As of Aug. 2, a chatbot that lets you believe you’re talking to a person is illegal across 27 countries. So is an unlabeled synthetic video of a politician. The European Union’s AI transparency rules are in force and enforceable, and the penalties attached to them are the part the industry noticed.
Fines run up to €15 million or three percent of global annual revenue for companies. EU institutions, bodies and agencies face up to €750k. National authorities handle enforcement.
What has to carry a label now
Article 50 of the AI Act splits the requirement in two, and the distinction matters more than the headlines suggest.
Content that machines merely assisted in creating gets an “AI” mark. That covers “authentic-looking” deepfake images, audio and video, plus published text.
Content generated fully by a machine gets a stronger “AI-Generated” label. Regulators name their examples plainly: “fully AI-generated deep fake videos with politicians or fictional events,” “full AI-composed music or art” and “AI-generated news summaries.”
The labeling net reaches past generated media. Models that recognize your emotions or process biometric data have to be disclosed too.
“The goal is simple: whenever AI plays a vital role, you are entitled to know,” the European Commission said in a video explainer.
The reasoning, in the Commission’s own words
“These obligations are intended to foster trust and integrity in the information ecosystem,” the Commission wrote in its new guidelines. “People should know when they are interacting with AI or exposed to AI-generated content. This will help them make informed decisions, calibrate their trust and reliance on AI and avoid mis information or deception.”
Read that closely and you’ll notice the goal isn’t to stop synthetic content. It’s to stop synthetic content from passing as the other kind.
The carve-outs are narrower than you’d hope
Personal content is exempt. Your group chats aren’t covered.
So are “evidently artistic” satirical and fictional works, which is the clause that will keep lawyers busy, because “evidently” is doing a great deal of work in a sentence about content designed to look convincing.
Proportionality applies to small and medium-sized companies and small mid-cap companies. The three percent figure is a ceiling, not a flat rate.
Model builders have a separate list
Providers of general-purpose AI models answer to the Commission directly. They’re required to document information, publish training data summaries and implement copyright policies. Non-compliance goes to national authorities.
That training data summary requirement is the one worth watching. Companies that have spent two years declining to say what they trained on now have a regulator asking in writing.
Industry says it’s a drag, Brussels says it’s a shortcut
The criticism from AI companies is the familiar one: undue burden, slowed innovation. Europe is now the most prominent regulator of the technology, and that position comes with a standing complaint attached.
Europe’s tech regulator makes the opposite argument, that a single market with harmonized rules is simpler to build for than 27 separate ones. There’s a real case there. Compliance across a fragmented bloc costs more than compliance with one rulebook.
“It also addresses potential risks to people’s health, safety, and fundamental rights, while safeguarding democracy and the rule of law,” the Commission wrote. “Ensuring its effective implementation is now a key priority for the Commission.”
If you build or publish anything that touches generated media in Europe, the practical question isn’t whether the rules apply. It’s whether your output carries the right one of two marks, and whether you can tell the difference between assisted and fully generated on your own pipeline. That’s the audit to run this week.
