The European Code of Practice on AI Content Transparency

What the Code of Practice on Article 50 asks of inference pipelines, APIs and client interfaces.

Introduction

The regulatory picture for artificial intelligence in Europe is changing. On 2 August 2026 the transparency obligations in Article 50 of the AI Act begin to apply, and the Code of Practice on Transparency of AI-Generated Content arrives alongside them as the main route to demonstrating compliance. The aim behind it is simple: people are entitled to know when the content in front of them has been generated or manipulated by artificial intelligence.

The work lies in applying it. Provenance metadata disappears the moment a CDN transform re-encodes a file. Watermarking standards are still undeveloped, particularly for text. The Code answers each of those problems at specific points in the architecture and five of its provisions will shape roadmaps over the next two release cycles.

The framework in practice

1. Multi-layer marking

The Code starts from an explicit technical finding: with the current state of the art, no single marking technique satisfies all four requirements of Article 50(2) at once, namely effectiveness, interoperability, robustness and reliability. Signatories therefore apply two layers, signed and timestamped metadata together with an imperceptible watermark embedded in the content itself, so that each layer covers the weak points of the other.

2. Detection as an obligation

Marking only makes practical sense where a detection mechanism exists to go with it. Signatories have to make a detection solution available free of charge, whether as a public specification, distributable software or an API, and open it to the public wherever the public may be exposed to the content.

3. A candid treatment of the limits of text

Text is the weakest link in provenance marking, and the Code acknowledges as much. Content under 200 tokens falls outside the watermarking obligation, because current techniques cannot mark passages that short with even basic reliability.

Free-form text above that threshold does have to carry a watermark, though the Code itself accepts that its reliability will be lower than for longer text. A practical consequence follows: while reliability stays low, a provider may limit access to the text detector to verified expert users, among them regulators, media, fact-checkers and researchers, because a low-confidence result shown to the general public misleads more easily than it clarifies.

Methods that recognise content carrying no prior mark, working only from the traces the model leaves behind, are what the Code calls “forensic detection”. It keeps them as an optional measure and records that, at the time of publication, they fell short of the maturity needed to meet the requirements of Article 50(2). For an engineering team the reading is direct: the mark goes in at generation, because no alternative exists today that would let you add it afterwards.

4. Binding interoperability

Signatories have to implement an interoperability solution for their detection mechanisms by 2 February 2027. Measure 3.4 opens four routes:

• A standard access method that routes each query to the detection mechanism it belongs to.

• A publicly readable signpost embedded in the content, indicating which detector to consult.

• A shared detection solution, run by a consortium that any signatory can join, SMEs included.

• Any other alternative that reaches an equivalent degree of interoperability.

This is the only hard deadline in the document, and whether verification proves workable rests on it.

5. Disclosure specifications

Section 2 of the Code turns the disclosure duty into rules on design and placement. The label consists of the capitalised acronym “AI”, with a freely available EU icon in the “AI GENERATED” and “AI MODIFIED” variants. User testing across several Member States showed that the variants carrying explicit text work better than the icon on its own.

Accessibility rules are mandatory in every case: the European Accessibility Act, WCAG 2.1 Level AA, screen-reader detectability and haptic cues for audio.

One point worth flagging is that AI-generated published text falls outside the labelling duty once it has been through human review and an identified natural or legal person takes editorial responsibility, with contact details published (Article 50(4)).

Conclusion

• Two-layer marking sets a defensible baseline.

• Free, continuous access to detection turns marking into verifiable transparency.

• The interoperability deadline of February 2027 decides whether that verification works at scale.

• The design and placement specifications give front-end teams a clear standard, and the testing behind the EU icon lends it credibility.

One caveat. The Code states expressly that adherence does not constitute conclusive evidence of compliance with Article 50. Signatories gain a structured, predictable route, along with a clear indication of what market surveillance authorities will examine, while ultimate responsibility for the underlying obligations stays with the provider or the deployer.

Marking, detection and labelling are becoming baseline infrastructure for anyone placing a generative system on the European market, and the timelines are short. Teams that begin the pipeline work now, ahead of settled standards and benchmarks, will pick their own implementation order rather than inherit one.

The full text of the Code, along with the material published alongside it, is available on the European Commission’s official page: https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content

Source: Code of Practice on Transparency of AI-Generated Content, Sections 1 and 2; Article 50 of the AI Act.