An advertising image from an image generator, a voiceover from a voice tool, a chatbot that answers customer questions in the evening: This is everyday life for many marketing and communications teams. And every time, the same question arises: Do we need to label this, and if so, how?

The short answer: In Switzerland, there is no general requirement to label AI-generated content. However, you must not use it to mislead others. Fair trading laws prohibit misleading statements, and the FDPIC interprets the Data Protection Act to mean that chatbots and altered images or voices of identifiable individuals must be recognizable. Furthermore, as soon as your content is used in the EU, Article 50 of the EU AI Act applies.

On December 2, 2026, the next step will take place: The transition period for technical labeling of existing AI generators will end, and new bans on intimate deepfakes will take effect. Our articles on the Swiss AI Act and the EU AI Act for Swiss companies provide an overview of the legal situation.

Do I have to label AI-generated content in Switzerland?

Not in general. As of October 2026, there is no law requiring a disclosure for AI-generated content. However, existing laws are technology-neutral, and they apply precisely where AI-generated content becomes problematic: in cases of deception, when real people are involved, and when there is misuse.

  • Unfair Competition Law (UWG): Any deceptive conduct that influences the relationship between suppliers and customers is considered unfair, particularly false or misleading statements about one’s own goods and services. An AI-generated product image that makes your product appear better than it actually is may fall under this category, with or without a disclaimer.
  • Data Protection (DSG): In May 2025, the EDÖB stated that users have a legal right to know whether they are speaking to a machine. According to the EDÖB, the use of programs that alter the faces, images, or voice messages of identifiable individuals must also always be clearly recognizable.
  • Rights of Personality (Swiss Civil Code): Anyone who replicates the face or voice of a real person must obtain that person’s consent; otherwise, they risk infringing on that person’s rights of personality, which can only be justified by an overriding interest or by law. According to the Federal Council, this includes the right to one’s own image and voice.
  • Criminal Law (Swiss Criminal Code): Abuse is already a criminal offense, for example, as defamation, fraud, or identity theft. Pornographic depictions of sexual acts involving minors are also criminal offenses even if the acts are “not actual”—that is, if they are artificially created.
  • Copyright Act (URG): It does not require any specific labeling. However, according to the Institute for Intellectual Property, anything generated by AI in response to a simple command is not protected.

Politically, this issue is evolving. In May 2025, the National Council rejected a proposal for general regulation of deepfakes; the Federal Council saw no gaps in criminal or civil law. In June 2026, at the request of the Federal Council, the National Council adopted a motion on sexualized deepfakes: A law would require providers of generative AI to assess risks and implement protective measures before bringing their products to market. The Council of States has not yet made a decision. For more information, see our press release on sexualized AI deepfakes.

In August 2026, the Federal Council stated in a response to an interpellation that measures to combat deepfakes are being considered in the proposed AI bill. It refers to the Council of Europe’s AI Convention, which, according to the Federal Council, cites the labeling of AI-generated content as an example of transparency. Our consultation tracker provides an update on the status of the bill.

This means that a Swiss labeling requirement may be introduced. Currently, for content produced entirely in Switzerland, the deciding factor is whether someone is misled or their rights are infringed.

What the EU AI Act Requires

The AI Act applies to Swiss companies if their AI system is used in the EU. The European Commission interprets this broadly in its July 2026 guidelines on Article 50: An operator outside the EU is covered if it intends to distribute the system within the EU, and simply publishing a deepfake on the freely accessible internet may be sufficient to meet this criterion. Content that enters the EU only through unforeseeable channels beyond your control is not covered. The guidelines are not binding, but they indicate how the Commission interprets the law.

Article 50 divides the responsibilities into two roles. Providers develop an AI system or have it developed and bring it to market or put it into operation under their own name—for example, the manufacturer of an image generator. Operators use an AI system professionally under their own responsibility—for example, an agency using it for its campaigns. According to the guidelines, anyone who merely commissions an advertisement without determining how the AI is used is not considered an operator.

An overview of the four obligations:

  • Chatbots (Paragraph 1, Providers): AI systems that interact directly with people must be designed so that users are made aware that they are interacting with an AI, unless this is already obvious.
  • Machine-readable markup (Paragraph 2, Provider): Generative systems must mark up their image, video, audio, and text outputs in a machine-readable format. This works like the security feature on a banknote that you can only see under UV light: invisible in everyday use, but verifiable. Exceptions include functions that only perform supporting tasks, such as spell-checking or color correction.
  • Deepfakes (Paragraph 4, Service Providers): Anyone who uses AI to create or alter image, audio, or video content that resembles real people, objects, places, or events and falsely appears to be genuine must disclose this fact. In the case of works that are clearly artistic, creative, satirical, or fictional, a notice that does not interfere with the enjoyment of the work is sufficient.
  • Texts of Public Interest (Paragraph 4, Operator): AI-generated texts that inform the public about matters of public interest must be disclosed. Exception: A human has reviewed the content, and a person bears editorial responsibility.

Timing is crucial: The notice must be provided clearly and unambiguously no later than the first interaction or suspension. According to the Commission, the provider’s machine-readable marking is not sufficient for this purpose because people do not notice it. Anyone who violates Article 50 risks fines of up to 15 million euros or 3 percent of global annual revenue; for SMEs, the lower amount applies.

The voluntary code of conduct for labeling AI-generated content offers practical guidance. It describes how providers should label content and how operators should mark it, and includes EU symbols that anyone is free to use: The core symbol is the abbreviation “AI,” while variations indicate whether content is fully generated or partially modified.

Schedule

The following dates apply for labeling, as summarized in the AI Act Service Desk schedule:

  • August 2, 2026: Article 50 takes effect. Since then, chatbots, deepfakes, and affected texts must be labeled.
  • December 2, 2026: End of the transition period for providers of generative systems that were on the market before August 2, 2026: They, too, must now label their outputs in a machine-readable format. On the same day, the bans on AI systems that generate intimate deepfakes without consent or depictions of child sexual abuse will take effect.
  • First half of 2027: Launch of the public consultation on the Swiss AI bill, according to the Federal Council's schedule.

What does December 2 mean for you? Your obligations as an operator have been in effect since August. Now, even established content generators must label their content; your job is to obtain these labels. Anyone who offers an image or video tool in the EU must also have effective safeguards in place against intimate deepfakes of identifiable individuals.

Three Typical Cases from Switzerland

Three simplified scenarios show how to examine both legal systems one after the other.

The AI-generated ad image for a client in the EU

Here's an example: An agency in Bern is designing a campaign for a sporting goods retailer in southern Germany. The main image shows a female runner who doesn't actually exist—she was created entirely using an image generator.

Is this a deepfake? Probably yes. According to the guidelines, it is sufficient that a person could plausibly exist and that the image appears to be real; advertising is not automatically considered an obviously creative work. The agency is the operator and visibly labels the image, for example with the “AI” symbol in a corner without any overlapping elements. If, on the other hand, the image shows a real product photo against an AI background without misleading viewers about the product, the Commission states that it is not a deepfake.

If the campaign were run only in Switzerland—for example, on billboards—there would be no requirement to include a disclosure; however, the Unfair Competition Act (UWG) still applies in that case. Specify in the contract who is responsible for including the disclosure and ensure that media partners do not remove it, as recommended by the Commission for long distribution chains.

The chatbot on the website

Here's an example: An online store in Thurgau hires an agency to build a chatbot based on a language model. The chatbot matches the store's design and also answers questions from customers in Germany.

Because the store has the chatbot developed and uses it under its own name, it is the provider. As of August 2, 2026, the chatbot must clearly state right from the start that an AI is responding—for example, in the first message and with a notice next to the input field. According to the guidelines, a single sentence in the terms and conditions, the mere designation “assistant,” or a misleading human profile picture are not sufficient. According to the EDÖB, users in Switzerland are also entitled to this notice.

The AI voice in the explanatory video

Here's an example: A software company in Lucerne produces explanatory videos for clients in Switzerland and Austria. The videos are narrated by a voice clone of the company's CEO.

According to the Commission’s examples, a cloned voice of a real person is generally considered a deepfake. The Code of Conduct requires a brief spoken disclaimer at the beginning of the audio and, where a screen is present, the visible symbol as well. A neutral synthetic voice that does not imitate a specific person is not automatically a deepfake; the decisive factor is whether the audience could be misled as to its authenticity.

In Switzerland, the managing director must give consent—preferably in writing—specifying the purpose and duration. Furthermore, according to the EDÖB, any alteration of voice messages from identifiable individuals must be clearly recognizable. A disclaimer at the beginning of the video is therefore the safest option under both legal systems.

Here's How to Label Things Correctly

Labeling is less a matter of the symbol than of the process. These six steps form a basic framework:

  1. Create an inventory. List where AI generates content: images, videos, voices, text, chatbots. Note the tool, purpose, and whether the content appears in the EU.
  2. Classify. Is it a deepfake, a text of public interest that hasn't undergone editorial review, a chatbot, or a case that involves only Swiss law?
  3. Clearly label it. Where necessary, include the notice the first time it appears: the “AI” icon in the upper-right corner of the image, at the beginning of videos, and, if possible, at regular intervals; for audio-only content, include a brief spoken notice. Use simple language and alternative text to make it understandable to everyone.
  4. Received technical marking. Generators that implement the code of conduct attach signed metadata and an invisible watermark to their results. An open standard for such provenance data is the C2PA’s Content Credentials, a kind of nutritional information label for digital content: they record how a file was created and edited. Check your tool’s documentation to see if it sets such data, and do not remove it during export. Providers who sign the code commit to prohibiting the intentional removal of this data in their terms of use.
  5. Establish approval procedures. Determine who will review AI-generated content before publication. For text, a spell-check alone is not sufficient: According to the Commission, an exemption from labeling requires a substantive review that includes fact-checking. When real people are involved, their consent is required for approval.
  6. Establish guidelines. Create a brief AI policy outlining when you’ll label content, which symbol to use, and who is responsible. The AI Strategy Package can help you get started.

It is important to note that a disclosure does not make impermissible content permissible. The Commission explicitly states that the labeling requirement does not permit illegal deepfakes, such as misleading advertising or intimate images shared without consent.

Frequently Asked Questions

Do I have to label AI-generated images?

In Switzerland, this is not generally the case, as long as no one is deceived or their rights are infringed upon. For images used in the EU that appear to be genuine, the AI Act requires a visible notice as of August 2, 2026. According to the Commission, obviously fantastical subjects, such as a dragon, are not considered deepfakes.

Do I have to label AI-generated text?

Only in limited cases. The AI Act covers texts that inform the public about matters of public interest—such as politics, health, or the environment—and whose content has not been reviewed or edited by anyone. According to the Commission, advertising copy and product descriptions that do not contain statements regarding health, safety, or sustainability are not covered by the Act.

Do I need to retroactively label older content?

No. According to the Commission, you are not required to retroactively label deepfakes created before August 2, 2026, or texts published before that date. If you are only now publishing an older AI-generated text, the requirement applies. The Commission nevertheless recommends labeling older deepfakes whenever possible without significant effort.

Conclusion

Labeling AI content is not yet a general requirement in Switzerland, but it’s rarely a matter of choice. Anyone who replicates real people, uses deepfakes or chatbots in the EU, or could mislead their audience has no choice but to provide a clear label. December 2, 2026, is a good opportunity to clearly define the criteria, symbols, and approval process once and for all. How does your team handle this?

In a nutshell:

  • In Switzerland, there is no general requirement to label AI-generated content; however, the Unfair Competition Act (UWG), the Data Protection Act (DSG), personality rights, and criminal law still apply.
  • According to the FDPIC, chatbots, as well as altered images and voices of identifiable individuals, must be recognizable.
  • If content is used within the EU, Article 50 of the AI Act has applied to chatbots, deepfakes, and certain texts since August 2, 2026.
  • Starting December 2, 2026, older generative systems will also be required to label their output, and new bans on intimate deepfakes will take effect.
  • The notice must be visible or audible no later than the first encounter; metadata alone is not sufficient.
  • Measures to combat deepfakes are being considered in the Swiss AI bill; the public comment period is scheduled for the first half of 2027.

As of October 7, 2026. This article is intended as a general overview and does not constitute legal advice; the assessment depends on the specific circumstances of each case.

Sources: AI Act, consolidated version on EUR-Lex · European Commission, Guidelines on Article 50 · European Commission, Code of Conduct · AI Act Service Desk, Timeline · FDPIC on the DSG and AI · Curia Vista 26.3458 · Curia Vista 26.3607