AI Transparency and Disclosure: When Must You Tell People They Are Interacting With AI?

What AI Transparency and Labelling Actually Mean

Transparency is the principle that people should know when and how artificial intelligence is involved in an interaction or a piece of content. Labelling is the practical implementation: a visible notice, a watermark or a machine-readable signal. Both aim to enable informed decisions and to prevent deception.

A key distinction is roles: providers (those who develop or place an AI system on the market) and deployers (those who use it) carry different duties. If you embed a generative AI into your own website, you are usually the deployer and must ensure disclosure toward your own users.

Article 50 EU AI Act: the four transparency obligations

Article 50 groups the transparency obligations for certain AI systems regardless of their risk class. The core points:

• Chatbots and assistants: systems that interact directly with people must be designed so users are aware they are talking to an AI, unless it is obvious (provider's duty).
• Synthetic content: providers of generative AI must mark outputs (text, image, audio, video) in a machine-readable format as artificially generated or manipulated, technically robust and as reliable as feasible.
• Emotion recognition and biometric categorisation: deployers must inform the persons exposed about the use (with narrow exceptions such as permitted law enforcement).
• Deepfakes and public-interest text: deployers must disclose that image, audio or video content is artificially generated or manipulated. AI text published to inform the public on matters of public interest must also be disclosed; exceptions apply for clearly artistic or editorially controlled works.

Timeline and who is affected

The transparency obligations of Article 50 apply from 2 August 2026. The EU AI Office is expected to support implementation through voluntary codes of practice and guidelines, for instance on how machine-readable marking can look technically. Until then the exact technical standard remains in flux; C2PA and Content Credentials have become a de facto approach for provenance.

The territorial scope is extraterritorial: the EU AI Act also covers providers and deployers outside the EU when the output of the AI system is used in the EU or the offering targets people in the EU. For Swiss companies with customers, users or markets in the EU, the obligations are therefore directly relevant.

Swiss perspective: revFADP, FDPIC and the Council of Europe Convention

Switzerland has no dedicated AI law with a direct labelling duty like Article 50. What applies first is the revised Data Protection Act (revFADP), which requires transparency when processing personal data and imposes information duties; automated individual decisions carry additional notice obligations. The supervisory authority is the Federal Data Protection and Information Commissioner (FDPIC). For Swiss law one consistently refers to revFADP/nFADP, not the GDPR.

In addition, Switzerland has signed the Council of Europe Framework Convention on Artificial Intelligence and human rights, which anchors transparency as a principle. In 2025 the Federal Council decided on the regulatory course, favouring the use of existing frameworks and implementation of the Convention over a broad horizontal AI law in the short term. Swiss firms should nonetheless plan around the EU AI Act as a de facto standard.

Labelling AI in practice: concrete patterns

• Chatbot notice: an opening line at start ('You are chatting with an AI assistant') plus a persistently visible label. A handover to a human should also be clearly signalled.
• Visible content label: a note such as 'Made with AI' or 'AI-generated image' directly on the asset, not hidden in the imprint.
• Machine-readable provenance: embedded metadata and watermarks, practically via C2PA/Content Credentials, so platforms and search engines can detect origin automatically.
• Deepfake disclaimer: for synthetic persons, voices or scenes, add an unambiguous notice; for satire or art in a way that does not devalue the work but still rules out deception.
• Documentation: record internally which systems are used, who holds the provider and deployer role, and how disclosure is implemented technically. This eases evidence and audits.

Building trust beyond the obligation

Labelling is the floor, not the goal. Trust grows when transparency is explanatory rather than merely formal: why AI is used, which data feeds in, where humans retain responsibility, and how people can object or request a human review. Clear, always findable information beats a one-off pop-up.

For credible practice, a concise AI transparency statement on the website, consistent labels across all channels, and an honest description of the deployed models' limits are advisable. This reduces reputational risk and is a competitive advantage in the DACH region, where data protection and seriousness carry weight.