AI labelling obligations: how to label chatbots, deepfakes and other AI content correctly

What do the AI labelling obligations under Art. 50 AI Act require, and since when have they applied?

Since 2 August 2026, Art. 50 of the AI Act (Regulation (EU) 2024/1689) has set out various transparency obligations for providers and deployers of AI systems. In common usage, they are usually summarised as the AI labelling obligation. Three constellations are particularly relevant for companies: interactive AI systems such as chatbots, AI-generated or AI-manipulated content including deepfakes, and the use of emotion recognition or biometric categorisation.

Chatbots and interactive systems

Providers of chatbots and other AI systems that interact directly with natural persons must ensure that users can recognise that they are communicating with an AI system.

As a rule, the information must be provided in a clear and distinguishable manner at the latest at the time of the first interaction. It may be omitted if the AI nature is obvious to a reasonably well-informed, observant and circumspect person, given the specific circumstances. According to the EU Commission's guidance, this exception is to be interpreted narrowly. In particular, the context of use and the group of persons addressed must be taken into account.

The obligation lies with the provider of the AI system. If you integrate an external solution, you should therefore check which company acts as provider and which as deployer in the specific case.

Deepfakes and synthetic content

For AI-generated or AI-manipulated audio, image, video or text content, Art. 50 first of all contains a technical labelling obligation. Providers of such AI systems must ensure that their outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. As far as technically feasible, the marking must be effective, interoperable, robust and reliable.

Note: For AI systems that were placed on the market before 2 August 2026, a transitional period until 2 December 2026 applies to the technical labelling obligation under Art. 50(2).

In addition, deployers have disclosure obligations for certain published content. This concerns in particular deepfakes and AI-generated or AI-manipulated text published with the purpose of informing the public on matters of public interest.

A deepfake is AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and could falsely appear to be authentic or truthful. Whether these conditions are met depends on the specific content and its context.

For artistic, creative, satirical or fictional works, less stringent requirements apply to the manner of disclosure. For texts on matters of public interest, the disclosure obligation does not apply if the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility.

Emotion recognition and biometric categorisation

Deployers of AI systems for emotion recognition or biometric categorisation must, as a rule, inform the persons exposed to them about the operation of the system.

In addition, the data protection requirements for the respective processing of personal data continue to apply.

Who is required to label AI content: provider or deployer?

Art. 50 distributes the transparency obligations between providers and deployers of AI systems. The respective role determines which obligations a company must fulfil itself.

• Provider: Anyone who develops an AI system or has it developed and places it on the market or puts it into service under their own name or trade mark is a provider. This applies regardless of whether the company is established inside or outside the EU, provided the conditions for the application of the AI Act are met.
• Deployer: Anyone who uses an AI system under their own authority in the course of a professional activity is a deployer. Purely personal, non-professional use is excluded.
• Freelancers and self-employed persons: Natural persons can also be deployers if they use an AI system in the course of a business, professional or freelance activity. According to the guidance, use that is regularly aimed at economic gain may in particular indicate this.
• Employees: Employees who use an AI system under the responsibility and control of their employer do not, as a rule, become deployers themselves as a result. The deployer role generally lies with the company.

In the case of an externally developed chatbot, a company may therefore be the deployer of the system while the manufacturer remains the provider. Depending on the setup, however, a company may also assume provider obligations itself, for example if a system is put into service under its own name or trade mark.

You should therefore check the allocation of roles separately for each use of AI.

Which content is not covered by the technical labelling obligation?

The technical labelling obligation under Art. 50(2) does not cover every output of an AI system. The EU Commission's guidance specifies several categories of cases that fall outside its scope.

These include in particular:

• short sequences of numbers, symbols or letters without independent synthetic content,
• source code,
• outputs intended exclusively for automated machine-to-machine communication and not accessible to any natural person,
• certain outputs used exclusively within closed industrial or product-related development environments,
• standard editing where the AI system performs an assistive function, as well as editing that does not substantially alter the input data or its semantics.

Pure spelling or grammar correction, for example, may fall under the exception for standard editing. In the case of more extensive changes, the scope and effect of the AI editing are decisive.

There is no general exception for artistic, creative, satirical or fictional works. Where a deepfake is involved, however, the disclosure can be made in a way that does not hamper the display or enjoyment of the work.

What exactly must AI labelling look like?

The specific implementation depends on whether technical marking by the provider or disclosure to natural persons is required.

What applies to machine-readable marking?

Providers of generative AI systems must design their systems so that the relevant outputs are marked in a machine-readable format and detectable as artificially generated or manipulated.

Art. 50 does not prescribe a specific technical standard for this. In practice, provenance information and technical provenance techniques in particular can be used. Standards such as C2PA may play a role here, but the AI Act itself does not require their use.

For implementation, the EU Commission has published the Code of Practice on Transparency of AI-generated Content. It specifies suitable measures for technical marking and detectability. Participation in the Code is voluntary.

The Commission and the AI Board have recognised it as a suitable instrument for demonstrating compliance with the transparency obligations under Art. 50(2), (4) and (5). Other suitable compliance solutions remain possible. However, adherence to the Code does not in itself constitute conclusive proof of compliance.

What must disclosure to people look like?

For deepfakes and texts subject to labelling, the information must be perceptible to natural persons. It must be provided in a clear and distinguishable manner at the latest at the time of first exposure.

The Code of Practice contains more detailed requirements on design. The EU Commission also provides three different icons: a general symbol for AI involvement, a symbol for fully AI-generated content and a symbol for partially AI-manipulated content. Each of the icons is available in several display variants.

Use of the EU icons is voluntary. Other suitable forms of disclosure can also meet the requirements of Art. 50.

When using the EU icons, the Commission recommends in particular a clearly visible placement, an understandable additional text label and an accessible design. The icon should be visible at the latest at the time of first exposure and, as a rule, should also be retained when the content is shared or downloaded.

If a company uses third-party AI tools, the technical marking therefore generally lies with the provider, while disclosure of deepfakes and certain texts lies with the company as deployer.

Which deadlines apply to the AI labelling obligation?

The transparency obligations under Art. 50 have generally applied since 2 August 2026.

For technical marking under Art. 50(2), there is a transitional rule for existing systems. Providers of AI systems that generate synthetic audio, image, video or text content and that were placed on the market before 2 August 2026 must meet the requirements of Art. 50(2) by 2 December 2026 at the latest.

The other transparency obligations under Art. 50 are not covered by this transitional rule.

According to the guidance, there is no retroactive labelling obligation for content that was generated before 2 August 2026. The Commission recommends subsequent labelling where this is practically feasible.

The transitional period until 2 December 2026 therefore applies solely to technical marking by providers of existing systems.

Practical catalogue: five typical use cases of the AI labelling obligation in companies

The following five situations are typical in medium-sized and large companies. They show how the abstract obligations under Art. 50 translate into specific AI applications.

What else changes from 2 December 2026 in addition to the labelling obligation?

On 2 December 2026, two new prohibitions introduced into Art. 5 of the AI Act by the Digital Omnibus on AI come into force. The first concerns certain AI systems for generating or manipulating realistic non-consensual intimate depictions of identifiable persons. The second covers certain AI systems for generating or manipulating depictions of child sexual abuse.

The prohibitions concern whether the respective AI practice is permissible and apply alongside the transparency obligations under Art. 50. They are of practical relevance in particular for providers and deployers of image, video or audio generation systems.

What happens in the event of infringements of the labelling obligation?

Infringements of Art. 50 can be punished with fines of up to EUR 15 million or, for companies, up to 3 per cent of total worldwide annual turnover in the preceding financial year, whichever is higher. For SMEs, including start-ups, the lower of the two maximum amounts applies in each case. Special caps also apply to small mid-caps.

In Germany, the Federal Network Agency (Bundesnetzagentur) is generally responsible for market surveillance. Special responsibilities exist for individual areas, for example for certain regulated products and the financial sector.

Conclusion: check the AI labelling obligation in three steps

When implementing the AI labelling obligation, what matters most for companies is the specific use of the AI system and their own role as provider or deployer.

Providers must in particular ensure transparency for interactive AI systems and the technical marking of generated content. Deployers are subject to information and disclosure obligations for emotion recognition and biometric categorisation, as well as for deepfakes and certain AI-generated texts.

Three steps are recommended for the review in your company:

1. determine your role as provider or deployer,
2. identify the relevant use case under Art. 50,
3. define the required technical marking or disclosure.

The EU Commission's guidance and the Code of Practice specify the requirements and provide additional orientation for practical implementation.

Frequently asked questions

#1 Do the labelling obligations also apply to AI-assisted texts or only to fully AI-generated content?

That depends on whether the text is published on matters of public interest and whether its content has been reviewed by a human beforehand.

Art. 50(4) covers AI-generated or AI-manipulated text published with the purpose of informing the public on matters of public interest. Disclosure is not required if the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for the publication.

According to the guidance, the human review must relate to the content. A mere spelling, grammar or formatting check is not sufficient for this. The disclosure obligation therefore depends on the type of AI use, the purpose of the publication, any human review or editorial control and whether editorial responsibility has been assumed.

#2 Do I also have to label AI-generated images on social media?

Visible disclosure is required if the published image is a deepfake. Technical marking generally lies with the provider of the AI system.

Providers of such AI systems must, as a rule, ensure that the generated content is marked in a machine-readable format and detectable as artificially generated or manipulated. The publication channel used, such as a website or a social network, generally does not change this classification.

#3 Do we have to retroactively label AI content generated before 2 August 2026?

For content generated before 2 August 2026, there is generally no obligation to label it retroactively. The EU Commission recommends voluntary subsequent labelling where this is practically feasible.

#4 Does the labelling obligation also apply to small and medium-sized enterprises?

Yes. Art. 50 AI Act generally provides no exemption from the transparency obligations based on company size. Special rules for small and medium-sized enterprises exist in particular regarding the calculation of possible fines.

#5 What has the Digital Omnibus changed about the AI labelling obligation?

The transparency obligations under Art. 50 have generally applied since 2 August 2026. The transitional rules amended by the Digital Omnibus on AI mainly concern high-risk AI systems: their requirements generally apply from 2 December 2027 for Annex III and from 2 August 2028 for Annex I.

For the technical marking of AI systems placed on the market before 2 August 2026, a transitional period until 2 December 2026 applies.

#6 Is a general notice in the legal notice or the general terms and conditions sufficient as labelling?

A general notice in the legal notice (Impressum) or the general terms and conditions is generally not sufficient. The information must be provided in a clear and distinguishable manner at the latest at the time of first interaction or exposure and must be sufficiently linked to the respective AI use or content.