Labeling
of
AI content.
Labeling
of
AI content.
of
Ab dem 02.08.2026 gelten die Transparenzpflichten der KI-Verordnung. Aber wann und wie muss man KI generierte Inhalte kennzeichnen? Die EU-Kommission gibt eine Handreichung und stellt EU-Icons zur Kennzeichnung bereit.
Labeling AI-Generated Content
The transparency requirements of the AI Regulation will take effect on August 2, 2026. Under these requirements, certain content generated by AI or modified using artificial intelligence must be labeled as such. In this context, “labeling” means making it clear to people that a piece of content was generated or modified by artificial intelligence.
The AI Regulation assigns this task to two roles. Providers of AI systems must technically mark their output—that is, in a machine-readable format using watermarks and metadata—so that AI-generated content can be identified later. Operators—that is, individuals who publish content generated or modified by AI—must, in certain cases, disclose this information in a way that is visible or audible to the audience. This article focuses solely on this obligation, which applies to any company or individual that distributes such content.
What types of content actually need to be labeled?
Not all AI-generated content is subject to this requirement. There are two categories covered. The first category concerns so-called deepfakes. These are AI-generated or altered image, audio, or video content that resembles existing people, objects, places, institutions, or events and would appear to viewers to be genuine or truthful. The second category concerns AI-generated or AI-altered text that is published to inform the public about matters of public interest. However, this applies only if the text was not subject to human review or editorial control and no natural or legal person has assumed editorial responsibility for it.
The Three New EU Icons
As part of the Code of Conduct on the labeling and tagging of AI-generated content, the European Commission has now provided three icons for labeling purposes.
This icon displays only the abbreviation “AI” and serves as a building block when AI was involved in the creation of a piece of content or when a separate text label or an interactive second layer is to be implemented. It is used, for example, when a deepfake video is accompanied by a supplementary text label such as “voices generated with” placed before the icon.
This icon indicates content that was generated entirely by AI, with no human-created elements and no human editorial oversight other than prompting. It is suitable, for example, for fully AI-generated deepfake videos, music or art composed entirely by AI, and AI-generated news summaries.
This icon identifies existing, human-created content that has been partially altered using AI, thereby becoming a deepfake or text about matters of public interest. Typical examples include replacing a face in a real photograph or adding furniture to an authentic image of an empty apartment after the fact.
You can download ZIP files containing all icons in all variations in SVG and PNG formats.
Are the new EU icons mandatory?
No, the EU icons are merely a suggestion and not mandatory. The only requirement is the disclosure itself. The Code of Conduct, which provides the icons, is voluntary and is binding only on its signatories. Even for signatories, compliance with the Code does not constitute conclusive proof of legal compliance. Those who sign the Code commit to implementing the disclosure either through the EU icon or an equivalent icon or label that meets the design and placement requirements. Thus, even for signatories, the EU icon is only one of several permissible methods of implementation. It offers a simple and uniform way to fulfill obligations but does not replace the operator’s independent responsibility. However, it makes sense to use these icons, as one can always argue that they meet the requirements.
The General Placement Principles
The same principles apply to all types of media. The labeling must allow for immediate recognition without interaction and without sustained attention. It must remain visible at least until it can be noticed under normal viewing conditions. Furthermore, it must be directly embedded in the content unless an equivalent alternative—such as an overlay in the user interface—is available. It must be clearly recognizable and distinguishable no later than the moment it is first perceived.
The general disclosure principles do not apply to deepfakes that are part of an obviously artistic, creative, satirical, or fictional work. Instead, a more flexible framework applies, one that balances disclosure with the undisturbed enjoyment of the work. While the artificial origin must still be disclosed, it must be done in a way that does not impair the presentation, normal exploitation, or enjoyment of the work. The disclosure must be clear, distinguishable, and accessible, and must occur no later than the first time the work is viewed, for example in accompanying notes, in the description, or in the opening or closing credits.
The specific placement of the disclosure depends on the particular distribution channel. For digital or interactive formats—such as on websites, in apps, or other user interfaces—the icon or label may be placed outside but adjacent to the image, video, or audio field, or embedded within user interface elements, as long as it remains visible without requiring any additional action on the part of the user. Possible options include brief notes, a discreet icon that displays additional information when clicked or hovered over, or an adjacent disclaimer. For non-digital presentations—such as in exhibitions, galleries, movie theaters, at festivals, or for works on physical media—disclosure at the point of entry or sale is sufficient, for example on an exhibition flyer, a ticket, or the packaging.
Image
For images, visual disclosure is possible. The icon or a similar label should be placed in a suitable location—such as the upper-right corner—without any overlapping elements. It is important that the chosen label be clear, accessible, and understandable.
Video
Visual disclosure is also possible for videos, although additional time-related requirements apply. The label must be displayed at the beginning of the video and, to the extent possible, repeated at regular intervals. It should be displayed at least after interruptions, such as commercial breaks. This takes into account the fact that different people may view a video at different times—for example, with live content—and that excerpts may be shared further. Where possible, it is recommended that the label be displayed throughout the video or at least during the manipulated segment. To date, there is no explicitly separate, stricter rule applicable solely to live videos.
Sound
Pure audio content is probably the only case in which visual disclosure is not possible. The decisive factor here is not whether a sound originates from AI, but whether it falls under the definition of a deepfake—that is, whether it bears a deceptive resemblance to a real person, voice, or event. If the requirement applies, a brief audible notice in plain language must be provided at the beginning, either in the language of the content or in English. For long-form or live content, this notice must be supplemented by reminders at regular intervals, at least following interruptions. If a screen is also available—for example, in a vehicle or on a smartphone—a visual disclosure via an icon must be provided in addition to the audible notice. Instead of a spoken notice, other audible solutions such as an earcon (a short, designed sound signal that conveys a specific meaning) are also permitted, provided they are accompanied by explanatory measures.
AI Music
For AI-generated music, the assessment involves several steps. First, one must determine whether a deepfake exists at all. Music composed entirely by AI with no connection to a real person or voice does not meet the definition of a deepfake and, according to the code, does not trigger any labeling obligation on the part of the operator. The exception for artistic works does not apply in this case, because there is no obligation from which to exempt the operator. The situation is different if a real voice is replicated or a genuine recording is altered in such a way that it can be deceptively attributed to a real person. This is likely to be the norm. In that case, an audio deepfake exists, and the disclosure requirement applies.
If this audio deepfake is also part of an obviously artistic work, the obligation is limited to appropriate disclosure that does not impair the enjoyment of the work. The audible notice does not necessarily have to be embedded in the song itself, but can be included, for example, in the track description, in the credits, or at the beginning of the track. In the case of music released via streaming services such as Spotify or similar platforms, it can generally be assumed that the content constitutes an artistic work, so this simplified regime should apply.
Text
For published text on matters of public interest, visual disclosure is possible and mandatory, unless an exception applies. The label must be placed in a clearly visible location, such as above or at the beginning of the text, near the headline, or in the header. Where appropriate, only the AI-generated or modified portion may be labeled. For very short texts where labeling within the text itself would impair readability, the requirement remains in effect but may be fulfilled through a contextual note in the user interface.
This requirement does not apply to text that has undergone human review or editorial oversight and for which a natural or legal person bears editorial responsibility for its publication.
Conclusion
The icons provided by the EU are a useful resource and save each user the trouble of developing their own solution.
At the same time, however, some questions remain unanswered. The main component of the label is the English abbreviation “AI.” The use of the respective national language is only intended when national language regulations exclude English. It remains to be seen whether the English abbreviation “AI” will be sufficient and understood in all EU member states. The code does not yet provide for a truly multilingual adaptation.
For audio-only content, there is currently no ready-made, EU-wide audio solution, nor is there an audio-only version of the icon; both are apparently still in development. The timeframes for videos remain somewhat vague, with phrases such as “as far as possible.”
We are happy to
advise you about
tHEMA ki!
