EU AI Act Article 50(2): AI Content Marking — C2PA, Watermarks & Machine-Readable Compliance Guide

Every AI-generated image, video, audio clip, and text must carry a machine-readable mark identifying it as AI-created. Article 50(2) of the EU AI Act makes this a legal requirement — with C2PA content credentials as the primary standard and watermarking as a durable complement. This guide explains the technical requirements, the Code of Practice, and exactly what providers must implement.

1. What Article 50(2) Requires

“Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated. Providers shall ensure their technical solutions are effective, interoperable, robust and reliable as far as this is technically feasible, taking into account the specificities and limitations of various types of content, the costs of implementation and the generally acknowledged state of the art…” — EU AI Act, Article 50(2)

This is a provider-side obligation. If you build, develop, or place on the market an AI system that generates synthetic content — images, videos, audio, or text — you must ensure that every piece of output carries a machine-readable mark identifying it as AI-generated. The mark must be detectable by tools, not just visible to humans.

The obligation is qualified by three important modifiers: technical feasibility (what is possible with current technology), cost of implementation (proportionate to the provider’s resources), and state of the art (the best currently available techniques). These modifiers mean the obligation is not absolute — but they also mean that as technology improves, the compliance standard rises with it.

2. Who Must Comply

Article 50(2) targets providers of AI systems that generate synthetic content. This includes:

3. Content Types Covered: Images, Video, Audio, and Text

Article 50(2) covers four categories of synthetic content, each with different technical maturity for marking:

The “state of the art” qualifier in Article 50(2) implicitly acknowledges these differences. Providers are expected to implement the best available marking for each content type, not achieve uniform detection rates across all four.

4. The Three Marking Mechanisms

The Code of Practice on AI Content Marking (finalized June 2026) endorses three complementary mechanisms. Providers are expected to use at least the first two; the third is optional but encouraged.

Mechanism 1: Digitally Signed Metadata (Primary)

Provenance information embedded in the content’s metadata, cryptographically signed and timestamped in a tamper-evident manner. This is the core compliance mechanism. The standard endorsed by the Code of Practice is C2PA (see Section 5).

Mechanism 2: Imperceptible Watermarking (Complementary)

Invisible signals embedded directly in the content — pixel patterns for images, waveform modifications for audio, token distribution patterns for text. Watermarks survive transformations that destroy metadata (screenshots, re-uploads, format conversions). They are “robust but lower-bandwidth” — they can confirm AI generation but carry less detailed provenance information.

Mechanism 3: Fingerprinting with Registry (Optional)

Perceptual hashing of generated content, logged to a queryable database. This allows anyone to check whether a piece of content was generated by a registered AI system. The Code of Practice treats this as supplementary rather than baseline, due to the operational demands of maintaining a registry.

5. C2PA Content Credentials: The Primary Standard

The Coalition for Content Provenance and Authenticity (C2PA) is an open technical standard developed by Adobe, Microsoft, Intel, and other major technology companies. The EU Code of Practice identifies C2PA as the sole technology currently meeting the requirements for digitally signed metadata under Measure 1.1.

How C2PA Works

C2PA Content Credentials are cryptographic records attached to a piece of content. They include information about how the content was created (AI-generated, AI-edited, camera-captured), who created it (the provider’s identity), and when it was created (timestamp). The records are digitally signed, making them tamper-evident — any modification to the content after signing is detectable.

What C2PA Records Contain

6. Imperceptible Watermarking: The Durable Complement

Watermarking embeds signals directly into the content itself, making them survive transformations that strip metadata. For images, this means subtle pixel-level patterns invisible to the human eye but detectable by algorithms. For audio, inaudible frequency modifications. For video, per-frame or temporal watermarks.

Watermarking by Content Type

The Code of Practice recognizes that no single watermarking technique is perfect. Providers are expected to use the best available technology and to complement watermarking with metadata marking. The combination provides stronger detection than either mechanism alone.

7. Special Challenge: Marking AI-Generated Text

Text presents the most challenging marking problem. Unlike images or audio, text has no inherent signal space in which to embed watermarks. Copy-pasting text strips all metadata. Paraphrasing defeats statistical watermarks. These technical realities are reflected in the “technical feasibility” qualifier of Article 50(2).

Current Approaches

API-level metadata: When text is generated through an API, the API response can include provenance metadata (model name, timestamp, generation parameters). This metadata travels with the API response but is lost when the text is copied into a document or email.

Statistical watermarking: Some providers embed subtle statistical patterns in token selection — slightly biasing word choices in ways that are detectable by the provider’s detection tool but imperceptible to human readers. These watermarks are fragile and can be defeated by paraphrasing.

Visible disclosure: Where machine-readable marking is not technically feasible, the Code of Practice recommends visible disclosure as a supplementary measure. For example, AI-generated articles can include a visible label: “This text was generated with AI assistance.”

8. The Code of Practice (June 2026)

On June 10, 2026, the European AI Office published the final Code of Practice on Marking and Labelling AI-Generated Content. The Code is voluntary — providers may sign up to demonstrate compliance — but non-signatories must prove compliance through alternative means.

Key Commitments

Commitment 1 (Marking): Signatories must implement machine-readable marking using digitally signed metadata (C2PA) as the primary mechanism, complemented by imperceptible watermarking. Metadata must include at minimum an attestation of AI generation, cryptographically signed and timestamped.

Commitment 1.2 (Preservation): Signatories must retain and not intentionally alter or remove existing metadata markings when AI systems use content as input. Terms of service must prohibit users from stripping metadata.

Commitment 1.3 (Richer Metadata): Signatories are encouraged to include additional provenance information using CAWG metadata assertions within C2PA manifests — such as the AI model used, generation parameters, and input provenance chains.

Commitment 2 (Detection): Signatories must make detection mechanisms publicly available through open specifications, downloadable software, or free cloud-based APIs. By February 27, 2027, an interoperable detection solution must be implemented.

Visual Labels

The Code of Practice also defines standardised EU visual labels for deployers to use when displaying AI-generated content. These labels use “AI” (English), “KI” (German), or “IA” (French, Italian, Spanish, Portuguese) in a standardised icon format. Visual labels complement but do not replace machine-readable marking.

9. Exemptions: What Does Not Need Marking

Article 50(2) includes several exemptions from the marking obligation:

10. Detection and Interoperability Requirements

Marking content is only half the equation. Article 50(2) requires that markings be “detectable” — meaning tools must exist to verify whether content carries AI-generation marks.

The Code of Practice requires providers to offer free public detection through at least one of three channels: open specifications that anyone can implement, downloadable software for local verification, or cloud-based APIs available at no charge.

By February 27, 2027, providers must implement an interoperable detection solution using publicly available industry standards. The goal is to prevent a fragmented landscape where each provider’s markings can only be detected by that provider’s proprietary tools. Instead, a common verification infrastructure — built on open standards like C2PA — should enable anyone to check any piece of content.

11. Timeline: August 2 vs December 2, 2026

The Digital Omnibus (adopted June–July 2026) introduced a split timeline for Article 50(2):

12. Implementation Checklist

13. Penalties and Enforcement

The European AI Office has direct enforcement authority over general-purpose AI model providers under Article 50(2). National market surveillance authorities oversee other providers. Providing incorrect or misleading information to authorities carries a separate fine of up to €7.5 million or 1% of turnover.

The Code of Practice provides a compliance pathway: signatories who implement its measures are presumed to comply with Article 50(2). Non-signatories must demonstrate compliance through alternative means — which may require more extensive documentation and evidence.

14. Frequently Asked Questions

Do I need to mark AI-generated content used only internally?

If the content is purely machine-to-machine or remains within a closed development environment, the exemption may apply. However, if internally generated content is later published, shared externally, or viewed by people who might not know it is AI-generated, marking is required from the point of generation.

What if my API provider already marks content?

If the upstream API provider (e.g., OpenAI, Anthropic) already embeds C2PA metadata or watermarks, your obligation is to preserve those markings. You must not strip them during processing. If you modify the content significantly, consider adding your own provenance record to the chain.

Does the Code of Practice apply to non-EU providers?

The Code of Practice is voluntary and open to any provider worldwide. However, Article 50(2) applies to all providers whose AI-generated content is available in the EU, regardless of where the provider is headquartered. Non-EU providers can sign the Code of Practice or demonstrate compliance independently.

What about AI-assisted content where a human heavily edited the output?

If an AI generates a first draft and a human substantially rewrites it, the final product may no longer qualify as “synthetically generated.” However, the line between assistance and generation is not clearly defined. Where AI contribution is significant, marking is the safer approach.

Can users opt out of watermarking?

No. Article 50(2) does not provide a user opt-out for machine-readable marking. Providers must mark all outputs regardless of user preference. Terms of service should clarify this requirement and prohibit the intentional removal of markings.