Since August 2, 2026, Article 50 of the AI Act has applied without any easing, unlike Article 4, which was rewritten this summer by Regulation (EU) 2026/1744. This text does not only concern model publishers: a company that puts an assistant online, publishes generated visuals, or disseminates AI-produced texts is directly subject to it.
Four Obligations, Two Types of Actors
Article 50 distinguishes between providers, who develop an AI system or place it on the market, and deployers, who use it in their activities. Providers of systems intended to interact directly with people must ensure that they know they are addressing an AI, unless this is obvious to a reasonably informed person.
Providers of systems that generate text, images, sound, or video must also mark these outputs in a machine-readable format, so they can be detected as artificial. Deployers of emotion recognition or biometric categorization systems must, for their part, inform the people exposed to them.
Finally, deployers who disseminate hyper-realistic fake images, sounds, or videos, meaning deepfakes, must indicate that they have been generated or manipulated. The same rule applies to texts published to inform the public on matters of general interest, unless they have been reviewed by a human and a person assumes editorial responsibility for them.
Only One Flexibility, Regarding Marking
Generative systems placed on the market before August 2, 2026 have until December 2, 2026 to apply the marking, according to Regulation (EU) 2026/1744. The other obligations did not benefit from any grace period. A code of good practices on the transparency of generated content was finalized this summer. It is not mandatory, but it proposes a method.
It is often said that these obligations primarily concern OpenAI, Microsoft, or Mistral. This is true for marking the outputs of their models, but much less so when a company assembles its own assistant, for example a Copilot Studio agent or a chatbot built on an API and deployed under its own name. It can then be considered the provider of that system, and informing the user falls to it. The deployer obligations, relating to deepfakes, public interest texts, and emotion recognition, remain its responsibility in all cases.
What Needs to Be Checked
Every conversational agent exposed to customers or employees must announce itself from the first exchange, for example with the phrase ‘You are interacting with an automated assistant,’ and not at the bottom of a legal notice page. Voice agents are concerned in the same way. Generated visuals, voices, or videos that realistically depict people, places, or events must carry a notice as soon as they could pass for authentic. For a clearly creative or satirical work, a discreet notice is sufficient.
Published texts, blog posts, press releases, or newsletters must be flagged if they are generated without human review or an identified editorial manager. Documented review remains the best protection. Furthermore, emotion recognition in the workplace is prohibited by Article 5, except for medical or safety reasons, which calls for close examination of an HR or call center tool that ‘measures engagement.’ For purchased tools, publishers must be asked if marking of their outputs is in place or planned before December 2.
What Is at Stake
A breach of Article 50 is punishable by a fine of up to 15 million euros or 3% of global turnover, whichever is higher. For SMEs, the lower of the two amounts applies. A company that has already inventoried its AI uses for Article 4 has a working base. It just needs to add four questions: does the tool interact directly with people, does it produce content disseminated externally, who is its provider under the regulation, and what notice is in place?
