Since Aug. 2, a chatbot designed to leave you thinking you’re speaking with a human has been illegal in 27 countries. The same goes for a synthetic video of a politician that carries no label. The European Union’s transparency rules for AI are now both in force and enforceable, and it’s the penalties bolted onto them that caught the industry’s attention.
Companies can be fined as much as €15 million or three percent of worldwide annual revenue. For EU institutions, bodies and agencies, the ceiling is €750k. Enforcement sits with national authorities.
What has to carry a label now
The requirement is split in two by Article 50 of the AI Act, and that split matters more than the headlines let on.
Where machines only assisted in producing the content, an “AI” mark is required. That applies to “authentic-looking” deepfake imagery, audio and video, as well as published text.
Where a machine produced the content entirely, the stronger “AI-Generated” label applies. Regulators spell out their examples without hedging: “fully AI-generated deep fake videos with politicians or fictional events,” “full AI-composed music or art” and “AI-generated news summaries.”
Generated media isn’t the whole of it, either. Systems that read your emotions or handle biometric data must be disclosed as well.
“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.”
Look at that carefully and the aim isn’t the elimination of synthetic content. It’s preventing synthetic content from being mistaken for the real thing.
The carve-outs are narrower than you’d hope
Content that’s personal falls outside the rules. Your group chats are safe.
So is anything “evidently artistic” in the satirical or fictional vein — the clause that will generate no shortage of legal billing, since “evidently” is carrying enormous weight in a sentence about material built to look persuasive.
Small and medium-sized firms, along with small mid-cap companies, benefit from proportionality. Three percent is an upper limit rather than a fixed charge.
Model builders have a separate list
Those providing general-purpose AI models report to the Commission itself. Their obligations include documenting information, publishing summaries of training data and putting copyright policies in place. Failures get referred to national authorities.
The training data summary is the obligation to keep an eye on. Firms that have spent two years refusing to disclose what went into their models are now facing a regulator putting the question in writing.
Industry says it’s a drag, Brussels says it’s a shortcut
AI companies are making the usual argument: excessive burden, innovation held back. Europe has become the technology’s most visible regulator, and that role arrives with a permanent complaint attached.
Europe’s tech regulator argues the reverse — that building for one harmonized market beats building for 27 separate ones. The point isn’t without merit. Complying across a fragmented bloc is costlier than complying with a single 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.”
For anyone building or publishing anything involving generated media in Europe, the practical question isn’t whether they’re covered. It’s whether what you put out carries the correct one of the two marks, and whether your own pipeline can distinguish assisted from fully generated. Run that audit this week.
