Most EU AI Act obligations take effect August 2, requiring chatbots to identify themselves and deepfakes to be marked

A new era for AI transparency

On 2 August, the bulk of the EU's AI Act comes into force, marking the world's first comprehensive AI regulation. From this date, providers of AI systems must inform users they are interacting with an artificial entity, and all content generated or manipulated by AI for professional purposes must be clearly identified. The European Commission's AI Office will start supervising general-purpose AI models, with powers to demand information, evaluate systems, and restrict them in case of non-compliance.

Artificial intelligence is a transformative technology that can bring extraordinary benefits to our citizens and businesses. But we are also seeing that it can cause harm if not designed and used properly, and that the most advanced models pose risks of a completely new magnitude.

Identifying chatbots and deepfakes

The most visible change for consumers: chatbots and virtual assistants must disclose that they are machines. Any AI-generated image, video, or audio used professionally must carry a watermark or label. The regulation also covers deepfakes more broadly, requiring that they be identifiable. However, a stricter mandatory labelling scheme for certain synthetic media, as well as the explicit prohibition of sexual deepfakes and child pornography generation by AI, will only take effect on 2 December 2026.

Banned practices already in force

Since 2025, a range of AI applications deemed to carry "unacceptable risk" have been banned across the EU. These include systems that use subliminal or manipulative techniques, exploit the vulnerabilities of minors or disabled persons, deduce race or sexual orientation via biometric categorisation, or implement social scoring. Also prohibited are emotion inference in workplaces and schools, mass scraping of internet images to create facial recognition databases, and real-time remote biometric identification in public spaces for law enforcement. Fines for using such systems can reach 35 million euros, with 15 million euros for violations related to high-risk AI.

Oversight and enforcement

The European AI Office will share supervision with national authorities. It will focus on providers of general-purpose AI models and systems under the Commission's competence, while member states monitor the rest. The European Data Protection Supervisor will also play a role in enforcement. Providers can face corrective orders and even market restrictions if they fail to comply.

Europe anticipated this evolution. With the AI Act, we have established a clear, risk-based and durable framework for trustworthy AI, which offers legal certainty to innovators while protecting the public interest. With the start of its implementation, we are taking an important step towards AI that people and companies can understand and trust, and whose benefits are widely shared throughout our society.

Phased implementation until 2028

The Act, approved in 2024, is being rolled out in stages. The 2 August deadline triggers most transparency duties, rules for many AI systems, and obligations for both providers and users. Additional requirements will follow in December, with full implementation expected by 2028. The Commission has also adopted a "Digital Omnibus" simplification measure to streamline some obligations.