The EU has begun to implement rules on labeling AI content: What editorial teams need to know.

The new rules in the EU do not mean that every journalistic text that has utilized artificial intelligence (AI) in its preparation must automatically be labeled. It is crucial whether the content has undergone substantial human review or editorial control and whether a physical or legal entity has assumed editorial responsibility for its publication. However, deepfake content must be clearly labeled even when it has passed editorial control.

Information about the use of AI must be clear, easily noticeable, and accessible to the audience at the latest during their first encounter with the content.

The Code of Practice is not law, but it shows how the law is applied

To facilitate the implementation of new obligations for providers and users of generative systems, the European Commission has published the Code of Practice on Transparency of AI-Generated Content. The document was prepared by independent experts in a process that involved providers of generative systems, labeling and detection technique experts, civil society organizations, the academic community, and other interested parties, with the support of the EU Office for Artificial Intelligence.

The Code is voluntary, but the obligations under Article 50 of the Act are not. This means that companies and organizations do not have to adhere to the Code, but they must demonstrate that they meet the transparency rules in another equally appropriate manner. The European Commission and the European Artificial Intelligence Board have assessed the Code as a suitable practical framework for demonstrating compliance, thus providing its signatories with greater legal certainty and predictability across all EU member states.

Not every use of AI warrants a label

There is an important exception for texts. Text generated or modified using AI, published to inform the public about issues of public interest, does not need to be labeled if it has undergone appropriate human review or editorial control and if a physical or legal entity has taken editorial responsibility for its publication.

This is particularly significant for editorial offices that use AI as an aid in research, data processing, transcription, translation, or language editing. The mere use of a tool at some stage of the work does not automatically create an obligation to label the final text. It is decisive that the journalist or editor verifies the facts, reliability of sources, context, and meaning of the content, and that the editorial office assumes responsibility for the published material.

However, the European Commission emphasizes that superficial checks are not sufficient. Spell-checking, grammar checking, or technical appearance of the text is not considered substantial human review or editorial control. The editor must have the authority to approve, modify, or reject the content based on its substance, alongside verifying the information and credibility of sources.

This exception applies to text but does not exempt the editorial office from the obligation to label deepfakes. If it publishes a photograph, audio, or video that accurately mimics an existing person, object, place, organization, or event using AI and could falsely appear authentic, the editorial office must clearly state that the content has been generated or altered. When a deepfake is part of an obviously artistic, creative, satirical, or fictional work, a label should still exist but can be positioned in a way that does not interfere with the display or experience of the work.

The label must be visible at the first encounter with the content

The Code stipulates that the label should be clear, legible, noticeable, and accessible to various audience groups. For photographs or videos, it may be placed directly on the content, while for videos, it should be displayed at the beginning, and in longer formats, preferably throughout the video, especially after advertisements or other interruptions. For audio content, the notification should be heard at the beginning, while for text, it should be placed above the text, alongside the title or in another location that the reader will see before they start reading.

It is also recommended that the editorial office explains whether the content as a whole was generated or just modified and, when possible, which part has been changed, such as a face, voice, or background. The European Union has prepared a set of free icons for fully generated and partially modified content, but their use is not the only permitted method of labeling.

However, the label does not indicate whether the content is true. It communicates its origin and manner of creation to the audience but does not serve as a confirmation of accuracy nor proof that the labeled content is misleading.

Editorial offices need internal rules

The Code recommends documented procedures through which the organization will determine when content should be labeled, which label is used and what it means, where it is placed, who checks whether the labeling is correctly implemented, and how to correct any omitted or incorrect labels. The rules should cover employees and external collaborators, as well as training, procedures for reporting omissions, and accountability for addressing them.

In Serbia, a specific law that comprehensively regulates the development and use of artificial intelligence, including its application in the media, has not yet been adopted. The Code of Journalists of Serbia, however, requires that the media use AI transparently, responsibly, and proportionally, to remain fully accountable for the published content, and to inform the public when content has been created using AI-based tools.

The national law on artificial intelligence in Serbia is currently in the preparation and harmonization phase with European Union standards. Although Article 50 of the Act has not yet become part of domestic legislation for local use, Serbia is expected to adopt identical transparency rules through the harmonization process in the near future.

The absence of a specific law should therefore not be a reason for editorial offices to delay establishing rules. Internal policies can predefine which tasks allow the use of AI, which data must not be entered into external tools, how results are verified, when content is labeled, and who makes the final decision on publication. Editorial offices can also introduce a higher standard of transparency than the legal minimum.

The Independent Association of Journalists of Serbia has adopted a Policy on the Use of Artificial Intelligence at NUNS and prepared a model document tailored for editorial offices. An example of the policy is freely available for download on the NUNS website, and each editorial office can adapt it to its size, mode of operation, technical capacities, and types of content it produces.

Source: NUNS