The European Union has published a voluntary Code of Practice for AI content labelling ahead of binding transparency duties taking effect on 2 August 2026. The code offers a recognised route to compliance, but companies remain subject to Article 50 of the AI Act whether they sign it or not.
With less than two months between publication and enforcement, providers and deployers must translate a high-level obligation into metadata, interfaces and editorial processes that work across text, image, audio and video.
Providers and deployers carry different responsibilities
The companies building generative models are expected to mark output in a machine-readable form. That provenance must survive ordinary distribution so downstream services can detect that material was generated or manipulated.
Deployers—the organisations placing models inside products—handle visible disclosure. Deepfakes and AI-generated or altered text on matters of public interest require labels, particularly where public-interest text is published without human editorial review.
Interactive systems must also tell people when they are dealing with a machine. A customer-service agent should disclose its nature early enough for the user to understand the interaction and decide whether to continue.
The code relies on open technical standards and a common EU icon to reduce inconsistent labelling. Shared cues can help users, but they need accessible text and placement that remains visible on small screens and syndicated content.
Compliance requires a chain from generation to publication
A label added only in the final interface can disappear when media is downloaded, cropped or reposted. Machine-readable provenance should travel with the asset, while content systems preserve the origin, model, edit history and human-review state.
Organisations need an inventory of generative features and publishing routes. Marketing, support, newsroom, employee and partner workflows may all produce public material through different tools.
Editorial review must be defined rather than assumed. A person clicking “approve” after a superficial glance may not provide meaningful control. Teams should document who checks factual claims, manipulation risk and whether the label remains appropriate after substantial human revision.
Exceptions and edge cases require legal analysis. The Commission is still preparing separate guidelines, so organisations should record interim interpretations and update them as official detail arrives.
Quality assurance should test whether metadata survives export, compression and platform hand-offs. It should also verify that visible labels are understandable across languages and assistive technologies.
Transparency works only when labels carry useful meaning
Over-labelling can make users ignore every marker, while narrow labelling leaves deceptive content unaddressed. The system should distinguish synthetic creation, material alteration and ordinary technical editing where the law and guidance require it.
Labels are not a substitute for preventing fraud, defamation or manipulation. A clearly marked deepfake can still cause harm, and malicious actors may remove provenance deliberately. Platforms need detection, reporting and enforcement alongside disclosure.
Companies should establish ownership now: product teams implement technical marks, content teams manage visible disclosure, legal teams interpret scope and security teams test resistance to removal or spoofing.
The voluntary code may also influence procurement. Customers can ask model and platform vendors how they create, preserve and expose provenance instead of building every control downstream.
Training is needed beyond compliance teams. Designers, editors and developers make everyday decisions about disclosure placement and metadata, so they must recognise when a workflow has crossed into regulated synthetic content.
The EU is attempting to make synthetic origin legible without forbidding useful creation. Its August deadline turns that ambition into an engineering and operational requirement.
The strongest implementations will not treat a small icon as the finish line. They will preserve evidence through the content lifecycle and give people a clear explanation at the moment it matters. Transparency becomes credible when the label is consistent, durable and connected to accountable publishing decisions.
