EU AI Act Article 50: Swiss Chatbot Compliance Duties

Since August 2, 2026, Article 50 of the EU AI Act has been in force, imposing transparency obligations on providers and deployers of chatbots, emotion-recognition and biometric categorisation systems, and any tool that generates or manipulates synthetic text, image, audio, or video. Unlike the Act's headline "high-risk system" provisions, these duties apply regardless of risk classification: a customer-support chatbot, a marketing deepfake generator, and a biometric categorisation tool are all in scope purely because of what they do, not how dangerous a regulator judges them to be. Swiss companies with no EU legal entity are not exempt — the Act's extraterritorial reach captures any organisation offering AI systems to users located in the EU, which covers the substantial share of Swiss fintech, e-commerce, and SaaS businesses selling across the border. The compliance clock is no longer theoretical: national market surveillance authorities can now act on non-compliance, with fines reaching €15 million or 3% of global annual turnover, whichever is higher.

What Article 50 actually requires

The obligations sit in four paragraphs of the Article, split between providers and deployers. Providers must design chatbots and other systems that interact directly with people so that users are informed they are dealing with an AI system, unless that is obvious to a reasonably informed person (Article 50(1)). Providers of systems that generate synthetic audio, image, video or text must mark the output in a machine-readable format so that it is detectable as artificially generated or manipulated (Article 50(2)). Deployers of emotion-recognition and biometric categorisation systems must inform the people exposed to them (Article 50(3)). Deployers must also disclose deepfakes as artificially generated or manipulated and disclose AI-generated text published to inform the public on matters of public interest, unless the text has undergone human editorial review with someone holding editorial responsibility (Article 50(4)). For evidently artistic, creative, satirical or fictional works the deepfake duty is limited to disclosure in a way that does not hamper the display of the work. All of this must be delivered clearly, at the latest at first interaction or exposure (Article 50(5)).

Generative AI systems already on the market before the August 2 deadline receive a narrow reprieve: machine-readable watermarking specifically has a grace period running to December 2, 2026, while the visible-disclosure duties for chatbots and deepfakes apply immediately with no equivalent grace period. Organisations that read "grace period" as covering the whole Article are exposed to enforcement action on the disclosure obligations that are already live.

The Swiss compliance gap: nDSG overlap and enforcement reality

Swiss organisations already navigating nDSG fairness and transparency obligations will find meaningful overlap here, but not equivalence. The nDSG's transparency principle concerns how personal data is processed; Article 50 concerns whether a system's AI nature and synthetic outputs are disclosed, a distinct duty that applies even where no personal data changes hands — a purely text-generation chatbot with no data-protection footprint is still squarely inside Article 50's scope. Swiss legal and compliance teams treating this purely as a data-protection adjacent issue risk missing an entirely separate enforcement track, one with its own fine ceiling and its own market-surveillance authority structure across each EU member state where the company has users.

Enforcement sits with national market-surveillance authorities in each member state where the company has users, and Article 50 infringements carry the fine ceiling cited above. Whatever the allocation of proof in a given proceeding, a Swiss company that has not documented its disclosure mechanisms, marking implementation and editorial-oversight process for AI-assisted public content will struggle to show that its practice was compliant.

Practical marking mechanisms and their limits

The European Commission published its final Code of Practice on marking and labelling AI-generated content on 10 June 2026. It is voluntary, but it is likely to become the main reference for demonstrating compliance with Article 50, and it points to a multi-layered approach: no single technique is enough, so the Code expects at least two layers of machine-readable marking where necessary, for example metadata combined with watermarks, and it provides guidance on displaying labels, including an EU icon or an equivalent label. In practice this is an engineering decision, not a legal afterthought. Marking must survive common transformations such as re-encoding, cropping or social-media re-upload to remain useful for downstream detection, so Swiss marketing, communications and product teams need to validate that their approach survives the channels their content passes through, rather than assuming a one-time implementation satisfies an ongoing obligation.

• Inventory every chatbot, conversational agent, and generative-AI tool that reaches EU users, regardless of whether it is classified as high-risk elsewhere in the AI Act.
• Implement clear, front-loaded AI-disclosure notices at the start of every chatbot interaction, not buried in terms of service or a dismissible popup.
• Apply visible synthetic-content labels to all AI-generated or AI-manipulated media, and validate that machine-readable markers survive your actual distribution channels.
• Establish a documented human-editorial-review process for any AI-generated text published on matters of public interest, with named accountability.
• Prioritise machine-readable watermarking retrofits for pre-August-2026 generative systems ahead of the December 2, 2026 deadline.
• Maintain compliance documentation proactively — disclosure mechanisms, watermarking evidence, editorial-oversight records — so that you can show compliance if an authority asks.
• Coordinate legal review between nDSG transparency counsel and AI Act compliance leads, since the two obligations overlap but are not interchangeable.

Article 50 is the first EU AI Act provision most Swiss organisations will actually encounter in production, precisely because it applies without regard to risk classification. The companies treating it as a documentation exercise now, before an enforcement inquiry forces the issue, will be the ones able to demonstrate compliance rather than scramble to reconstruct it after the fact.