From 2nd August 2026, the AI Act regulations regarding transparency obligations for the use of artificial intelligence come into force. In practice, this means, among other things, the obligation to inform users when they are interacting with an AI system and to label certain content generated or altered with the help of artificial intelligence.
These regulations impose obligations on both providers of specific AI systems and entities that use them. In practice, they are particularly relevant for businesses using chatbots and companies utilising generative AI in marketing, advertising, media, e-commerce, or when creating texts, graphics, audio recordings, and video materials.
This does not mean that from 2nd August every graphic generated in Midjourney, text prepared with the help of ChatGPT, or photo retouched using AI requires a special label. The scope of obligations depends on how the AI was used and what the final output is.
The user should know they are talking to AI
One of the obligations laid down in the AI Act is to ensure that people using AI systems intended for direct interaction with users know that they are communicating with artificial intelligence. This obligation rests on the providers of such systems and applies, amongst others, to customer service chatbots, voicebots, virtual assistants, or AI agents.
The information should appear no later than at the first interaction with the user. Therefore, if a customer opens a chat on an online shop’s website and starts a conversation with a bot, they should know from the outset that there is no human on the other side.
Above all, this means verifying whether the implemented system actually transmits the appropriate information to the user and whether it does so at the right moment. Hiding a mention of AI solely within the terms and conditions or the privacy policy may fail to achieve the purpose of these regulations.
Labelling content generated by AI
The mere fact of using artificial intelligence does not automatically trigger an obligation to label the material. A text prepared with the help of ChatGPT must be approached differently from a realistic human image generated for an advertising campaign, and differently again from a photograph where AI has merely improved sharpness or removed noise.
In the case of images, audio recordings, and video materials, it is of particular importance whether the created content can be classified as a deepfake. According to the AI Act, this refers to content generated or manipulated by AI in the form of image, sound, or video that resembles existing persons, objects, places, entities, or events and may falsely appear authentic or true. The disclosure obligation applies precisely to such content.
A deepfake does not necessarily mean only a video with a politician’s or celebrity’s face swapped. In practice, the problem may also arise in much more everyday applications of generative AI: a realistic human image generated by AI, a synthetic voiceover, a generated realistic scenery, or a product photo placed by AI into a new environment.
However, not every AI intervention in a material will lead to such a result. Tools utilising artificial intelligence are nowadays part of ordinary photo, graphic, or sound editing software. Quality enhancement, noise reduction, or other technical actions should not automatically be treated the same way as creating a new, realistic situation that did not actually take place in reality.
What should the labelling look like?
The information about the use of AI should be clear and noticeable to the recipient no later than upon their first contact with the given content. It is therefore not about hiding information in metadata, terms and conditions, or at the end of a long description that the recipient might never read.
In the case of graphics, the label can be placed directly on the image or presented in an equivalent manner within the interface. For video material, it should be visible to the person watching the material, and the method of labelling audio content should allow the recipient to know that the heard material has been generated or modified by AI.
The European Commission has prepared a set of EU icons to facilitate content labelling. They distinguish, amongst others, between materials generated entirely by AI and content that existed previously but was subsequently partially modified using artificial intelligence. For instance, a realistic image generated from scratch will fall into the fully generated content category, whereas a real photograph of an empty flat that AI subsequently “furnished” is given by the Commission as an example of partially modified content.
EU icons for labelling AI content prepared by the European Commission:
The use of EU icons is voluntary. One can also apply their own understandable designation, such as “image generated by AI” or “voice generated using AI”. Choosing one’s own label does not exempt one from the necessity of ensuring that the information is legible and meets the requirements of the AI Act.
Korzystanie z unijnych ikon jest dobrowolne. Można zastosować również własne, zrozumiałe określenie, np. „obraz wygenerowany przez AI” czy „głos wygenerowany przy użyciu AI”. Wybór własnego oznaczenia nie zwalnia z konieczności zadbania o to, aby informacja była czytelna i spełniała wymagania AI Act.
Texts prepared using AI
The rules regarding texts look different. The AI Act does not introduce a general obligation to label every email, product description, LinkedIn post, or article just because generative AI was used in preparing it.
The obligation applies to texts generated or manipulated by AI and published with the aim of informing the public on matters of public interest. In practice, these might be, for example, materials concerning politics, security, public health, the environment, or other matters significant from the point of view of public debate. Ordinary sales or marketing texts will not, as a rule, automatically fall into this category.
At the same time, the AI Act provides an important exception. Labelling is not required if the AI-generated text has undergone human review or editorial control, and editorial responsibility for the publication is assumed by a natural or legal person.
However, such control should not be reduced solely to correcting typos and commas. If AI prepares material concerning matters of public interest, human verification should be of a genuine nature, meaning it should encompass content assessment, fact-checking, and the final decision to publish.
Responsibility for labelling
The AI Act distinguishes between the obligations of AI system providers and the entities that use those systems. Providers are responsible, amongst others, for certain technical mechanisms enabling the detection of content generated or modified by AI, whilst deploying entities face obligations such as disclosing deepfakes and specific texts concerning matters of public interest.
In practice, cooperation with advertising agencies, production studios, or freelancers requires special attention. If materials are prepared by an external entity, it is wise to establish in advance whether and to what extent they may use AI, as well as how content requiring labelling will be identified.
Technical content labelling by AI providers
As for the obligations of AI system providers, these primarily concern the technical labelling of content generated by those systems. On the providers’ side, the obligation is therefore primarily of a technical nature and does not have to be visible to the recipient. Article 50(2) of the AI Act requires systems generating synthetic text, image, audio, or video to mark the outputs in a machine-readable format, enabling the detection that they have been generated or manipulated by AI. The regulations do not mandate a single technology: these could be, for example, digital watermarks, provenance metadata, cryptographic solutions, or fingerprinting.
For systems placed on the market before 2nd August 2026, however, the AI Act provides a transitional period; their providers have until 2nd December 2026 to adapt to this obligation. For systems placed on the market from 2nd August 2026, this obligation applies from the moment of their introduction.
AI system providers are already beginning to implement such solutions in practice. An example is Anthropic, which announced the use of two types of markings in Claude: user-invisible watermarks embedded in the generated text and digitally signed provenance metadata for supported graphic files. In the case of text, this is not about adding invisible watermarks or hidden information to a ready-made response. The mechanism operates already at the generation stage. The way the model chooses subsequent words creates a specific pattern that the user is unable to notice during normal reading, but which can later be detected using an appropriate tool, which Anthropic has announced it is already working on. The tool is intended to allow verification of whether the pattern characteristic of Claude is present in the analysed text, and thus assessment of whether the content was processed by the model. However, the result of such an analysis itself will not constitute unequivocal proof of the text’s origin.
Content published before 2nd August 2026
Companies do not have to revisit all materials generated and made available before 2nd August 2026 solely to label them in accordance with the new rules. The Commission has confirmed that the obligation is not retroactive in this respect, although it encourages voluntary labelling of earlier content where possible.
It is worth approaching materials prepared earlier but still used after the new regulations come into force with caution, however, especially in actively run campaigns. In such cases, the manner of applying the rules may require an assessment of the specific scenario.
Sanctions
Infringing transparency obligations can incur administrative fines reaching up to EUR 15 million or 3% of the company’s total worldwide annual turnover for the preceding financial year.
However, a fine under the AI Act is not the only risk associated with the improper use of AI-generated content. Unlabelled material may also cause reputational issues or lead to allegations related to misleading communication.
On top of this, entirely different areas of law may apply. Using a specific person’s likeness may require analysis from the perspective of personality rights, processing personal data from the GDPR perspective, and using other people’s materials when creating content may raise copyright questions. Simply adding an “AI Generated” label does not automatically mean that the way the material is used complies with the law.
Transparency of AI usage
The AI Act does not prohibit the use of generative artificial intelligence in marketing, customer service, or content creation. The new regulations are primarily intended to ensure that, in specific situations, the recipient knows they are in contact with artificial intelligence or that the material presented to them has been generated or modified with its help.
For enterprises, this means above all the need to organise how they use AI: determining where synthetic content is created, who is responsible for its assessment, and how information obligations will be fulfilled in practice.
Author:
Wojciech Kulig
He supports clients in matters related to new technology law, in particular in the areas of personal data protection, intellectual property and issues related to artificial intelligence.
