A new obligation landed across the European Union this week, and it changes how millions of people will experience the internet. Since 2 August 2026, the bloc’s AI transparency rules require chatbots and other interactive systems to tell users plainly that they are talking to a machine rather than a person.
The European Commission’s AI Office, working alongside national regulators, began enforcing the wider Artificial Intelligence Act on the same date. The transparency layer sits at the heart of that push. Companies that deploy conversational agents can no longer let users assume a human sits behind the screen.
The rules reach further than chat windows. Deepfakes, meaning images, video or audio that AI generates or alters, must now carry a label. Providers must also embed machine-readable marks in synthetic content so that platforms and researchers can spot manipulation before it spreads. The Commission frames the measures as a defence against deception ahead of a crowded election calendar.
Regulators argue that clarity protects choice. When a shopper argues with a support bot or a voter watches a clip of a politician, they deserve to know what they are dealing with. The Commission’s own guidance ties the obligations to informed consent, a principle it has pressed since the Act cleared the Parliament.
Industry did not wait for penalties. More than 180 organisations have signed a voluntary Code of Practice on the transparency of AI-generated content, a scheme that turns the legal text into concrete engineering steps. Signing buys goodwill and a clearer route to compliance, though it does not shield a company that ignores the underlying law.
The same date switched on enforcement for general-purpose AI models, the large systems that power most consumer tools. Providers now shoulder binding duties on documentation, copyright and systemic-risk assessment. The Commission has opened a complaints channel and a whistleblower tool so that users and downstream developers can flag breaches directly.
Critics inside the tech sector warn that the definitions still leave room for argument. When does a lightly edited photograph become a deepfake? How prominent must a chatbot’s disclosure be? The guidelines offer examples, yet the first real answers will come from enforcement decisions rather than press releases.
Smaller developers voice a different worry. They fear the paperwork will favour large incumbents that can absorb compliance teams, while start-ups burn scarce engineering time on labelling pipelines. The Commission counters that the Code of Practice exists precisely to lower that cost and to spread proven templates.
The stakes stretch beyond Europe. Because global platforms rarely build one product for Brussels and another for the rest of the world, the AI transparency regime will likely shape interfaces far outside the single market. That extraterritorial pull, familiar from the bloc’s privacy and data rules, gives this week’s deadline weight well past the continent.
Enforcement will now test whether disclosure changes behaviour or simply adds a line of small print. For the moment, the message to anyone chatting online in Europe stays blunt: assume nothing, and read the label.
