Skip to main content
Provenance metadata declares how creative content was produced — whether AI was involved, which tools were used, and what disclosure obligations the declaring party believes apply. Regulations like the EU AI Act and California SB 942 place disclosure obligations on deployers and covered platforms; AdCP carries the structured provenance signals those parties rely on through the programmatic supply chain so every participant can declare, transmit, and verify the same data. Provenance attaches to creative manifests, individual assets, or content-standards artifacts. It is a claim by the declaring party — receiving parties verify claims independently using their own detection tools.
EU AI Act Article 50 enforcement begins August 2026. California SB 942 is already in effect. Major platforms mandate AI content labeling today. AdCP carries the structured, machine-readable provenance and disclosure metadata these regulations rely on through the programmatic supply chain, where no standard for it previously existed. The protocol moves the data; the legal obligation still sits with the deployer making the disclosure to the end user.

​
The provenance object

Provenance is an optional object that can attach to creative assets, creative manifests, individual typed assets, and content artifacts. No fields are required at the provenance level — each section is independently useful. Schema: provenance.json

​
Minimal example

Most provenance declarations answer one question: was this AI-generated, and does it need a disclosure label?
That’s it. digital_source_type says how the content was produced. disclosure.required says whether it needs a label. Everything else — tool details, C2PA references, jurisdiction-specific render guidance, verification results — is optional context that supply-chain participants can add when they need it. For content with no AI involvement, provenance is even simpler:

​
Full example

​
Digital source type

The digital_source_type enum classifies AI involvement in content production, aligned with the IPTC digitalsourcetype vocabulary. Schema: digital-source-type.json

​
Choosing the right value

For mixed-production creatives, choose the value that best describes the overall creative at the level where provenance is attached. If you need to distinguish AI involvement per-asset, attach provenance at the individual asset level instead (see Inheritance below). Common patterns:
  • AI image + human copy: Attach trained_algorithmic_media to the image asset, digital_creation to the text asset, and composite_with_trained_algorithmic_media at the manifest level
  • DCO with AI-generated headlines: data_driven_media at the manifest level, trained_algorithmic_media on the AI-generated text assets
  • Human photographer + AI background removal: composite_with_trained_algorithmic_media at the manifest level

​
Human oversight

The human_oversight enum describes the level of human involvement in an AI-assisted creation process. This field is relevant when digital_source_type indicates AI involvement. For non-AI content, omit it.
human_oversight and disclosure.required are independent fields, and the protocol does not derive one from the other. Some regulations carve out human-edited or human-directed AI output from disclosure obligations (for example, EU AI Act Article 50(4) where a human takes editorial responsibility), and the protocol leaves that determination to the declaring party’s legal analysis. Asserting human_oversight: edited or directed does not by itself justify disclosure.required: false — the carve-out has factual prerequisites the schema cannot evaluate. Sellers and governance agents may treat the combination as an audit-worthy claim and require corroborating evidence from the declaring party.
Governance agents surface this audit-worthy pattern through get_creative_features.audit_observations[] using OVERSIGHT_DISCLOSURE_CARVEOUT_CLAIMED. That observation is not a rejection code by itself; it lets sellers retain or route the claim for audit while the creative continues through normal review.

​
Inheritance

Provenance attaches at three levels in the creative hierarchy. The most specific provenance wins, and replacement is full-object — there is no field-level merging.

​
Resolution rules

  1. If an individual asset has provenance, use it
  2. Otherwise, if the manifest has provenance, use it
  3. Otherwise, if the creative asset has provenance, use it
  4. Otherwise, no provenance is declared for that asset

​
Example: mixed creative

A creative where the image is AI-generated but the copy is human-written. The manifest-level provenance covers the overall creative. The image asset overrides with its own, more specific provenance.
In this example:
  • banner_image uses its own provenance: trained_algorithmic_media with full AI tool details
  • headline inherits the manifest-level provenance: composite_with_trained_algorithmic_media
  • clickthrough_url also inherits the manifest-level provenance
Note that the image’s provenance is a complete replacement. Even though the manifest-level provenance has declared_by, the image asset must re-declare it in its own provenance object if that information should carry through.

​
Artifact inheritance

For content artifacts (publisher content), the same pattern applies:
The article text inherits digital_creation from the artifact. The illustration overrides with its own trained_algorithmic_media provenance.

​
Trust model

Provenance is a claim by the declaring party. It is not proof. The enforcing party should verify independently.
In advertising, the party declaring provenance and the party enforcing it have competing incentives. A buyer submitting a creative has reason to claim the content is human-made — AI-generated creatives may face placement restrictions, mandatory disclosure labels, or outright rejection on certain inventory. A seller accepting that creative has the opposite incentive: publishing AI-generated content without proper disclosure creates regulatory liability for the publisher, not the advertiser. AdCP handles this tension by treating provenance as a claim, not a fact. The buyer declares; the seller verifies. Verification happens at each enforcement point independently, using AI detection services (via get_creative_features), C2PA manifest validation, or both. No party needs to trust any other party’s assertion. The protocol provides the structure for claims and the integration points for verification — the supply chain provides the adversarial pressure that keeps both honest. The declared_by field identifies who attached the provenance claim. The verification array carries any detection results the declaring party wants to disclose. But the party enforcing a provenance requirement runs its own verification through existing governance infrastructure.

​
Declaring party roles

​
Buyer-attached verification

The verification array on the provenance object lets the declaring party share detection results for transparency. Multiple services can independently evaluate the same content:
These results are supplementary. A seller that requires provenance verification runs its own detection through get_creative_features rather than trusting the buyer’s attached results. Verification results use one of four outcomes:

​
Example: provenance through a campaign

Acme Brands is running a spring campaign. Their agency, Meridian Media, uses an AI image generator to produce a set of display banners — photorealistic product shots with AI-generated backgrounds. Meridian attaches provenance to the creative manifest: digital_source_type is composite_with_trained_algorithmic_media, ai_tool identifies the generator, and disclosure.required is true with eu_ai_act_article_50 and ca_sb_942 listed as applicable regulations. For the EU jurisdiction, Meridian sets render_guidance.persistence to continuous with positions preferring overlay — expressing the EU AI Act’s requirement for persistent labeling. The campaign is submitted to Pinnacle Publishing through AdCP. Pinnacle’s ad operations platform checks the provenance claim, then runs the creative through its verification pipeline via get_creative_features. The AI detection service returns ai_modified with 0.94 confidence — consistent with the declared source type. With the claim verified, Pinnacle applies its own jurisdictional policy — as the covered platform under SB 942, it makes its own disclosure determination rather than relying on Meridian’s disclosure.required: true as authoritative — reads the render guidance for the serving jurisdiction, applies the disclosure label with the specified persistence, and clears the creative for serving. The provenance metadata, the detection result, the render guidance, and the disclosure decision are all recorded and auditable. If Meridian had declared digital_capture instead — claiming no AI involvement — Pinnacle’s detection service would have flagged the inconsistency. The creative would be held for review, not served.

​
C2PA integration

The c2pa field provides a soft reference to C2PA Content Credentials — the cryptographic provenance standard developed by the Coalition for Content Provenance and Authenticity.

​
Why a URL reference

C2PA bindings are typically embedded in the media file itself. But ad tech pipelines routinely transcode, resize, and reformat creative assets — breaking file-level C2PA bindings in the process. A URL reference to the original C2PA manifest store survives this transcoding, preserving the chain of provenance through the supply chain. The reference is a pointer, not a replacement for C2PA. Any party in the chain can fetch the manifest from the URL and verify the original content credentials, even after the media file has been transcoded.

​
Usage pattern

  1. Creator generates content and produces a C2PA manifest
  2. Creator uploads the manifest store to a stable URL
  3. Creator attaches the manifest_url in AdCP provenance
  4. Downstream parties (agencies, platforms, sellers) can verify the original credentials at any time by fetching the manifest

​
Disclosure requirements

The disclosure object declares regulatory obligations for AI-generated content.

​
Render guidance

The render_guidance object on each jurisdiction expresses the declaring party’s intent for how the disclosure should be rendered based on the regulation’s requirements. Different regulations have different persistence requirements:
  • continuous — The disclosure must remain visible or audible throughout the content display duration. For video/audio, the full playback. For static formats (display, DOOH), the full display slot. For DOOH, “content duration” means the ad’s display slot within the rotation, not the screen’s full rotation cycle.
  • initial — The disclosure must appear at the start for a minimum duration before it may be removed. Pair with min_duration_ms to specify how long — without it, the duration is at the publisher’s discretion.
  • flexible — Disclosure presence is sufficient; the publisher controls timing and duration.
When multiple sources specify persistence for the same jurisdiction (e.g., brief required_disclosures[].persistence and provenance render_guidance.persistence), the most restrictive mode applies: continuous > initial > flexible. The positions array is an ordered preference list. The first position that the serving format supports should be used. For example, ["overlay", "subtitle"] means “prefer overlay, fall back to subtitle if overlay is not available.” Not all position-persistence combinations are meaningful. Positions with inherently bounded duration — end_card, pre_roll — cannot satisfy continuous persistence because they appear only for part of the content. Creative agents should not request continuous on these positions, and formats should not claim continuous support for them. For audio-only environments (podcast, streaming audio, smart speakers), only audio, pre_roll, and companion positions are applicable. Visual positions (overlay, footer, subtitle) are undefined without a screen. Creative agents building for audio formats should restrict render_guidance.positions to audio-compatible values. Render guidance travels with the creative through the supply chain. At serve time, the publisher reads the guidance from provenance and renders accordingly. Governance agents can audit whether the publisher followed the declared guidance.

​
Multi-asset aggregation

When a creative is assembled from multiple assets with different render_guidance for the same jurisdiction (common in DCO), the most restrictive persistence applies across the assembled creative: if any asset requires continuous, the assembled creative requires continuous. This follows the same precedence as conflict resolution: continuous > initial > flexible.

​
Enforcement vs self-reported compliance

For formats where the publisher controls the rendering surface (hosted video, display banners, SSAI), the publisher can enforce render guidance directly — rendering an overlay, controlling its duration, and verifying compliance. For opaque, self-rendering creatives (MRAID, JavaScript tags, VPAID), the creative controls its own viewport. The publisher cannot inject or enforce disclosure rendering inside the creative’s sandbox. In these cases, disclosure compliance depends on the creative agent embedding the disclosure during build. The format’s disclosure_capabilities should reflect this: only claim persistence modes the format’s rendering layer can verify or enforce, not modes that rely on creative self-compliance. Governance agents can verify self-rendered disclosures post-hoc via get_creative_features by rendering the creative in a headless environment and inspecting for disclosure presence.

​
Known regulation identifiers

Regulation identifiers are conventions, not a closed enum. New regulations can be referenced without protocol changes.

​
Embedded provenance and watermarks

c2pa.manifest_url is a sidecar reference: a URL pointer to a detached cryptographic manifest. It is durable across the network, but file-level C2PA bindings do not survive ad-server transcoding, resize, or re-encode. For pipelines where assets pass through intermediaries before reaching the publisher, two transcoding-resilient fields carry provenance evidence inside the asset itself.

​
embedded_provenance[]

Provenance metadata carried within the content stream — either as a manifest embedded in the file container (e.g., a JUMBF box in JPEG, a C2PATextManifestWrapper in plaintext per C2PA Section A.7) or as invisible markers in the content that encode or reference a provenance record.
verify_agent SHOULD be present for methods where the receiver cannot self-verify (e.g., provenance_markers); it MAY be omitted for self-verifiable embeddings such as a C2PA text manifest with a public key the receiver already trusts.

​
watermarks[]

Content watermarks that encode an identifier or fingerprint within the asset. Distinct from embedded provenance: a watermark carries an identifier (who generated it, who owns it), while embedded provenance carries or references a structured provenance record (the full chain of custody). A single asset may carry both.

​
The verify_agent shape

verify_agent is the buyer’s representation that this layer can be verified by a governance agent the seller already accepts. The seller is the verifier-of-record: it publishes which agents it will call on creative_policy.accepted_verifiers[] (returned by get_products), and the buyer’s verify_agent.agent_url MUST be a canonicalized match of one of those entries. Off-list URLs are rejected with PROVENANCE_VERIFIER_NOT_ACCEPTED before any outbound call. This is buyer-supplied evidence, not buyer-driven routing. The seller chooses which agent it actually invokes, and may substitute a different on-list agent than the buyer nominated; it does not call buyer-asserted endpoints outside its allowlist. Buyers MAY omit verify_agent for self-verifiable embeddings (e.g., a C2PA text manifest with a public key the seller already trusts) — in that case, the seller selects an agent from accepted_verifiers at evaluation time. Inheritance follows the same rule as the parent provenance object: most-specific wins, replacement is full-object, no field-level merging.

​
Creative policy enforcement

Sellers express provenance requirements on creative-policy — a field on every Product, surfaced through get_products so buyers see requirements before submitting.
provenance_required: true means the creative MUST carry some provenance object, anywhere in the inheritance chain. provenance_requirements refines that with field-level expectations. accepted_verifiers[] publishes the governance agents the seller operates or has allowlisted — the seller is the verifier-of-record, and buyers’ verify_agent references MUST be a canonicalized match of one of these agent_url values. Field-level requirements are seller-enforced; JSON Schema does not validate them. Sellers that publish a requirement MUST enforce it. sync_creatives rejects non-compliant submissions with one of: error.field MUST point at the resolved provenance path that was inspected. PROVENANCE_CLAIM_CONTRADICTED error.details is limited to the audit-safe allowlist { agent_url, feature_id, claimed_value, observed_value, confidence } plus substituted_for when the seller called an on-list agent other than the one the buyer nominated — sellers MUST NOT forward arbitrary verifier extension fields, detail_url, or any verifier response shape that may carry cross-tenant or PII data. These codes are correctable: a buyer’s orchestrator reads them, fixes the creative, and resubmits. Auto-retry without correction will not pass.

​
Where provenance attaches

​
Schema reference