Article 50 of the EU AI Act starts applying on August 2, and the first test is simple: if people are dealing with AI, they need to be told.
Starting Sunday, every company shipping a chatbot, a deepfake tool, or an AI text generator into the European Union has a legal obligation it didn't have last week. That's new. Article 50 of the EU AI Act applies from August 2, 2026, and Brussels has already put people behind the rule. According to the Associated Press, the European Commission is adding 38 staff to its AI Office to monitor companies from startups to OpenAI and DeepSeek, and it has launched whistleblower and compliance tools for reports of suspected breaches. The machinery is in place. Now the question is whether companies are ready.
The Commission published its Article 50 transparency guidelines on July 20, less than two weeks before the rules started to apply. That wasn't generous. The guidance covers direct AI-to-human interaction, AI-generated content including deepfakes, emotion recognition and biometric categorisation, and AI-generated text on public-interest matters. If you're building or deploying any of those systems in Europe, the old habit of hiding AI behind a friendly interface is over.
The core rule is blunt. People must be told when they're directly interacting with an AI system unless that fact is already obvious from the context. AI-generated or manipulated audio, image, video and text must be marked in a machine-readable format when the provider is in scope. Deployers also have to disclose deepfakes and certain AI-generated public-interest text. The Commission's own Article 50 summary says the information has to be clear, distinguishable and accessible at the time of first interaction or exposure. That is not buried compliance language. It is a user-facing duty.
The label has to follow the content #
The public will notice the rule most in places where synthetic material already looks normal: chatbots, ads, political clips, customer service avatars, publishing tools and image generators. The Guardian reported Friday that synthetic text, images, video and audio designed to look truthful must be visibly marked as AI-generated and carry a digital watermark showing their artificial origin. Fines can reach €15 million or 3% of worldwide annual turnover. For a large AI lab, 3% is not a nuisance fee.
There is some breathing room, but not enough to ignore the deadline. Under the EU's transition provision, generative AI systems already on the market before August 2 have until December 2, 2026, to meet the machine-readable marking obligation. That grace period is narrow. It doesn't erase the broader disclosure duties that start Sunday, and it doesn't cover content made after the deadline as if nothing changed. If your EU-facing chatbot still pretends to be a person next week, you're exposed.
The rule on public-interest text is the one media companies and political operators should read twice. Article 50 requires disclosure when AI-generated or manipulated text is published to inform the public on matters of public interest. There is an exception where the text has gone through human review or editorial control and a natural or legal person takes editorial responsibility for it. That's a real carve-out. But it isn't a magic phrase you add after publication. You need an editorial process that can be shown if a regulator asks.
Look at the edge cases and you can see why industry groups are already nervous. The Guardian quoted CCIA Europe's AI policy lead, Boniface de Champris, warning that the Commission's July guidance could turn the label into something close to a cookie banner, everywhere and easy to ignore. He pointed to examples as different as a field in an advert and a manipulated political speech. He has a point. If everything gets a label, the label gets weaker.
The enforcement risk is no longer theoretical #
OpenAI and DeepSeek are useful names here because they show the range of targets. AP reported that the new AI Office staff will monitor American and Chinese tech firms as well as newer companies. DeepSeek already drew European regulatory attention after Italy's data protection authority asked questions about its handling of personal data and then blocked the app in early 2025, as TechCrunch and other outlets reported at the time. OpenAI faces a different problem: scale. When millions of users touch a product, even a small disclosure failure can become a large compliance file.
The AI Office's Whistleblower Tool changes the pressure inside companies too. The Commission says people professionally connected to providers can submit reports anonymously in any EU language through a secure inbox. That's not decoration. It gives engineers and trust-and-safety workers - and the policy staff who sit between them - a route to flag suspected breaches without waiting for an outside audit. If you're running compliance as a slide deck, that should worry you.
The harder part is that Article 50 lands before the market has settled on one clean technical standard. Google has SynthID. Meta has labeling policies. The Commission has created black-and-white label icons and a voluntary code of practice, which it says more than 180 organisations have signed. Those are useful pieces, not a complete answer. Companies that don't sign the code can still comply, but they will have to show their own measures are adequate.
Frankly, the serious firms should have seen this coming. The AI Act was adopted in 2024. The transparency date was known. And the Commission has now written down the operational expectations. What changed this week is not the existence of the rule. It is the arrival of staff, tools, fines and a live deadline.
Also read: Perplexity charges $200 a month for an AI agent that does Microsoft Copilot's job on Microsoft's own Windows • Larry Ellison staked Oracle on OpenAI and now Wall Street is asking what happens if he's wrong • Moonshot AI's Kimi K3 runs on 20,000 Nvidia chips through Alibaba as Washington alleges export-control breach