The AI Act tightens its rules: AI transparency is now mandatory in the EU
Since August 2, 2026, the transparency obligations of Article 50 of the European Artificial Intelligence Regulation (AI Act) are fully applicable. Chatbots, AI-generated content and deepfakes must be clearly identified, and companies that fail to comply face penalties of up to 3% of their global turnover. The era of invisible AI has ended in Europe.
The AI Act entered into force on August 1, 2024 and since then it has been applied in phases: the prohibitions of practices considered unacceptable and the obligation of AI literacy arrived in February 2025, and the governance rules and the duties of general-purpose models, in August 2025. The next big stage was precisely the one that has just been activated: the transparency obligations of Article 50, the operational heart of the regulation.
What Article 50 requires
Article 50 obliges providers and deployers of certain AI systems to inform people when they are interacting with a machine. If a user talks to a customer service chatbot, they must know they are not talking to a human, unless it is evident. It also requires labeling synthetic content: an image, video or voice generated or artificially manipulated must be marked as such, and deepfakes dealing with matters of public interest cannot be presented as authentic.
In addition, when AI is used for emotion recognition or biometric categorization, the affected people must be informed, unless legal purposes concur. The European Commission published on June 10 a Code of Good Practice on transparency of AI-generated content, with a dual-layer marking system and watermarks. Adherence to the code is voluntary, but the obligations of Article 50 are legal requirements, not recommendations.
The high-risk postponement does not postpone transparency
It is worth clarifying a common misunderstanding. The so-called Digital Omnibus, approved by the European Parliament on June 16, has delayed the rules for high-risk systems: the autonomous systems of Annex III move to December 2027 and those integrated into products with sectoral regulation, to August 2028. That respite has led many companies to think that the AI Act waits again. That is a mistake: the postponement does not suspend the prohibited practices, nor AI literacy, nor the obligations on general-purpose models, nor Article 50.
In fact, there is a transitional period until December 2, 2026 for marking content generated by systems already marketed before August 2, but it refers to a specific perimeter: for everything else, transparency is already enforceable and punishable.
The real risk: the AI already in the company
The greatest danger for many organizations is not the big AI project presented to the management committee, but the AI that is already installed without anyone having mapped it: smart features activated in a CRM, a scoring module in a marketing tool, a recruiting assistant, a content generator in communications or an office copilot deployed without control.
The latest Deloitte report on the state of AI in the enterprise shows that fewer than six out of ten workers with access to corporate AI tools use them daily. That means many companies do not even know which AI they use, who uses it and for what. Without that map, it is impossible to comply with the AI Act, and since August 2 the exposure is legal, not theoretical. Article 4, in addition, obliges ensuring a sufficient level of AI knowledge among the workforce: training is no longer just a competitive advantage, it is a regulatory duty.
What companies should do now
Compliance does not require a compliance department or large investments. The Commission does not impose a rigid curriculum or a minimum number of hours. It is enough to be able to demonstrate three things: a basic inventory of the AI tools in use, an internal note with rules per team or position, and a training record, even if half-yearly. The pattern recalls the GDPR of 2018: years of ignored warnings, first fines and last-minute scrambles.
The AI Act has not come to complicate companies’ lives, but to force what should already have been done: knowing which AI is used, understanding its risks, warning users and training the people who handle the tools. The date has already passed. Now it is time to demonstrate, not promise.
