C2PA AI-Generated vs AI-Edited — Granular Provenance

Direct answer

C2PA can distinguish fully AI-generated media from AI-edited media by combining digitalSourceType, action history, ingredient relationships and Regions of Interest. Version 2.4 adds c2pa.ai-disclosure, which can describe model provenance and whether the workflow was fully autonomous, prompt-guided or human-validated. These are structured provenance statements, not automatic judgments about materiality, truth or disclosure law.

Applies when

Use this evidence when you need to verify which AI-generation or AI-editing distinctions the signer recorded at asset or action level.
Use this evidence when you need to verify which regions, source ingredients and oversight state are represented in the credential.

Doesn't apply when

Do not treat this evidence as proof that the recorded AI change is legally material or immaterial.
Do not treat this evidence as proof that every unrecorded part of the history was human-created.

Worked decision

Example

A real packshot receives an AI-generated beach background and is approved by a human editor.

Interpretation

The credential can identify the photographed ingredient, AI editing action, affected background region and human_validated state without calling the product synthetic. Use component-level disclosure: “Background generated with AI. Product photographed.” and keep the verification result.

Next steps

Required action

Classify the source of the asset and each material action using the appropriate IPTC digital source type.
Relate prompts, references and source media as ingredients; use Regions of Interest when AI affected only part of the asset.
Add supported AI disclosure fields, verify the manifest and translate technical assertions into accurate audience-facing wording.

Publishing layer

Platform differences

Original asset

A C2PA-aware verifier can inspect the signed assertions associated with the exact file it receives.

Transformed derivative

Screenshots, resizing, transcoding, metadata stripping or unsupported export can remove readable provenance or invalidate its binding.

Publishing platform

Native labels and legal disclosure remain separate layers; their thresholds can be broader than the technical provenance signal.

Evidence

Source

C2PA Technical Specification (2.4 series)Coalition for Content Provenance and Authenticity · last checked 2026-09-12
A New Implementation Guide for Content CredentialsC2PA · last checked 2026-09-12
Use of Content Credentials to Identify Synthetic and Non-Synthetic ContentC2PA Technical Working Group · last checked 2026-09-12

Freshness and accountability

Last verified

Last reviewed 2026-09-12 · Current

Run this exact scenario

A representative asset scenario is prefilled. Confirm how provenance was created, whether it survived the final export and where the asset will be published.

Open prefilled checker

How to read the signal

Key facts

Fully generated	`trainedAlgorithmicMedia` with a creation action can identify an asset generated as new model output.
AI-edited composite	`compositeWithTrainedAlgorithmicMedia` and action-level source types can identify AI applied within a broader human or captured-media workflow.
Regions of Interest	Assertion metadata can localise affected image areas, text ranges, audio intervals or video time segments instead of labelling only the whole file.
Human oversight	`c2pa.ai-disclosure` defines `fully_autonomous`, `prompt_guided` and `human_validated` as distinct declared workflow states.

Quality gate

Common failure modes

Marking the entire product image AI-generated when only the background was generated.
Using human_validated to imply that a lawyer, regulator or platform approved the content.
Assuming precise region metadata means the modification was immaterial or legally exempt.

Disclara treats technical provenance as evidence about process, not as proof of truth, authorship or legal compliance. This page is source-backed information, not legal advice. Review the sources and methodology.