Why is transparency regarding content generated by AI systems so important?
Why is transparency regarding content generated by AI systems so important?
Various AI systems can generate large volumes of synthetic content, which people are finding increasingly difficult to distinguish from human-generated and authentic content. The widespread availability and increased capabilities of these systems have a significant impact on the integrity of the information ecosystem and trust in it, as new risks of disinformation and large-scale manipulation, fraud, impersonation and consumer deception.
Furthermore, some AI systems are designed for direct interaction with individuals and may pose specific risks of impersonation or deception.
Transparency in the field of artificial intelligence is therefore crucial, as it fosters trust, ensures the accountability of AI system providers and implementers, enables informed acceptance of content and decision-making, and is aligned with ethical and legal guidelines, thereby contributing to the responsible and ethical use of artificial intelligence.
Who is the provider of an artificial intelligence system, and what transparency obligations apply to them?
In accordance with Article 3(3) of the AI Act, AI system providers are natural or legal persons, public authorities, agencies or other bodies that develop AI systems or have them developed, and place them on the EU market or put them into service under their own name or trademark. This applies regardless of whether these providers are established or located in the EU or in a third country.
Provisions of the AI Act also apply to AI system providers established within or outside the EU if the outputs of their AI system are used in the EU.
Providers must ensure that their artificial intelligence systems comply with the relevant transparency obligations set out in Articles 50(1), (2) and (5) of the AI Act before placing those systems on the market or putting them into service.
In accordance with Article 50(1) of the AI Act, providers must design and develop AI systems that communicate directly with natural persons (chatbots, AI agents and avatars) with a notice stating that they are communicating with artificial intelligence.
In accordance with Article 50(2) of the AI Act, providers of AI systems that generate or modify synthetic content, such as image, video, audio or text content, must ensure that the output is labelled in a machine-readable format (e.g. watermarks, digitally signed metadata) and that such output is also recognisable as having been artificially generated or modified. They must also ensure that the outputs of their generative AI systems are labelled with effective, reliable, robust and interoperable machine-readable labels that enable the identification of outputs generated or manipulated by AI systems.
[1] Transparency obligations under Article 50 of the Artificial Intelligence Act
Who is a deployer and what transparency obligations apply to them?
Operators of AI systems are natural or legal persons, public authorities, agencies or other bodies that use AI systems within their remit, except where such use is for personal, non-professional activities.
Where a natural person uses an AI system in their own name, for example to create and disseminate fake content on social media, this is considered a personal, non-professional activity. Such use is excluded from the scope of the AI Act. However, if the activity involves regularly deriving economic benefit (or is otherwise part of a business, commercial, professional or liberal activity), it is considered a ‘professional’ activity. In this case, the natural person is considered to be a user of that artificial intelligence system.
Where the operator of the AI system is a legal person under whose authority the system is used (e.g. an advertising company), individual employees acting on the instructions and under the supervision of that legal person (e.g. digital animators, web designers, content creators, journalists) are not regarded as separate users of that system. The legal entity remains the operator even if third parties (e.g. contractors, self-employed individuals) are involved in the operation of the system on its behalf and under its responsibility and supervision.
Operators of AI systems for emotion recognition and biometric categorisation must clearly inform individuals when they are subject to the operation of these systems and, of course, ensure compliance with the GDPR.
Implementers of AI systems that generate or manipulate content constituting deepfakes (images, video, audio) or publish text intended to inform the public about matters of public interest must disclose that such content has been artificially created or altered.
Relevant Article: 50
[1] Transparency obligations under Article 50 of the Artificial Intelligence Act
What does the Code stipulate?
The Code of Conduct on the Transparency of Content Generated by Artificial Intelligence (hereinafter the Code of Conduct) was published in its final version on 10 June 2026, and the European Commission (EC) adopted an adequacy assessment for it on 9 July 2026 (more on this via the link).
The Code sets out practical and technical methods for labelling and identifying content created or modified by artificial intelligence. It focuses on the obligations of AI system providers under Article 50(2) of the AI Act and the obligations of implementers under Article 50(4) of the AI Act.
The measures envisaged include:
- protected and digitally signed metadata;
- watermarks;
- standardised icons;
- techniques for detecting content created or modified by artificial intelligence.
The Code is divided into two sections:
- The first section is intended for providers of artificial intelligence systems. It contains commitments regarding machine-readable labelling and enabling the detection of synthetic content that has been created or modified by artificial intelligence.
- The second section is intended for users of artificial intelligence systems. It contains rules for labelling deepfakes and text created or modified by artificial intelligence that is published with the aim of informing the public about matters of public interest. For the purposes of the second section of the Code, a set of ‘AI’ icons has also been developed, which implementers can use to label content created or modified by artificial intelligence.
Organisations may sign up to both sections of the Code or only to the one that is relevant to their role and activities.
Relevant article: 50(7)
[1] Code of Practice on the Transparency of Content Generated by Artificial Intelligence
[2] Icons for labelling AI-generated content
[3] Signing up to the Code for providers and implementers (with frequently asked questions and answers)
Why is the Code important for data controllers?
For providers and implementers of artificial intelligence systems, the Code primarily provides a clearer and more standardised way of demonstrating compliance. The Code of Conduct addresses the technical and practical aspects of implementing the obligations for providers and operators of AI systems set out in Article 50(2), (4) and (5) of the AI Act.
Signatories who actually implement its requirements will be able to demonstrate compliance with the obligations regarding transparency and content labelling by referring to their compliance with the provisions of the Code. The competent authorities will continue to verify compliance, but such verification may focus primarily on whether the signatory is honouring the commitments it has made.
Nevertheless, the Code is not the only possible means of ensuring compliance. Obligated parties may opt for their own technical and organisational solutions, but they will have to demonstrate that these achieve a comparable level of effectiveness and reliability.
Adherence to the Code and compliance with it may also be taken into account as a mitigating factor when determining the amount of any fine.
The European Commission has also published a list of frequently asked questions for potential signatories and other interested organisations, which further clarifies the content of the Code, the accession process and the implementation of individual commitments.
What do the guidelines set out?
The guidelines supplement the Code of Conduct on the Transparency of Content Generated by Artificial Intelligence (hereinafter the Code of Conduct), the final version of which was published on 10 June 2026 (more on this via the link).
The guidelines address Article 50 of the AI Act in its entirety, as the purpose of the guidelines is to provide providers and operators of AI systems, as well as competent authorities, with practical guidance for a uniform understanding and effective implementation of the transparency obligations for certain AI systems prescribed by the AI Act.
The guidelines constitute a non-binding legal interpretation of the entirety of Article 50 of the AI Act, the purpose of which is to clarify the content and scope of the prescribed obligations. They primarily contain explanations of legal concepts (e.g. directly interactive AI systems, synthetic content, deepfakes, text on matters of public interest) and define exceptions, such as the normal editing of content; they also clarify the role of stakeholders and address the protection of individuals’ rights. They are intended for all supervisory authorities, providers and implementers of AI systems, courts and other stakeholders to assist in interpreting obligations. These are practical guidelines for the consistent, effective and uniform implementation and understanding of these obligations.
When do the transparency obligations for generated content come into force?
Article 50 of the Artificial Intelligence Act applies from 2 August 2026. From that date onwards, providers and operators of AI systems must comply with the transparency obligations set out in that provision.
A limited transitional period is provided for only in respect of AI systems placed on the market before 2 August 2026, and only in relation to the obligation to label and disclose content generated by artificial intelligence (Article 50(2) of the Artificial Intelligence Act). Providers of such systems must comply with these obligations from 2 December 2026 onwards.
Content created before 2 August 2026 does not need to be labelled retrospectively. Nevertheless, the Commission encourages the relevant providers to do so wherever possible, as this contributes to the transparency objectives set out in Article 50 of the AI Act.
What is a ‘deepfake’ and when must it be labelled?
A deepfake refers to AI-generated or manipulated visual, audio or video content that resembles existing persons, objects, places, entities or events and which appears to the user to be authentic or real (Article 3(60) of the AI Act).
For content to constitute a ‘deepfake’, four cumulative criteria must be met:
- Similarity – a high degree of similarity between the deepfake content and the simulated subject.
- Existence – the simulated persons, objects, places, entities or events must resemble someone or something that exists, is likely to exist or could plausibly exist in reality.
- People, objects, places, entities or events – for content to constitute a deepfake, it must resemble existing people, objects, places, entities or events. Details are set out in the guidelines.
- False appearance of authenticity or truthfulness – content that would falsely appear to a person to be authentic or true.
This refers to a key characteristic of synthetic content and its potential to mislead a person regarding the authenticity or reality of the content.
To assess the last criterion above, the Guidelines on the Transparency of Content Generated by Artificial Intelligence allow stakeholders to take into account:
- the degree of similarity,
- the potential message conveyed by the content,
- the anticipated and foreseeable contexts of use,
- the intended and reasonably foreseeable composition of the audience and their expectations.
Providers of AI systems that generate or adapt content constituting deepfakes (images, video, audio) or publish text intended to inform the public about matters of public interest must disclose that such content has been artificially created or altered.
The guidelines contain examples of when content constitutes deepfakes and when it does not, and of when matters of public interest are involved in the creation of text intended to inform the public (e.g. matters relating to public services, consumer protection, democratic processes, fundamental rights) and when AI-generated texts do not need to be disclosed (if they are subject to human verification or editorial review, and where a natural or legal person bears editorial responsibility for the publication of the content) and the conditions for this.
The guidelines also clarify when an exception applies to content that forms part of a clearly artistic, creative, satirical, fictional or similar work or programme (where transparency obligations are limited to disclosing the existence of such created or adapted content in an appropriate manner that does not hinder the display or enjoyment of the work) and when such an exception does not apply.
[1] Transparency obligations under Article 50 of the Act on Artificial Intelligence
Which text generated or adapted by AI must providers clearly label?
In accordance with Article 50(2) of the AI Act, providers of AI systems that generate or adapt textual content shall ensure that the output is labelled in a machine-readable format (e.g. watermarks, digitally signed metadata) and that such outputs are also identifiable as having been generated or adapted by AI. The outputs of their generative AI systems must be labelled with effective, reliable, robust and interoperable machine-readable labels that enable the identification of outputs generated or manipulated by AI systems. These include:
- text in containers (‘Containerised text’): For text embedded in structured file formats (e.g. PDF, Word or HTML), providers must include origin information in digitally signed and time-stamped metadata.
- free-form text: For unformatted free-form text longer than 200 tokens, providers must mark the content with an imperceptible watermark.
Providers are not required to mark the text in the following cases, for example:
- Very short texts: Texts shorter than 200 tokens, as the technology does not yet allow for reliable watermarking.
- Standard editing and minor corrections: The use of a text editor to check spelling, correct grammar, apply technical formatting or make stylistic corrections that do not substantially alter the content, meaning or style of the text.
- Source code and technical labels: Programme code (scripts, SQL, JSON, YAML), output data for machine-to-machine communication, and individual words, image captions or user interface elements.
- Criminal prosecution: UI systems that have a legal basis for the detection, prevention and investigation of criminal offences.
Relevant Article: 50
[1] Transparency obligations under Article 50 of the Artificial Intelligence Act
[3] Code of conduct on the transparency of content generated by artificial intelligence
Which text generated or modified by AI must operators clearly label?
In accordance with Article 50(4) of the AI Act, operators of AI systems must clearly label text that has been generated or modified by artificial intelligence and is published for the purpose of informing the public on matters of public interest. For text to fall within the scope of this obligation, three criteria must be met.
The text must be:
- published
- intended for the public (e.g. published on freely accessible websites)
- on matters of public interest, such as:
- politics and democratic processes,
- public administration and services,
- the judiciary and law enforcement,
- fundamental rights,
- public safety,
- public health,
- environmental protection,
- consumer safety
- any economic, financial, political, scientific or cultural development that may be a matter of public debate.
However, publishers are not required to label the published text if any of the following exceptions apply:
- Human review and editorial responsibility: If the text has undergone a content review or editorial oversight by a human being prior to publication (e.g. fact-checking) and a specific legal or natural person assumes editorial responsibility for the publication (e.g. traditional media, professional blogs with an editorial team, officially verified statements).
- Personal and non-professional use: Natural persons who create or publish texts solely for private or non-commercial purposes.
- Unpublished texts: Private or business correspondence between individuals and internal communication within organisations (e.g. posts on an intranet).
- Content not relating to the public interest: fantasy novels, product descriptions in shops (without claims relating to health, consumer protection or environmental claims) or summaries of news items generated in a chat room that are visible only to the user who issued the prompt.
- Criminal prosecution: The publication of text permitted by law for the purposes of investigating and preventing criminal offences.
Relevant Article: 50
[1] Transparency obligations under Article 50 of the Artificial Intelligence Act
[3] Code of conduct on the transparency of content generated by artificial intelligence
We have recently started using AI tools to create content for our website, so we are interested in the current legal framework regarding this in Slovenia. We would like to know whether we are required to include information or notes regarding content created using AI methods (text, graphics, video, etc.)?
From 2 August 2026, content created using AI tools will need to be labelled. To make it easier to understand how to comply with the transparency obligations for generated content, the European Commission has issued Guidelines on the implementation of transparency obligations for certain artificial intelligence systems under Article 50 of Regulation (EU) 2024/1689 of the Artificial Intelligence Act and a Code of Conduct on the transparency of content generated by artificial intelligence. Both documents will assist both providers and implementers in fulfilling the transparency obligations set out in the AI Act, and will help the competent authorities in monitoring the labelling of AI-generated content.
Further details on the transparency of AI-generated content are provided in the other frequently asked questions.
Relevant Article: 50
[2] Code of conduct on the transparency of content generated by artificial intelligence