EU AI Act · Article 50
Be the agency that hands over AI creative with proof.
Since August 2, 2026, the EU AI Act's transparency rules have put marking duties on AI model providers and disclosure duties on the organizations that publish AI content. In practice, brands and platforms now ask the agencies in between for provenance. Archibal is that documentation: signed C2PA credentials, a content fingerprint, and a provenance record for the composite deliverable no single tool's marking covers.
Does this apply to you?
If any of these sound like you, the clients and platforms you work with can now ask you for provenance, and some already do.
You run an agency or studio
You generate images, video, or copy with AI for client campaigns that reach EU audiences.
You are a brand in-house team
Your team produces AI-assisted creative and publishes it to the market yourselves.
You are a production house
You deliver AI content to clients who need to prove how it was made.
You run a platform
Users publish AI-generated content through your product to an audience.
One obligation on a widening timeline
Article 50 is in force. Other AI-labeling rules apply already or have dates set; the timeline below shows each one with its source.
EU AI Act · Article 50
AI media reaching EU audiences has to carry machine-readable disclosure.
Synthetic image, audio, video, and text needs an embedded, verifiable marking that a platform or regulator can check.
All EU-facing agencies, brands, and platforms
Most compliance tools tell you what to do. Archibal produces the proof.
Law firms
Interpret the regulation and send you a memo. Useful once. It doesn't touch a single asset.
Consultancies
Produce a gap-analysis report and leave. You still have to build and prove the marking yourself, per campaign.
Archibal
Generates the signed manifests, cryptographic fingerprints, and records that document exactly how your AI creative was made - the provenance your clients and the platforms you publish to now require - automatically, every time an asset is created or delivered.
From brief to provable delivery
The record builds itself as you work - no separate documentation step.
Brief
Accept the project brief with its territories and channels.
Create
Build with your existing AI tools. Archibal captures what was used.
Review
Risk assessment runs automatically. Issues surface before delivery.
Deliver
Signed assets, a provenance record, and a verification link in one step.
Done
The client has proof. You have a permanent, verifiable record.
Who each law binds - and how Archibal helps
Most of these obligations fall on the AI model providers and the platforms, not on agencies - but they drive the provenance your clients now demand. Here is what Archibal hands you for each.
EU AI Act - Article 50(2)
Who it binds / what it requires
Binds AI model providers (Midjourney, OpenAI, Adobe) to mark synthetic image, audio, and video machine-readably. Not an obligation on your agency.
How Archibal helps
Signs C2PA credentials with the cryptographic provenance method Recital 133 names - covering the composite deliverable no single tool marks.
EU AI Act - Article 50(4)
Who it binds / what it requires
Binds deployers to disclose deepfakes and certain AI content to the audience in a clear, human-readable way (a caption, label, or overlay).
How Archibal helps
Documents that the content is AI-generated and how - backing the audience-facing label you place at distribution. The label itself stays your responsibility.
EU Code of Practice on AI-generated content
Who it binds / what it requires
A voluntary technique for layered marking: metadata plus provenance credentials.
How Archibal helps
Signed manifest, provenance metadata, and a perceptual fingerprint - all three layers.
California SB 942
Who it binds / what it requires
Binds large generative AI providers (1M+ users) to watermark output and offer a public detection tool. The duty is on the provider, not the agency using the tool.
How Archibal helps
Marks your composite deliverable and makes it publicly checkable - the provenance clients and platforms increasingly require.
China - AI content labeling
Who it binds / what it requires
Binds platforms and providers to apply audience labels and invisible traceability metadata for Chinese audiences.
How Archibal helps
Supplies the signed manifest and traceability metadata your distribution partners are required to carry.
India - IT Rules amendment
Who it binds / what it requires
Binds intermediary platforms to label AI content and carry traceable metadata where feasible.
How Archibal helps
Gives the platforms you publish through the provenance they need to surface.
South Korea - AI Basic Act
Who it binds / what it requires
Binds providers and platforms to label AI content and watermark synthetic output.
How Archibal helps
Signed credentials and a fingerprint that travel with the asset into those channels.
German advertising guidance (2026)
Who it binds / what it requires
Advises advertisers to label AI visuals in ads that resemble authentic content.
How Archibal helps
A territorial risk-register entry plus a signed record documenting the disclosure decision.
C2PA Conformance Program
Who it binds / what it requires
A valid signed manifest for each media type.
How Archibal helps
Conformant signed output per asset, embedded or as a sidecar.
| Regulation | Who it binds / what it requires | How Archibal helps |
|---|---|---|
| EU AI Act - Article 50(2) | Binds AI model providers (Midjourney, OpenAI, Adobe) to mark synthetic image, audio, and video machine-readably. Not an obligation on your agency. | Signs C2PA credentials with the cryptographic provenance method Recital 133 names - covering the composite deliverable no single tool marks. |
| EU AI Act - Article 50(4) | Binds deployers to disclose deepfakes and certain AI content to the audience in a clear, human-readable way (a caption, label, or overlay). | Documents that the content is AI-generated and how - backing the audience-facing label you place at distribution. The label itself stays your responsibility. |
| EU Code of Practice on AI-generated content | A voluntary technique for layered marking: metadata plus provenance credentials. | Signed manifest, provenance metadata, and a perceptual fingerprint - all three layers. |
| California SB 942 | Binds large generative AI providers (1M+ users) to watermark output and offer a public detection tool. The duty is on the provider, not the agency using the tool. | Marks your composite deliverable and makes it publicly checkable - the provenance clients and platforms increasingly require. |
| China - AI content labeling | Binds platforms and providers to apply audience labels and invisible traceability metadata for Chinese audiences. | Supplies the signed manifest and traceability metadata your distribution partners are required to carry. |
| India - IT Rules amendment | Binds intermediary platforms to label AI content and carry traceable metadata where feasible. | Gives the platforms you publish through the provenance they need to surface. |
| South Korea - AI Basic Act | Binds providers and platforms to label AI content and watermark synthetic output. | Signed credentials and a fingerprint that travel with the asset into those channels. |
| German advertising guidance (2026) | Advises advertisers to label AI visuals in ads that resemble authentic content. | A territorial risk-register entry plus a signed record documenting the disclosure decision. |
| C2PA Conformance Program | A valid signed manifest for each media type. | Conformant signed output per asset, embedded or as a sidecar. |
Archibal produces the evidence; your legal and compliance function decides how it is used. See how the record itself is built, from capture to verification.
Audit your existing library, free
Drop in any image, video, audio, or PDF. Archibal checks whether it already carries C2PA content credentials and what AI signals it contains - no account, no upload kept.
Check a fileArticle 50, answered
Does Article 50 apply to our agency if we are not based in the EU?
The machine-readable marking duty in Article 50(2) sits on the AI model providers, wherever you are based. What can reach your agency is Article 50(4) - the deployer duty to disclose deepfakes and certain AI content to EU audiences - and, in practice, your clients' and platforms' demand for provenance. Both follow the content into the EU market regardless of your headquarters.
What counts as "AI-generated" for disclosure purposes?
The obligation covers synthetic and meaningfully AI-manipulated audio, image, video, and text. A useful default: if a reasonable person could mistake it for authentically captured or human-authored content, treat it as in scope.
Do we label every AI-touched asset, or only fully synthetic ones?
The safest posture is to mark anything with material AI generation or manipulation. Archibal records provenance on every asset you run through it, so the marking exists whether or not a given piece turns out to need it.
What is the difference between machine-readable marking and a visible label?
A visible label is the "made with AI" a person reads - the human-readable disclosure Article 50(4) asks deployers to place. Machine-readable marking is the embedded, verifiable record (C2PA credentials, a fingerprint) a platform or regulator can check - the method Recital 133 names and Article 50(2) puts on the model providers. Archibal produces the machine-readable layer; you place the visible label at distribution.
Does an Archibal record count as Article 50 compliance evidence?
Article 50(2) machine-readable marking is an obligation on the AI system providers - Midjourney, Adobe, OpenAI - not on your agency. Archibal's C2PA manifest uses the cryptographic provenance method that Recital 133 of the AI Act names as a satisfying technique. For Article 50(4) - the deployer duty to disclose AI content to audiences - Archibal provides the documented record that you knew the content was AI-generated and delivered it with full provenance. The audience-facing label at distribution (a caption, an end-card) remains your responsibility.
What if a platform strips the C2PA manifest - is the provenance lost?
No. The perceptual fingerprint is derived from the content itself, so a stripped or re-encoded copy still matches back to its original signed record in the registry. The provenance survives even when the embedded metadata does not.
How long do we need to keep provenance records?
Plan for the statute of limitations in your publishing markets - commonly several years. Archibal keeps the record in a durable archive so the record is there when a question arrives long after the campaign ended.
Win the work. Prove the provenance.
Start free and deliver your next AI project with a provenance record the client can verify themselves.