Chatbots and AI content: what has been mandatory since 2 August 2026

A deadline maintained despite the Digital Omnibus

Since 2 August 2026, the transparency obligations of Article 50 of the AI Act — the European regulation on artificial intelligence (Regulation (EU) 2024/1689) — have applied. The “Digital Omnibus AI” regulation (Regulation (EU) 2026/1744 of 8 July 2026, published in the Official Journal on 24 July and in force since 27 July 2026) postponed several parts of the text, notably Annex III high risk to 2 December 2027. Article 50, for its part, was maintained — and the Commission published its final guidelines on that article on 20 July 2026.

Swiss companies are concerned. The AI Act applies to third-country actors as soon as their artificial intelligence (AI) systems touch the market of the European Union (EU): European customers, or outputs of the system used in the EU (art. 2). A French-speaking Swiss small or medium-sized enterprise (SME) whose chatbot answers French visitors, or which publishes generated content read by European customers, falls within scope.

Who is concerned

Targeted are the providers of these systems, but also the deployers — the companies that use them under their authority. Concretely: you operate a chatbot facing your customers, you publish AI-generated texts, images or videos, you distribute realistic manipulated content, or you use a lawful emotion recognition system.

If your activity is strictly Swiss, with no link at all to the EU, art. 50 does not bind you. Flagging AI to your customers nonetheless remains advisable: the FADP (Federal Act on Data Protection) already requires transparency of processing (art. 19 FADP), and the preliminary draft of a Swiss AI law expected at the end of 2026 will very probably make transparency a pillar.

The four cases of article 50

Case 1 — the chatbot (art. 50(1)). Anyone interacting with a conversational AI system must know they are speaking to a machine, unless that is obvious from the context. The information must come at the start of the interaction, not in buried terms and conditions.

Case 2 — generated content (art. 50(2) and 50(4)). Content produced by AI (text, images, audio, video) must be marked in a machine-readable format — an obligation that falls first on the provider of the tool. For published information texts (articles, news), the deployer must in addition disclose the AI generation, unless a human has reviewed the content and editorial responsibility is assumed.

Case 3 — deepfakes (art. 50(4)). Any realistic generated or manipulated content depicting real people, places or events must carry a visible and unambiguous label.

Case 4 — emotion recognition and biometric categorisation (art. 50(3)). In lawful cases, the persons exposed must be informed beforehand. Important reminder: emotion recognition at work and in education has been prohibited since February 2025 (art. 5(1)(f)) — no notice can make it lawful.

Existing systems: the nuance of 2 December 2026

The Digital Omnibus introduced a targeted transition. For generative systems placed on the market before 2 August 2026, the machine-readable marking of content (art. 50(2)) benefits from a reprieve until 2 December 2026.

Watch the exact scope of that reprieve: it covers only the machine-readable marking of earlier systems, an obligation that falls chiefly on tool providers. Informing people facing a chatbot, labelling deepfakes and disclosing on published information texts apply, by contrast, since 2 August 2026, with no transition. As a deployer, do not count on December: rather, check with your providers that marking will be switched on in time.

The penalties

Failure to meet the transparency obligations is punishable by fines of up to 15 million euros or 3 % of worldwide annual turnover, whichever is higher (art. 99(4)). For a small or medium-sized enterprise (SME), art. 99(6) reverses the rule: it is the lower of the two ceilings that applies. The reputational risk — an undeclared chatbot, an unlabelled deepfake — can in any case cost more than the fine.

Getting compliant in a week

Day 1: inventory your public-facing AI touchpoints. Website chatbot, WhatsApp assistant, AI-written newsletters, visuals generated for social media, demonstration videos. For each, note the tool used and the provider.

Days 2 and 3: draft and display the notices. For a chatbot, at the start of the conversation, for example: “You are talking to a virtual assistant based on artificial intelligence from [company]. Its answers are generated automatically and may contain errors. To speak to a person, ask for an adviser at any time.” For published content, in a caption or footer: “This content was generated or assisted by artificial intelligence, then published under the responsibility of [company].” If your customer base is multilingual, plan the notices in French, German, Italian and English.

Day 4: contact your generative tool providers and ask for written confirmation that machine-readable marking of content is or will be enabled (art. 50(2)). It is their obligation, but your reputation depends on it.

Day 5: lock down the organisation. Document the human review of published information texts if you invoke the disclosure exemption (art. 50(4)). Train the people who publish — which also advances your AI literacy obligation (art. 4), in force since February 2025. Add the use of AI to your privacy policy (art. 19 FADP), applicable in Switzerland regardless of the AI Act.

Where to start

The simplest route is to begin from an honest inventory of your uses, then handle the public-facing touchpoints one by one, notice by notice.

AI-Karma's free tool offers a transparency module (/transparence) that generates these ready-to-paste notices, in four languages, tailored to your company name, with the checklist of accompanying steps. Its classification module (/aiact) allows you to check upstream, in about fifteen minutes, whether your use cases fall under art. 50, under high risk, or under neither.

This is not legal advice. The model notices are to be adapted to your situation and, in case of doubt, validated by specialised counsel.

Frequently asked questions