The EU AI Act takes its first big step: from today, chatbots must say they are machines
This Sunday, 2 August 2026, marks a before and after in the regulation of artificial intelligence: the first transparency obligations of the European AI Regulation (AI Act), the first comprehensive law on this technology in the entire world, come into force. From today, chatbots and virtual assistants must inform users that they are conversing with a machine, and content generated or manipulated by AI —including deepfakes— must be identifiable as such. Brussels is also assuming, for the first time, direct supervisory and sanctioning powers over large general-purpose models, such as those behind ChatGPT or Gemini.
What changes exactly from August 2
The law, approved in 2024, is being rolled out in stages and this phase activates Article 50, devoted to transparency. The obligations apply to any AI system used in four specific situations, and not only to those considered high risk:
1. Direct interaction with people
Those who design chatbots, virtual assistants or automated phone systems must ensure that the user always knows they are talking to an AI. AI agents also fall within the scope of this obligation: if the provider cannot predict whether the agent will interact with a human, it must be designed to reveal its nature in all such situations.
2. Synthetic content
Generative text, image, audio or video systems must label their outputs in a machine-readable format, so that it can be detected that they have been artificially created or manipulated. The European Commission is developing a standardized EU label for this marking.
3. Emotion recognition and biometric classification
Those who deploy this type of system must inform the people affected that they are being analysed.
4. Deepfakes and public-interest texts
Those responsible who generate deepfakes must indicate that the content has been manipulated, and those who publish AI-generated texts on matters of general interest —politics, health, economy, public safety— must disclose it. There is a key exception: if the text has been reviewed by humans and a person or entity assumes editorial responsibility, the notice is not required. That said, the Commission clarifies that a merely formal review, such as correcting spelling, does not count as editorial control.
Who watches and how much it costs to fail
The big institutional novelty is that the European AI Office takes over from today the supervision of general-purpose models: it will be able to require information from their providers, access the systems to evaluate them, order corrective measures and even restrict their availability if it detects non-compliance. Surveillance is shared with national authorities, which will control the rest of the tools deployed in each member state, while the European Data Protection Supervisor will do the same with the EU institutions.
As support, Brussels has set up a scientific panel of 60 independent experts, appointed a chief scientific adviser and enabled reporting channels for citizens and companies. Sanctions for failing to comply with the transparency obligations can reach 15 million euros or 3% of the infringing company’s annual worldwide turnover, whichever figure is higher; the authorities must take proportionality into account in the case of SMEs.
A calendar that continues
The roll-out of the law continues. On 2 December 2026 the transitional period for generative AI systems already marketed before today —the “AI Omnibus” package, agreed in May, grants them until that date to adapt the machine-readable marking— will end, and new prohibitions will come into force, such as those affecting systems capable of generating explicit sexual content without consent. In December 2027 the rules for high-risk AI in healthcare, education or employment will arrive, and in August 2028, those for systems embedded in regulated products, such as medical devices or vehicles.
Meanwhile, the sector is already positioning itself: 25 companies, including Google, Microsoft, Amazon, IBM and OpenAI, have voluntarily joined the code of good practice promoted by the Commission, while Meta and the big Chinese companies have decided not to join. The EU is betting on a model in which trust is the prerequisite for AI adoption; from today, that trust has a name: transparency.
