Article 50 of the AI Act governs transparency. It is the article that affects most companies, because it is not tied to a high-risk classification: even an entirely harmless system can be subject to transparency duties if people interact with it or it generates content. These duties have applied since 2 August 2026. The idea fits in one sentence: people should know when they are dealing with a machine and when content was generated by one.
Interacting with an AI system
Where a person interacts with an AI system, they must be informed of this, unless it is obvious from the circumstances. The most important practical case is a chatbot on a website or in a customer channel. Implementation is simple and still regularly done wrong. The notice belongs at the start of the interaction and in the interface, not in the terms of use. A footnote does not serve the purpose, because it does not reach the person before they start typing.
Labelling generated content
For synthetically generated content, whether text, image, audio or video, the article requires machine-readable marking. This requirement is addressed to the providers of generating systems and is technical in nature, usually implemented through metadata or watermarks. For existing generative systems, the transitional period for this marking ends on 2 December 2026. Separate from this is disclosure to people, which applies especially where published content depicts a real person or event without being real. An AI-generated marketing image without reference to a person raises different considerations than a video that makes a person say something.
What this means for a typical business
For most companies the duties come down to three manageable points. First, every chatbot in a customer channel needs a notice at the start. Second, where content that was essentially generated by a system appears publicly, that must be made recognisable, subject to the exceptions for human editorial review. Third, where a system recognises emotions or categorises people biometrically, additional information duties towards the people affected apply. The third point is most often overlooked, because such functions frequently sit as a side feature in a purchased product, for example in analytics for customer calls.
The most common implementation mistakes
Three patterns recur. The notice is in the terms of use instead of the interface. The chatbot uses a human name and becomes evasive when asked, which aggravates the breach rather than curing it. And labelling is implemented only on the company’s own website, not in the other channels where the same automation runs. A system that does not say it is a system when asked directly damages trust more than honest labelling would ever have cost.
Who carries the duty
With a bought-in chatbot, the duty follows the role: whoever operates the system under their own name in their own channel informs users, regardless of who supplied the technology. Pointing to the vendor does not relieve you. Labelling therefore has to be considered when integrating the product; standard products often offer a setting for it, but it is rarely active by default.
Transparency is cheaper than its reputation
The widespread concern that disclosing AI use lowers acceptance usually does not survive experience. What bothers users is not the machine but the deception, and a clearly labelled assistant that answers quickly is generally accepted. Article 50 is the duty with the best ratio of effort to effect in the whole act: a few lines in the interface, low risk and a gain in trust that justifies the measure on its own. How we build chatbots that meet these requirements is described in our offer for an AI chatbot on your own data; the rules are summarised in the glossary entry on Article 50 transparency duties.