Brussels: From 2 August, Europe expects the machines to own up. That day the transparency duties in Article 50 of the EU’s AI Act start to bite, forcing providers and deployers to tell people when they face a chatbot, a synthetic voice or an image a model conjured from nothing. The question is whether the new regime changes behaviour or simply adds a line of fine print.

The rules split the duty in two. Providers that build generative systems must mark their output in a machine-readable way across text, audio, image and video, so detectors and platforms can flag it downstream. Deployers who use those tools to create deepfakes must disclose that the content is artificial.

Where AI transparency gets hard

The obligations sound simple and prove awkward. A standardised EU label still sits on the drawing board, and the Commission is drafting a code of practice to spell out how marking should work in code. Until that lands, firms must interpret broad principles on their own, then defend those choices to regulators later.

The scope catches more than obvious fakes. The Commission’s draft guidelines, published in May, say content that looks or sounds like a real person needs a label even when no one meant to deceive and even when the person does not exist. That sweeps in satire, advertising and film work that once escaped scrutiny.

Brussels softened the edge with time rather than substance. Systems already on the market before 2 August get a grace period until 2 December to fit the technical marking and detection measures. New systems enjoy no such cushion, which pushes developers to bake labelling into products now rather than retrofit it.

Why enforcement will decide the outcome

Rules only matter if someone checks them. The Commission gains direct powers over general-purpose AI models on the same August date, giving it a lever over the foundational systems that sit beneath most consumer tools. That combination, transparency duties plus oversight of the biggest model providers, marks a shift from writing law to testing it.

Yet the practical hurdles are real. Watermarks can be cropped, stripped or degraded, and no marking standard survives every edit. A determined bad actor will find ways around a label, so the rules work best against careless or commercial misuse rather than deliberate disinformation. Officials know this, which is why they frame marking as one layer among several rather than a cure.

The compliance cost falls unevenly. Large model providers can absorb the engineering, while smaller studios and agencies must buy or build detection they barely understand. That gap risks entrenching the incumbents the EU says it wants to challenge, an irony that runs through much of the bloc’s digital rulebook.

For ordinary users, the payoff is modest but real. A labelled deepfake is easier to doubt, and a disclosed chatbot is harder to mistake for a human agent. Trust, once lost to a flood of synthetic media, costs a fortune to rebuild, and Europe is betting that early labels beat late regret.

The 2 August start will not settle whether the AI Act works. It will show whether transparency can move from principle to habit, and whether regulators can hold the line when the first hard cases arrive. On that answer rests much of Europe’s claim to govern a technology it does not build.