AI Act Article 50: Transparency and Labelling Obligations

Providers must design and develop AI systems intended to interact directly with natural persons so that the people concerned are informed that they are interacting with an AI system. This covers classic website chatbots as well as voice assistants on the phone, AI avatars in video calls and AI agents that answer emails.

The obligation only falls away where this is obvious from the point of view of a reasonably well-informed, observant and circumspect person, taking into account the circumstances and context of use. Apply this exception with care: a voice assistant with a natural-sounding voice is precisely not obvious as AI to many people. Special rules apply to systems authorised by law for law enforcement.

Timing matters: the notice must be given at the latest at the first interaction (para. 5). A sentence in the legal notice or terms and conditions does not do the job. In practice, a combination of greeting text and a permanent label in the chat window works well, so the notice stays visible when someone joins the conversation later.

Providers of AI systems – including general-purpose AI systems – that generate synthetic audio, image, video or text content must ensure that the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. This obligation is aimed at the makers of generative AI, not at the companies that use it to create images or text.

The technical solutions must be effective, interoperable, robust and reliable as far as technically feasible, taking into account the specifics of the content types, the costs of implementation and the generally acknowledged state of the art. In practice several techniques are combined – such as metadata, invisible watermarks and logging – because each on its own is easily lost; metadata, for instance, is stripped by many platforms on upload.

Exempt are systems that only perform an assistive function for standard editing or do not substantially alter the input data provided by the deployer or its semantics, as well as systems authorised by law for law enforcement. For systems placed on the market before 2 August 2026, the Digital Omnibus grants a transition period until 2 December 2026.

Deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake must disclose that the content has been artificially generated or manipulated. A deep fake is content that appreciably resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful.

Typical cases in business: a photorealistic image of a “customer” who never existed, a video in which the CEO speaks to staff in several languages via a voice clone, or a product photo in a setting that does not exist. The obligation is not limited to publication – an internal training video with a synthetic presenter is also covered if it could falsely appear authentic.

Where the content forms part of an evidently artistic, creative, satirical, fictional or analogous work, the obligation is limited to disclosure in an appropriate manner that does not hamper the display or enjoyment of the work. The Commission’s guidelines read this exception narrowly. They also stress that where it is foreseeable that people will not perceive content from the start – such as videos people join midway – a one-off notice at the beginning is not enough.

Deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest must disclose that the text has been artificially generated or manipulated. This covers news and reports on politics, health, the economy or public safety – including on company websites when they cover such topics in an informative way.

The key exception: the obligation does not apply where the AI-generated text has undergone a process of human review or editorial control and a natural or legal person holds editorial responsibility for its publication. For press offices and newsrooms with a clear approval process, that is the norm. Automatically published text, such as AI-generated market reports or news feeds without review, is the critical case.

Advertising copy, product descriptions and internal documents are generally not covered, because they do not serve to inform the public on matters of public interest. The line can blur, though – think of a corporate blog on health topics. If in doubt, a short voluntary note or a documented editorial process helps.

Deployers of an emotion recognition system or a biometric categorisation system must inform the natural persons exposed to it about the operation of the system. The provision also expressly requires that personal data be processed in accordance with the GDPR and other data protection rules.

Before thinking about notice texts, there is a prior question: is the use allowed at all? Emotion recognition in the workplace and in educational institutions is prohibited in principle under Art. 5(1)(f) AI Act, with exceptions only for medical or safety reasons. That includes features in meeting tools or call-centre software that analyse employees’ “mood”. Certain forms of biometric categorisation are prohibited too.

Where the use is permitted – for example in market research with consent or drowsiness detection in vehicles – the notice must be given before, or at the latest when, capture begins and must be perceivable by everyone affected, including people with disabilities. Since biometric data may be involved, a data protection impact assessment is usually part of the package.