Since Saturday, AI in Europe has to admit it's AI. The EU's new transparency rules under the AI Act kicked in on August 2, and they're the first piece of the law that ordinary users will actually see: chatbots must now say they're machines, and deepfakes need a label.
Two rules, plainly stated
The European Commission's announcement boils down to two duties for anyone building or deploying AI.
Rule one is about content. AI-generated images, audio, and video that look like real people, places, objects, or events — deepfakes — have to be labelled clearly and visibly, and the files themselves must carry machine-readable marks. The same goes for emotion recognition tools, biometric categorisation systems, and AI-written text that's published to inform the public without a human editor checking it first. The EU even designed an official set of icons so the labels look the same everywhere.
Rule two is about conversation. If you're chatting with a bot, an AI agent, or an avatar, it has to tell you. No more wondering whether the endlessly patient support rep is a person.
Why now
The Commission isn't subtle about its reasoning: AI got good enough that people can't reliably tell generated content from the real thing anymore. Its announcement lists what follows from that — misinformation and manipulation at scale, fraud, impersonation, and consumer deception. Labels won't stop any of it outright, but they give people a fighting chance to know what they're looking at.
Who enforces it, and what it costs to ignore
Enforcement is split three ways: national market surveillance authorities in each member state, the European AI Office for the systems it supervises, and the European Data Protection Supervisor when EU institutions are the ones running the AI. The fines are not small. Companies can be hit for up to €15 million or 3% of global annual turnover, whichever hurts more. EU bodies themselves face penalties up to €750,000. Smaller firms get proportionality, not immunity.
There's help for companies trying to do this right — the Commission has published compliance guidelines, and signing the code of practice on AI-content transparency counts as evidence of good faith.
The staged rollout, and where this fits
The AI Act technically entered into force back on August 1, 2024, but it applies in waves. Earlier waves covered banned practices and obligations for general-purpose models — important, but invisible to most people. This wave is different. It's the one that shows up in your chat windows and your feeds. And the Commission says effective implementation is now a top priority, which is regulator-speak for: we intend to actually enforce this.
What changes on your screen
If you're in the EU, expect disclosure notices in chat interfaces, visible tags on synthetic video and audio, and invisible marks in file metadata that platforms and researchers can scan even when a visible label got stripped. If you run a platform or ship an AI product, the grace period is over — the clock started Saturday, and the fine print is indexed to your global revenue.
Image: Marco, via Pexels
