The European Union has begun applying a new set of AI transparency requirements, giving regulators authority to penalize companies that fail to identify chatbots, label deepfakes or disclose certain other uses of artificial intelligence.
The provisions, which took effect on Sunday under Article 50 of the EU AI Act, require people to be informed when they are interacting directly with an AI system rather than a person. Providers of systems that generate or alter text, images, audio and video must also make synthetic output detectable, including through machine-readable markings.
Organizations deploying AI face additional disclosure duties. Deepfakes must be identified as artificially generated or manipulated, while the use of emotion-recognition and biometric-categorization systems must be disclosed to affected people. AI-generated text published to inform the public on matters of public interest generally requires a label when it has not received human review or editorial oversight.
The obligations potentially reach far beyond major technology developers. Banks using automated customer-service agents, employers and recruitment firms deploying workplace tools, media organizations publishing synthetic material, and other businesses offering AI-assisted services may fall within the rules. The requirements also apply to companies based outside Europe when their systems or outputs are offered or used in the EU.
Breaches of the transparency provisions can result in penalties of as much as €15 million or 3% of a company’s worldwide annual turnover, whichever is higher. Enforcement will primarily involve national market-surveillance authorities, while the European AI Office has responsibilities connected to general-purpose AI models and other areas governed at the EU level. Regulators can investigate suspected violations and seek information from companies.
The EU is expanding the Brussels-based AI Office with 38 additional personnel to monitor companies ranging from startups to large American and Chinese technology groups. It has also introduced channels through which technology workers and users can confidentially report suspected unlawful conduct.
“As enforcement begins, we are taking an important step towards AI that people and businesses can understand and trust,” European Commission Executive Vice-President Henna Virkkunen said.
The transparency rules are intended to make it harder to pass off machine-generated material as authentic human communication. The issue has become more urgent as generative systems produce increasingly realistic voices, photographs and video, creating new opportunities for impersonation, fraud, nonconsensual sexual imagery and political manipulation. Machine-readable indicators are also designed to help online services and detection tools recognize synthetic media at scale, although their effectiveness will depend on implementation and whether identifying information survives editing or redistribution.
The requirements include transitional arrangements for some generative AI systems that were placed on the market before Aug. 2. Content published before the provisions became applicable does not have to be relabeled retroactively.
The AI Act uses a risk-based framework rather than imposing one set of restrictions on every system. Some practices considered to pose unacceptable risks are prohibited, while high-risk applications in areas such as employment, essential services and critical infrastructure face stricter governance requirements. General-purpose model providers are subject to a separate set of duties, with additional obligations for models considered capable of creating systemic risks.
The law’s phased rollout has made compliance a continuing process rather than a single deadline. Companies must determine which role they occupy under the legislation—such as provider or deployer—and document how their products satisfy the relevant requirements. The latest provisions place those responsibilities directly in front of users by requiring visible disclosures and technical signals when AI is involved.