EU AI Act: creators get no grace period for labeling AI content

The transparency rules of the EU AI Act took effect today. We covered what changes for you as a reader last week. One part of it keeps getting lost. The regulation does not only bind OpenAI, Google and Meta. It binds you as well, the moment you publish something yourself.

Private use stays free, income changes everything

The European Commission draws the line at money, not at reach. Post an AI video purely privately and you are acting in a personal capacity, which falls outside the scope of the rules. Do it in a way that regularly earns you an economic benefit, or as part of a business, a profession or freelance work, and you count as a deployer. That is the term the regulation uses for anyone who runs an AI system under their own responsibility. A monetised YouTube channel will usually qualify, and so will an Instagram account with paid partnerships. Employees are not personally on the hook, because the company carries the duty.

What you have to label

The law names two cases. The first one covers deepfakes, meaning AI-generated or AI-altered images, audio and video that show real people, places or events and could plausibly pass as authentic. The second one covers AI-written text you publish to inform the public about a matter of public interest, so topics such as politics, public health or consumer safety.

There are exceptions. Obviously fantastical material is not a deepfake, and the Commission cites humans flying unaided as an example. For work that is clearly artistic, satirical or fictional, a disclosure that does not spoil the piece is enough. For text, the duty falls away if a human has reviewed the substance and someone carries editorial responsibility. A spell check explicitly does not count.

The provider watermark does not cover you

This is where it gets expensive. The machine-readable marking that ChatGPT, Gemini and Sora embed in their output does not discharge your duty. The Commission states plainly that deployers cannot fall back on it. Your disclosure has to be noticeable without any technical tools, so visible or audible, and it has to be in place by the time someone first encounters the content. A tiny note in a footer, a faint label or a line buried in your terms will not do. The EU offers three icons for exactly this purpose, free of charge, as SVG and PNG. They are not mandatory and the code of practice behind them is voluntary, but they save you the design question.

There is no grace period for you

The much-quoted deadline of 2 December 2026 applies to providers only, only to the machine-readable marking under Article 50(2), and only to systems that were already on the market before today. Your duties as a deployer start with no transition at all. Content you published before today does not have to be labelled retroactively, although the Commission asks you to do it wherever that is not a burden.

Enforcement is split up. Germany chose its Federal Network Agency as the central contact and complaints body on 29 July, but it only acts as a market surveillance authority within its own remit, and the media sector stays with the state media authorities. How hard the EU pushes the big platforms shows in the fight over AI assistants on phones. Fines reach 15 million euros or 3 percent of worldwide annual turnover, whichever is higher. Nobody is going to hit a channel with a few thousand subscribers with a sum like that. The complaints desk is open to everyone all the same.