C2PA and EU AI Act Article 50

Article 50 of the EU AI Act establishes several transparency obligations for AI systems. The obligation most closely related to C2PA requires providers of systems that generate synthetic audio, images, video, or text to mark their outputs in a machine-readable format and make those outputs detectable as artificially generated or manipulated. Separate obligations require deployers to give clear, human-facing disclosures for certain deepfakes and public-interest text.

C2PA can make an important contribution to this framework. A Content Credential can carry signed, machine-readable provenance about how an asset was created or changed, which system issued the record, which AI-related actions were declared, and whether the credential still validates against the asset. C2PA 2.4 also adds a dedicated AI Disclosure assertion and supports credentials for a broader range of media and structured-text formats.

But C2PA is not a legal opinion, an automatic Article 50 certificate, or a complete marking-and-labelling system by itself. Article 50 asks whether a system and its outputs fall within a legal scope, whether marking remains effective and detectable in practice, and whether people receive any required visible or audible disclosure. Those questions require policy, durable marking, testing, monitoring, presentation, and governance in addition to provenance metadata.

Article 50 contains several different transparency duties

Article 50 is not one general rule saying that all AI-related content must carry the same label. It assigns different duties to different actors and situations. C2PA is highly relevant to the provider marking obligation in Article 50(2), can support deployer disclosures under Article 50(4), and has little direct relevance to some other paragraphs. [1]

This separation matters operationally. The company that provides a generative image API may have to build marking and detection into the system. A customer that uses the API to publish a realistic synthetic advertisement may separately have to display a disclosure to viewers. The provider and deployer can be different organizations, and one technical control does not automatically satisfy both roles.

Providers and deployers have different responsibilities

Provider responsibilities

A provider is the person or organization that develops an AI system, has it developed, and places it on the market or puts it into service under its own name or trademark. A company can therefore become the provider of a downstream product even when it relies on a model supplied by someone else. Providers outside the European Union can also fall within scope when the output of their system is used in the EU. [1]

For Article 50(2), the provider owns the system-level capability. Marking cannot be treated solely as an optional action left to each end user after generation. The provider should determine which outputs are covered, apply the marking reliably, make detection possible, test the mechanism, document exceptions, and give downstream actors enough information to preserve and use the signals.

Deployer responsibilities

A deployer is an organization or professional actor using an AI system under its authority. Personal, non-professional use is excluded, while a company remains the deployer when employees, contractors, or freelancers operate the system on its behalf and under its responsibility. [2]

Deployers make contextual decisions that a provider often cannot make at generation time. They know whether an output will be published, whether it resembles a real person or event, whether it is intended to inform the public, whether substantive human review occurred, and how the audience will first encounter it. These facts determine whether a visible or audible disclosure is required under Article 50(4) and how that disclosure should be presented.

What Article 50(2) requires from providers

Article 50(2) applies to providers of AI systems, including general-purpose AI systems, that generate synthetic audio, image, video, or text content. Covered outputs must be marked in a machine-readable format and detectable as artificially generated or manipulated. The technical solution must be effective, interoperable, robust, and reliable as far as technically feasible, taking account of content-specific limitations, implementation costs, and the generally acknowledged state of the art. [1]

The requirement therefore has two connected parts. The output must carry or participate in a machine-readable marking scheme, and an appropriate detector must be able to recognize the relevant signal. A metadata structure that is never read, a watermark without an available detector, or a detector that works only on pristine laboratory samples would not address the complete operational problem.

Scope is narrower than every use of AI

The Commission guidelines and FAQ identify outputs that fall outside the marking duty, including short sequences of numbers, symbols, or letters; source code; outputs used only in machine-to-machine communication without human exposure; and certain closed-loop industrial or development outputs before they become final outputs. The obligation also does not apply to the extent that a system performs an assistive function for standard editing or does not substantially alter the deployer's input or its semantics. The guidelines also describe a narrow proportionality-based treatment for specified business-to-business or industrial contexts. [2]

These distinctions should be implemented as a controlled policy decision, not as a casual product toggle. A provider should record why a system, endpoint, content type, or transformation is considered in scope or exempt; who approved the classification; which guideline version was used; and what change would cause the classification to be reviewed.

C2PA support for a format does not determine whether Article 50 applies to that format. For example, C2PA 2.4 supports credentials for HTML and structured text, including source code and Markdown, while the Commission guidance treats source-code output as outside the Article 50(2) marking duty. Technical capability and legal scope are different questions.

Where C2PA fits

C2PA provides a standardized way to associate a digital asset with a signed Content Credential. The credential can identify the claim generator, record creation and editing actions, describe AI-related source types, preserve ingredient relationships, include a dedicated AI Disclosure assertion, carry signing and time evidence, and bind the provenance record to the asset. A verifier can then determine whether the credential is well formed, cryptographically valid, and trusted under a selected trust policy. [6]

Those properties make C2PA a strong candidate for the digitally signed metadata layer of an Article 50 system. A provider can use it to declare that an output was created or manipulated by an AI system, describe the relevant workflow with greater precision, and preserve a verifiable record of who issued the statement. Because C2PA is an open specification with interoperable verification tooling, downstream platforms do not need a proprietary decoder for every provider.

The C2PA signals most relevant to AI transparency

• Actions describe events such as creation, generative editing, opening, placement, transformation, and publication.
• digitalSourceType values distinguish generative creation, generative editing, algorithmic alteration, capture, composition, and ordinary digital creation at a particular stage of the workflow.
• Ingredients connect the output to parent media, prompts, models, datasets, references, or other computational inputs when disclosure is appropriate.
• The C2PA 2.4 AI Disclosure assertion can add structured information about model type, model identity, scientific domain, and the declared level of human oversight.
• The signing credential identifies the service or organization responsible for issuing the provenance claim under the verifier's trust policy.
• Hard bindings connect a standard manifest to a specific asset representation, while soft bindings can assist recovery after embedded provenance becomes separated from the content.

Why C2PA alone is not the whole solution

A valid Content Credential answers a provenance question: what signed statements are associated with this asset, and have the protected record and content binding remained intact? Article 50 asks broader operational and legal questions. It requires a provider to make covered outputs detectable, and it can require a deployer to communicate directly with people. A C2PA credential is valuable evidence, but it cannot answer those questions alone.

Metadata can disappear

Embedded manifests can be removed by social networks, content delivery networks, image optimizers, video transcoders, screenshots, copy-and-paste workflows, or applications that do not preserve the relevant format structures. C2PA is designed to reveal tampering with a credential, but it does not prevent the credential from being removed entirely. Durable Content Credentials use soft bindings, such as fingerprints or invisible watermark identifiers, to help recover a candidate manifest from a repository. [7]

A provider relying on C2PA should therefore test the real distribution paths its customers use. It should measure whether credentials survive common transformations, whether a durable signal remains detectable, whether the repository is available, and whether recovered candidates can be safely matched and validated. Coverage should be monitored continuously rather than assumed from a successful test at generation time.

A machine-readable credential is not a human-facing label

Article 50(4) requires certain deployers to disclose deepfakes and covered public-interest text to people. The Commission states that the disclosure must be understandable and perceivable without requiring specialized tools or dedicated actions. A deployer cannot satisfy that duty merely by relying on the machine-readable marking supplied by the provider. [2]

C2PA can still support this workflow. A platform can inspect the credential, determine that AI generation or manipulation was declared, and use that result as one input to a visible label. But the deployer must make the contextual classification, choose the appropriate wording or icon, ensure accessibility, and present the disclosure no later than first exposure.

C2PA does not decide legal scope or compliance

The C2PA specification does not know whether an output is a legal deepfake, whether a text concerns a matter of public interest, whether human review was substantive, whether an editing function was standard, or whether a particular exception applies. It also does not judge whether content is true or false. These are legal, factual, and contextual determinations made outside the credential.

For this reason, a product should not return a bare result such as Article 50 compliant based only on manifest validation. A more responsible result separates technical facts from policy conclusions: credential present, signature valid, trusted signer, AI disclosure present, durable marker detected, provider policy passed, deployer label required, or legal review needed.

The Code of Practice provides a practical route

The Code of Practice on Transparency of AI-generated Content is voluntary, but the Commission and the AI Board have confirmed it as an adequate tool for demonstrating compliance with the marking and labelling obligations in Article 50(2), (4), and (5). Signatories can rely on the Code's measures as a recognized compliance route. Organizations that do not sign must demonstrate that their alternative measures are adequately equivalent. [3]

The provider section of the Code reflects a layered view of marking and detection. Under the current state of the art, it describes digitally signed metadata together with imperceptible watermarking for most generated media, while fingerprinting and logging can provide supplementary support. It also addresses secure keys, preservation of existing markings, detection access, testing, interoperability, and richer provenance information. The deployer section addresses clear labels, icons or other disclosures, placement at first exposure, audio disclosures, and accessibility.

The Code does not make C2PA the only permitted metadata technology. C2PA is nevertheless a natural fit for the signed metadata component because it provides an open provenance format, cryptographic binding, standardized assertions, certificate-based signing, and interoperable verification. A provider can combine that layer with an appropriate watermark, optional fingerprinting, detection services, and retained logs.

A layered provider architecture

1. Classify the system and output. Determine whether Article 50(2) applies, which exception rules are relevant, and which policy version governs the decision.

2. Create the final output representation. Apply any content-affecting watermark or other durable mark before calculating the C2PA hard binding.

3. Construct and sign the Content Credential. Include the minimum accurate AI-related actions, source types, and disclosures required by policy; avoid unnecessary personal or confidential information.

4. Verify the completed output independently. Confirm that the manifest, signature, trust state, hard binding, and required policy fields validate after the final encoding step.

5. Register the evidence. Store the immutable asset hash, manifest, signing operation, certificate and time evidence, policy result, and any watermark or detector identifiers.

6. Make detection available. Operate or integrate detectors for the selected marks, publish technical documentation for relevant stakeholders, and version the detector and thresholds.

7. Monitor production and downstream transformations. Sample delivered content, test common conversions, group failures, alert on lost markings, and retain historical observations.

Deployer disclosure duties require a separate layer

Deepfakes

A deployer must disclose AI-generated or manipulated image, audio, or video content that constitutes a deepfake. The Commission guidance treats this as a contextual test involving resemblance to an existing or plausibly existing person, object, place, entity, or event, together with a false appearance of authenticity or truthfulness. Audience expectations and deployment context can matter. [2]

A C2PA credential may show that an image was generated by a model or that a voice was synthetically modified. That is useful evidence, but it does not by itself establish that the legal deepfake criteria are met. The deployer needs a review process that combines provenance facts with the subject, resemblance, message, context, and likely audience expectations.

When disclosure is required, it must be clear and distinguishable at first exposure. For evidently artistic, creative, satirical, fictional, or analogous works, Article 50 allows an appropriate disclosure that does not hamper the display or enjoyment of the work. C2PA can remain attached as detailed machine-readable provenance behind that lighter-touch presentation.

Public-interest text

Deployers must disclose AI-generated or manipulated text that is published to inform the public on matters of public interest, unless the content underwent substantive human review or editorial control and a natural or legal person holds editorial responsibility for the publication. The Commission FAQ emphasizes that superficial grammar or spelling checks do not amount to substantive review. [2]

A C2PA credential can preserve the generation and review history of a document, identify the system that created or edited it, and record a publisher's signed statement. It cannot determine whether the review was substantively adequate or whether the responsible person legally assumed editorial responsibility. Those facts need workflow approvals, documented review, and governance outside the manifest.

What evidence should be retained

A mature Article 50 programme should be able to reproduce not only the final output, but also the technical and policy decision that governed it. Useful evidence includes:

• the system, model, endpoint, and output type involved;
• the provider or deployer role assigned to the organization;
• the applicable Article 50 scope decision, exception, and policy version;
• the original and released asset hashes and immutable asset versions;
• the C2PA manifest, signing operation, certificate chain, trusted time evidence, and verification result;
• the watermark, fingerprint, or other durable-mark identifiers and detector versions;
• resilience and interoperability test results for relevant formats and transformations;
• the visible or audible disclosure presented to people, including placement and accessibility evidence;
• human-review or editorial-control records when an exception depends on them;
• monitoring observations, incidents, remediation, and approved exceptions; and
• the code, guideline, trust-list, and internal-policy versions used at the time.

The objective is reproducibility. Months later, the organization should be able to explain what the system produced, how it was marked, which detector recognized it, which policy applied, what the user saw, and why the organization considered the control adequate at that time.

Four practical scenarios

1. An image-generation API

A company provides a text-to-image API under its own brand to European customers. The company is likely acting as the provider of the AI system. For covered outputs, it can apply an imperceptible watermark, add a C2PA manifest describing generative creation and model-related disclosure, sign the result, verify the completed image, and retain a corresponding evidence record. The API customer may later become a deployer with a separate duty to display a label if it publishes content that qualifies as a deepfake.

2. A synthetic spokesperson in an advertisement

An advertising agency uses an AI video service to create a realistic synthetic spokesperson resembling a real or plausibly real person. The upstream service's machine-readable marking supports provenance and detection. The agency, as deployer, still needs to determine whether the result constitutes a deepfake and, if so, provide an understandable visible or audible disclosure at first exposure. A Content Credential alone is not enough for that presentation duty.

3. A newsroom article drafted with generative AI

A newsroom uses an AI system to prepare a draft about a matter of public interest. Editors investigate the claims, rewrite substantive sections, fact-check sources, approve the article, and the publisher assumes editorial responsibility. The public-facing labelling obligation may fall within the human-review and editorial-responsibility exception, depending on the facts. C2PA can still preserve internal and external provenance, but it should not automatically convert workflow telemetry into a legal exemption result.

4. AI-assisted standard editing

A photo application applies routine denoising, sharpening, color correction, or another assistive operation that does not substantially alter the user's input or its semantics. The Article 50(2) marking duty may not apply to the extent the function qualifies as standard editing under the guidelines. The provider should document that classification and monitor feature changes. C2PA may still record the action for provenance even when the legal marking obligation is not triggered.

Common misunderstandings

Article 50 readiness checklist

For providers

• Inventory every system, model integration, endpoint, and output format that may generate synthetic content.
• Document provider status, EU territorial scope, output classification, exclusions, and exception logic.
• Select a layered marking architecture appropriate to each content type and current state of the art.
• Apply content-affecting marks before C2PA signing and verify the final released output.
• Use accurate, minimal C2PA actions, source types, ingredients, and AI disclosures.
• Protect signing and watermark keys, separate test and production environments, and audit every signing operation.
• Provide usable detection capabilities and versioned documentation to relevant stakeholders.
• Test effectiveness, interoperability, reliability, and robustness across real transformations.
• Preserve existing markings when processing third-party content and prohibit unauthorized removal.
• Monitor marking coverage, detector performance, credential survival, repository availability, and policy drift.

For deployers

• Identify professional uses that may create deepfakes or public-interest text.
• Create documented decision criteria for deepfake scope, public-interest publication, artistic context, and human-review exceptions.
• Inspect available machine-readable provenance, but do not treat it as the only source of truth.
• Present required disclosures clearly, distinguishably, accessibly, and no later than first exposure.
• Retain evidence of the label, placement, language, audio treatment, icon, and accessibility behavior.
• Record substantive review and editorial responsibility when relying on the public-interest-text exception.
• Reassess disclosure when content is repurposed, distributed to a different audience, or placed in a new context.

Dates, enforcement, and the current transition

Article 50 has applied since 2 August 2026. A limited transition applies only to Article 50(2) for AI systems placed on the market before that date: providers of those systems must take the necessary steps to comply by 2 December 2026. Content generated before 2 August 2026 does not have to be labelled retroactively, although the Commission encourages voluntary labelling where possible. [4]

Enforcement will primarily involve national market-surveillance authorities, with defined roles for the AI Office and the European Data Protection Supervisor in specified circumstances. Article 99 allows administrative fines for Article 50 non-compliance of up to EUR 15 million or, for an undertaking, up to 3% of worldwide annual turnover for the preceding financial year, whichever is higher. For SMEs, the applicable cap is the lower of the amount and percentage, subject to the Regulation's other rules. [5]

These dates and consequences reinforce the need for historical evidence. An organization should be able to show which systems were already on the market, which transition applied, when technical controls became active, how effectiveness was tested, and how exceptions or incidents were handled.

The simplest mental model

• Article 50(2) is primarily a provider duty to mark covered synthetic outputs and make them detectable.
• Article 50(4) is primarily a deployer duty to give people clear disclosures for covered deepfakes and public-interest text.
• C2PA supplies signed, machine-readable provenance and evidence; it does not make the legal classification.
• Durable watermarking or fingerprint-based recovery can help when metadata is stripped.
• Visible or audible labels remain necessary when the deployer disclosure duty applies.
• Testing, monitoring, policy versioning, and audit evidence turn technical components into an operational compliance system.
• The safest product language reports technical facts and policy outcomes separately from legal conclusions.

C2PA can be a central part of an Article 50 architecture because it gives AI-generated and manipulated content a portable, signed provenance record. Its greatest value appears when it is combined with durable marking, detection, human-facing disclosure, monitoring, and reproducible evidence. In that layered model, Content Credentials do not claim to solve the law by themselves. They make the technical history of each output inspectable, verifiable, and governable.

Sources and notes

[1] EU AI Act, Article 50 , Article 3 definitions, and Article 2 scope. Provider and deployer roles, territorial reach, machine-readable marking and detectability, deepfake and public-interest-text disclosures, timing, and accessibility of information.

[2] European Commission, Guidelines on Article 50 transparency obligations and Article 50 questions and answers. Scope, exclusions, standard editing, provider and deployer responsibilities, deepfake analysis, visible disclosure, and substantive human review.

[3] European Commission, Code of Practice on Transparency of AI-generated Content. Approved voluntary compliance route for Article 50(2), (4), and (5), including provider marking and detection measures and deployer labelling measures.

[4] Regulation (EU) 2026/1744, Digital Omnibus amendment, and the Commission Article 50 FAQ. Application from 2 August 2026 and the limited 2 December 2026 transition for pre-existing systems under Article 50(2).

[5] EU AI Act, Article 99 - Penalties. Maximum administrative-fine framework for Article 50 non-compliance and SME treatment.

[6] C2PA Technical Specification 2.4, Content Credentials specification. Actions, digital source types, ingredients, AI Disclosure, claims, signatures, trust, structured-text support, and asset bindings.

[7] C2PA 2.4, Soft Binding Resolution API and C2PA Explainer. Watermark- and fingerprint-assisted recovery, manifest repositories, and the limits of provenance as a truth signal.