EU AI Act transparency rules: what the new labeling requirements mean

The European Union's AI Act, the bloc's landmark framework regulating artificial intelligence, has begun rolling out new transparency obligations designed to make it easier for people to tell when they are interacting with a chatbot or viewing content that was generated or altered by AI. The rules took effect this month as part of the law's phased implementation timeline, which introduces different obligations at different dates depending on the risk category assigned to each type of AI use.

At the centre of the new rules is a distinction between two kinds of companies: providers, meaning the companies that develop and market AI systems, and deployers, meaning the platforms and services that put those systems to use for their own customers or users. Some companies, particularly large technology firms that both build AI models and operate consumer-facing products, fall into both categories simultaneously and must meet obligations that apply to each role.

Providers are required to design AI systems so that synthetic audio, image, video or text content they generate can be identified as artificially produced or manipulated, typically through machine-readable markers embedded in the file, sometimes called watermarking, though the technical method is left partly to providers to implement as long as it achieves reliable detection.

Deployers, meanwhile, must ensure that people are clearly informed when they are interacting with an AI system rather than a human, unless that fact is already obvious from the context. This obligation is squarely aimed at chatbots and voice assistants used in customer service, healthcare information lines and similar settings where a user might otherwise reasonably assume they are speaking with a person.

A related obligation targets what the law calls deepfakes: AI-generated or manipulated image, audio or video content that depicts real people, places or events in a way that would appear authentic to a reasonable person. Anyone who generates or shares this kind of content is now required to disclose that it has been artificially created or manipulated, with narrower exceptions for clearly satirical or artistic works, provided the disclosure does not undermine the work's expression.

The European Commission has published a set of standard AI labels that companies can adopt directly rather than designing their own disclosure language and iconography from scratch, an approach intended to create some visual consistency across platforms and make the labels more recognisable to users over time, similar to how standardised nutrition labels function on food packaging.

Enforcement falls to national regulators in each EU member state, coordinated at the union level by the AI Office, which can levy fines for non-compliance that scale with a company's global annual turnover, mirroring the enforcement structure used for the EU's General Data Protection Regulation. Early enforcement priorities are expected to focus on the largest AI providers and platforms rather than smaller developers, though the obligations technically apply broadly.

The transparency rules are one part of a broader, staggered rollout under the AI Act, which separately bans a small number of AI uses considered to pose unacceptable risk, such as social scoring by governments, and imposes stricter obligations on AI systems classified as high-risk, including certain uses in hiring, credit scoring, law enforcement and critical infrastructure, with those obligations phasing in on a later timeline than the transparency rules that took effect this month.

Companies operating outside the EU are not exempt if their AI systems are used by people within the bloc, a scope provision similar to how GDPR applies to any company handling EU residents' data regardless of where the company itself is based. This has pushed many global AI providers to build EU-specific labeling and disclosure features into their products rather than maintaining separate versions for different markets.

For everyday users, the practical effect of the new rules should be a growing number of visible AI labels on generated images, clearer disclosures at the start of chatbot conversations, and disclosure statements attached to AI-manipulated media shared online, though how consistently and quickly platforms roll these features out in practice will depend heavily on how actively national regulators enforce the new requirements in the coming months.