What Article 50 really requires from August onward

On 2 August 2026, the transparency obligations in Article 50 of the European Union’s Artificial Intelligence Act (Regulation (EU) 2024/1689) take effect. There has been plenty of discussion about the Act in general, and little about what this specific article requires in practice. It will affect any business using AI with the public sooner than most other parts of the Act.

What it says, without legal jargon

Article 50 does not classify AI systems by how dangerous they are, as the rest of the Act does for high-risk systems. It sets out a much simpler principle: if someone interacts with AI or consumes AI-generated content, they must know it. It covers four specific situations:

Any system designed to interact with people, such as a chatbot or voice assistant, must be designed so that people know they are talking to a machine. Any system that generates or manipulates images, audio, video or text must mark that content so it can be detected as artificially generated. Emotion recognition or biometric categorisation systems must inform the people exposed to them. And any content that is a deepfake must be declared as such.

The detail people miss: when to give notice

The guidelines published by the European Commission in May 2026 clarify something many businesses are getting wrong: the notice cannot be hidden in the terms of use or shown just once at the start of the business relationship. It must appear before or at the beginning of each interaction. For a chatbot, that means showing the notice at the start of the conversation, not in a link at the bottom of the website. For generated audio or video, the notice must appear at the beginning of the clip.

The guidelines themselves give examples that do not meet the standard: tiny text in a website footer, an inconspicuous label on an image, or a notice that appears for a moment and disappears. The test is whether the information is clear and distinguishable: easy to notice, easy to understand and easy to identify as separate from the rest of the content.

The exception that raises the most questions

Article 50(1) waives the notice requirement when the artificial nature of the interaction is obvious to a reasonably informed and attentive person. The guidelines adopt the EU consumer protection law standard of the “average consumer” and add an important nuance: businesses must consider whether their target audience includes especially vulnerable people, such as children, older people or people with disabilities, and their level of digital literacy.

In practice, something obvious to a software developer, such as a coding assistant, may not be obvious to the general public calling a customer service line. Obviousness cannot be assumed. It has to be assessed in light of the actual audience.

Who is covered, even outside the EU

The Act applies not only to providers established in the European Union, but also to any provider outside the EU whose system is used within it. A recent legal analysis illustrates this with a straightforward example: a customer service chatbot serving European customers triggers the Article 50(1) obligation, wherever the company operating it is based.

The actual timeline, with its nuances

The general deadline is 2 August 2026. But one detail has changed in recent months: the provisional agreement reached by the Council and European Parliament on 7 May 2026 on the AI Digital Omnibus package introduces a transition period until 2 December 2026, but only for the technical marking of generated content (Article 50.2) in systems already on the market before August. The notice requirement for direct interactions (Article 50.1) does not get this extension and takes effect on the original date.

Penalties for non-compliance can reach €15 million or 3% of annual global turnover, whichever is higher.

Why this matters more than it seems

Article 50 is the first time a G7 jurisdiction has imposed binding and enforceable AI transparency obligations at mass-consumer scale. It does not require a complex technical audit like the one for high-risk systems. It requires something simpler and, in a way, harder to ignore: telling the truth about what is a machine and what is not at the exact moment someone needs to know, without hiding it in the fine print.