Europe's rulebook for artificial intelligence crossed a significant threshold on August 2, when the transparency obligations written into the EU AI Act became legally enforceable across the bloc. From that date, organisations operating AI systems in Europe must tell people when they are dealing with a machine rather than a human, and must mark the content their models produce so it can be recognised as synthetic.
The European Commission's AI Office, working alongside national regulators, began enforcing the wider AI Act on the same day. The Commission framed the change as a response to a problem that has grown faster than the tools to manage it.
The rapid development of generative and interactive AI systems is making it increasingly difficult to distinguish AI interactions and AI-generated content from human-created and authentic content, the Commission said in its transparency guidelines, arguing that people should know when they are interacting with AI so they can calibrate how much to trust what they see.
Providers and Deployers Carry Different Duties
The obligations sit in Article 50 of Regulation (EU) 2024/1689 and split responsibility between two roles. Providers, meaning the companies that build AI systems and place them on the market, must design those systems to disclose that a user is talking to software, unless that fact is already obvious. They must also embed machine-readable markings in synthetic audio, images, video and text, and supply a way to detect them.
Deployers, meaning the platforms and businesses that put those systems to work under their own authority, carry a different burden. They must label deepfakes and other AI-generated or manipulated media designed to pass as real, and must inform people subjected to emotion recognition or biometric categorisation systems. Large operators such as Meta fall into both categories at once.
There are carve-outs. Content that has passed through substantive human editorial review, with a named person taking editorial responsibility, is treated differently, and routine editing that does not substantially alter a work is exempt from the marking requirement.
Fines Reaching 3 Percent of Global Revenue
Non-compliance can trigger penalties of up to 15 million euros, roughly 17.2 million dollars, or 3 percent of worldwide annual turnover, whichever figure is higher. The reach extends well beyond Europe's borders: the AI Act binds providers, deployers, importers and distributors that place AI on the EU market or whose AI outputs are used inside the Union, regardless of where the company is headquartered.
The rules apply immediately to systems in scope, no matter when they launched, although material published before August 2 does not have to be retroactively labelled. One narrow transitional window remains. Providers of generative systems already on the market have until December 2 to satisfy the marking and detection requirement.
The Commission Made the Icons Itself
Rather than leave labelling design to each platform, the Commission published a set of disclosure icons alongside worked examples of when to use them. The icons resemble markers already rolled out by TikTok, Instagram and Facebook. Using the EU's specific artwork is voluntary; the underlying obligation to label is not, a distinction the Commission drew pointedly.
A voluntary Code of Practice on Transparency of AI-Generated Content offers a recognised route to demonstrating compliance, and the Commission has published a first list of more than 180 signatory organisations. Signing brings a measure of presumed conformity and a gentler enforcement posture; declining it means proving compliance some other way, under closer scrutiny.
Guidance adopted on July 20 gave companies less than two weeks of formal lead time before enforcement began, and the practical questions now shift to how consistently national authorities interpret terms such as obvious and substantively edited across 27 member states.
