must be accessible and comply with applicable EU law, in particular the European Accessibility Act and the Web Accessibility Directive. This includes application of:
- audio descriptions or alternative cues for visual disclosure elements;
- tactile and haptic cues for audio-only content, considering the needs of end-users with hearing impediments (for example a vibration alert before audio playing identifying that the audio contains deep fake content);
- high contrast icons or equivalent labels and screen-reader compatibility, including for end-users with colour vision deficiencies;
- detectability of the icon or equivalent label by assistive technologies.
If applicable to the products or services, recognised accessibility standards such as ETSI EN 301 549 TSI or the W3C‑Web‑Content‑Accessibility‑Guidelines 2.1 should be implemented. Providers of VLOPs or VLOSEs subject to the obligation to put in place risk mitigation measures under Article 35(1) DSA are encouraged to use the EU Icon or another equivalently compliant solution to enable deployers of AI systems to implement their labelling obligation in accordance with Article 50(4).
Measure 1.2: placement specifications
To meet the legal requirements under Article 50(5) of labelling in a clear and distinguishable manner at the latest at the time of first exposure to published text or deep fakes, the following placement specifications should be implemented, in particular across online platforms and search engines, in both offline and online environments.
Sub-measure 1.2.1: overarching principles for placement specifications
The following overarching principles applicable to all content – with the exception of deep fakes that are part of artistic, creative, satirical, fictional or analogous works (see Commitment 3) must be applied:
- the icon or equivalent label must be placed in a manner that ensures immediate recognition without requiring user action or sustained attention;
- the icon or equivalent label must remain visible at least for a sufficient duration to be noticed under normal exposure conditions to ensure perceivability;
- the icon or equivalent label must be directly embedded into the content; alternatively, a user interface overlay that for natural persons appears to be on the content can be used;
the icon or equivalent label must be clearly perceivable and distinguishable at the latest at the time of first exposure of a user to the deep fake or published text.
Sub-measure 1.2.2: placement specifications where visual disclosure is possible
The following placement specifications should also be implemented where visual disclosure is possible:
- The icon or equivalent label should appear in an appropriate place where no intervening overlay elements exist, for example in the top right corner of an image or video.
- The icon or equivalent label should appear at the beginning of a deep fake video, after interruptions (for example after commercial breaks) and, where possible, at regular intervals; if a video does not entirely consist of deep fake content, the icon or equivalent label should appear throughout the deep fake part of the video.
- If a deep fake is exclusively used internally (for example for training), the disclosure can be placed in the user interface, physical setting or any other readily available medium before exposure to the deep fake, as long as it is clear and distinguishable.
- For visual deep fakes, audible disclosures may only be implemented as an additional disclosure method.
- For audio deep fakes, when a screen is available (for example in a car or on a smartphone), an additional visual disclosure should be displayed.
- For published text, the icon or equivalent label should be placed above or near the headline or in the colophon at the beginning of the text, or only that part of the text that is AI-generated or manipulated can be labelled. For very short text (single words or brief phrases), labelling can be through a notice in the user interface (for example an icon adjacent to the output or a notice at the beginning of interaction).
Sub-measure 1.2.3: placement specifications where visual disclosure is not possible
If visual disclosure is not possible, an additional audible disclaimer should be included at the latest at the time of the first exposure to the deep fake. Additionally, audible cues such as tones or earcons can be used to support recognition. The initial audible disclaimer should, where appropriate, be supplemented with reminders at regular intervals for the entire duration of the audio deep fake to ensure ongoing awareness.
Measure 1.3: voluntary participation in a task force under the Code
Support is encouraged for the work and activities of a dedicated task force under the Code aimed at advancing the further development and usability of the EU icon as a minimum standard and addressing issues relating to detection and marking solutions. Signatories and relevant stakeholders from various sectors, including independent researchers, civil society organisations and national authorities, are encouraged to participate in the task force.
Commitment 2: internal processes
Internal processes, awareness measures and review mechanisms should be put in place or maintained to fulfil the obligations under Article 50(4) and (5) and the measures specified in this section of the Code.
In brief
Deployers must implement compliance processes and documentation to demonstrate how the disclosure obligation is met (for example by using the EU icon). If employees or external contractors are directly involved in the implementation of disclosure measures, AI literacy should be promoted and awareness of the obligations ensured, for example through training measures. The effectiveness of the implementation of the design and placement specifications must be verified through internal and external reviews. Contact channels should be provided for reporting missing or incorrect disclosures. Reported cases of non-compliance must be reviewed immediately and any non-compliance remedied.
[ENDE]
Measure 2.1: internal compliance process
Internal compliance processes and documentation must be put in place or maintained to show how the disclosure obligations are implemented using the icon or equivalent label. The documentation can include a general description and examples. It is recommended for information on the disclosure solution applied to be published in a clear and accessible manner.
If necessary, and particularly in cases of regular use of AI systems to create deep fakes or text, processes should be put in place to ensure that the design and placement specifications are implemented properly to mitigate risks of non-labelled or incorrectly labelled content.
Media service providers within the meaning of Article 2(2) of the European Media Freedom Act can apply existing procedures and standards, to the extent that they comply with the measures within this commitment and with Article 50(4) and (5).
Measure 2.2: awareness and literacy
Proportionate efforts must be made to ensure awareness of the disclosure obligations under Article 50(4) and (5) among the personnel, including employees and external contractors directly involved in the implementation of disclosure measures or overseeing compliance with the measures in this section of the Code, for example with training, including in the context of promoting AI literacy.
Measure 2.3: review, feedback and cooperation with authorities
The effective implementation of the design and placement specifications must be supported through internal review and external feedback.
It is recommended for channels to be provided that allow individuals or third parties (such as trusted flaggers, independent researchers, academics, fact-checkers) to flag missing or incorrect disclosures, where appropriate through existing reporting mechanisms (for example trusted flagger mechanisms, interfaces for third-party fact-checking services or notice and action mechanisms). Such reported cases must be reviewed immediately and measures taken to remedy any non-compliance.
Commitment 3: disclosure for artistic, creative and similar works
In accordance with Article 50(4), deep fakes that form part of artistic, creative, satirical, fictional or analogous work must be disclosed in a way that does not hamper the display or enjoyment of the work by:
- a) using an icon or equivalent label in accordance with Measure 1.1 and placing it in a manner appropriate to the type of work and the context;
- b) ensuring that such disclosure is clear, distinguishable and accessible to all persons and provided at the latest at the time of first exposure;
- c) ensuring that the icon or equivalent label is perceivable for a sufficient duration and easily noticed under normal conditions.
Where content is made available in a digital or interactive manner (for example on websites, in apps or on other user interfaces), the icon or equivalent label can be placed outside or adjacent to the deep fake content and integrated into user interface elements. Where content is made available in a non-digital or non-interactive manner (for example exhibitions, art galleries, cinemas, festivals or similar contexts, audio or video on a physical carrier), disclosures can be provided at the point of entry or as part of the accompanying information, such as exhibition leaflets, entrance tickets and packaging.
Commitment 4: human review and editorial control for published text
Media service providers within the meaning of Article 2(2) of the European Media Freedom Act who are already subject to editorial standards and to regulatory frameworks may invoke the exception to the disclosure obligation in Article 50(4) second subparagraph by applying their existing review and editorial procedures and established professional standards.
Otherwise, appropriate policies for human review or editorial control must be established, adapted or maintained to ensure that a natural or legal person holds editorial responsibility for the publication.
These internal policies can be based on existing processes, should be proportionate to the deployer’s size and resources and must include at least the following elements:
- the identification and publication of the natural or legal person with editorial responsibility (name, role and contact details);
- an overview of the concrete organisational measures and human resources allocated to guarantee adequate human review or editorial control and ensure editorial responsibility is assumed before publication of the published text.
The implementation of this commitment must in no way affect media freedom, editorial independence or protection of journalistic source information.
Annex
The Annex to the Code contains the EU icons that can be used by deployers to comply with Measure 1.1. Three main icons are provided: one icon for fully AI-generated deep fakes or published text (Figure 1), a second icon for AI-manipulated/partially modified deep fakes or published text (Figure 2), and a third basic icon enabling deployers to supplement it with an interactive layer with further information or an alternative textual label (Figure 3).
Figure 1 The EU icon developed by the AI Office for fully AI-generated content in several variations against a background (black, white, black transparent, white transparent).
The design and development of the EU icon were subject to empirical user-testing conducted across several Member States, which evaluated the perception of authenticity, noticeability and recognition, understandability and trust. The findings and methodology of the empirical study will be made public.
Figure 2 The EU icon developed by the AI Office for partially AI-modified content in several variations against a background (black, white, black transparent, white transparent).
These icons are publicly available for signatories to use in line with the placement specifications, without the need for attribution to the European Commission or the AI Office.
Figure 3 The basic EU icon developed by the AI Office in several variations against a background (black, white, black transparent, white transparent).
The task force will explore the integration of an interactive second layer with the aim to enrich the above static icons with provenance data and information on what has been modified. The development of this interactive layer will be in alignment with the technical capabilities of deployers and other actors distributing and/or verifying content. An audio-only EU icon will also be developed.
