Summary

• Article 50 duties begin applying across the European Union on 2 August 2026.
• Providers and deployers carry different obligations covering AI interaction, synthetic media, biometrics, and public-interest content.
• Compliance requires changes to interfaces, publishing processes, procurement, documentation, and supplier oversight.

The European Commission has published implementation guidance for the AI Act’s transparency provisions, giving technology providers and organisations using artificial intelligence a more precise account of duties that begin applying on 2 August 2026.

Rather than requiring one universal label for every use of AI, Article 50 distinguishes between several technologies and operating roles. Providers of systems that interact directly with people must make the artificial nature of the interaction clear, while providers generating or manipulating audio, images, video, or text must support machine readable detection of the content’s origin.

Organisations deploying AI acquire separate disclosure duties in specified settings, including deepfakes, emotion recognition, biometric categorisation, and some public-interest text produced without human review or editorial control. A supplier may provide technical marking, but the organisation publishing or operating the system can still retain responsibility for what the public sees.

Because these provisions have been known for some time, the guidance does not establish an unexpected direction for European regulation. It narrows the implementation questions that product teams, publishers, public bodies, procurement departments, and legal functions must settle before the obligations move into active supervision.

Provider and deployer duties do not always align

Enterprise AI usually reaches the public through a chain of suppliers, integrators, customers, contractors, and publishing systems. A model provider may insert metadata into generated material, while a customer edits the file, passes it through another platform, and distributes it through a channel that removes or ignores the original marker.

Under that structure, compliance cannot rest solely on a vendor’s product documentation. Organisations must test whether disclosures and machine readable signals survive compression, conversion, editing, export, and transfer between systems, particularly where several suppliers participate in the same workflow.

Direct interaction raises a similar problem because AI is often embedded inside a broader service rather than presented as a clearly labelled chatbot. Automated systems can draft replies, recommend decisions, speak on a telephone line, or guide a user through an application while the surrounding organisation remains the visible service provider.

Interfaces will need to disclose the artificial role without overwhelming people with notices that become routine and unread. The Commission’s Article 50 questions and answers provide detail on scope, but individual organisations must still decide where a disclosure appears and how it remains understandable across devices and service channels.

Systems already on the market receive a limited transition for certain technical marking and detection functions, which must be brought into line by 2 December 2026. That extension offers additional engineering time, although it does not remove the wider duties applying from August.

Procurement becomes part of transparency compliance

Contracts for generative and interactive AI will need to establish what a supplier actually provides. Buyers should understand which formats receive markers, whether exported content retains them, how product interfaces display AI interaction, and who must investigate when a technical control disappears downstream.

Where a service relies on subcontractors or external models, the principal supplier may not control every part of the content chain. Procurement teams will need enough visibility to identify those dependencies and determine whether a change of model, file processor, or content platform alters the compliance position.

Public bodies face an additional burden because material can move between internal systems, external contractors, communications teams, and citizen-facing services. A disclosure applied at the point of generation may be absent by the time the content appears in a public notice, automated reply, or administrative decision process.

The Commission has also supported a voluntary code of practice covering AI generated content. Following the code may provide a clearer route for documenting controls, but the underlying legal duties remain binding whether an organisation adopts the code or uses a different technical method.

Enforcement will determine how far providers and deployers must go when marking fails or content has been deliberately altered to remove it. National market surveillance authorities and courts will shape the practical boundaries, which may produce differences in supervisory intensity even though the legislation applies across the European Union.

Companies operating in several countries should consequently build a common evidence base around system inventories, risk ownership, testing, and supplier records. A policy document stating that AI content will be labelled offers little protection if production systems cannot show when a marker was added, preserved, or lost.

Transparency obligations can also reach tools bought initially for internal productivity. Customer service, recruitment, marketing, regulatory reporting, and public communications can expose AI generated material outside the organisation, even where the original purchase was treated as an employee assistant rather than a public system.

By August, affected organisations need to know which systems they provide or deploy, where people encounter them, and whether their technical controls survive ordinary use. The Commission has supplied a clearer rulebook, while implementation will depend on the less tidy interaction between software design, supplier contracts, human review, and production operations.