Content provenance and authenticity (Beta)

Introduction

The Coalition for Content Provenance and Authenticity (C2PA) is a collaboration among tech and media companies to combat online misinformation. It aims to establish standards and tools for verifying the authenticity of digital media, such as images and videos, by capturing and preserving information about their creation and editing. The goal is to prevent the spread of deceptive or manipulated content on the internet.

C2PA defines a specification to accomplish this goal. The tooling used to implement C2PA is provided by the Content Authenticity Initiative (CAI), which consists of various SDKs, a command-line tool, and an underlying Rust library. See the GitHub repo.

C2PA uses cryptography to bind provenance metadata, called assertions, directly to the asset binary. Together, those assertions record who created or edited an asset, which tools were used, and whether AI was involved. The resulting manifest is tamper-evident: anyone can verify the signature, and any change to the asset breaks that signature unless the asset is signed again. A manifest can still be stripped from an asset entirely, so C2PA is best understood as a transparency mechanism for cooperative participants rather than a way to prevent misuse.

Cloudinary preserves, extends, and verifies C2PA manifests across the whole media lifecycle: on upload, through transformation and optimization, and at delivery.

You can use this online validation tool to see the history of certifications that have been applied to an image.

Cloudinary and the C2PA standard

Cloudinary has been an active participant in content authenticity standards work for several years:

• Cloudinary joined the Content Authenticity Initiative (CAI) in December 2020 and supports the adoption of open standards for media provenance and authenticity.
• Since 2023, Cloudinary has been included on the C2PA Interim Trust List (ITL) as an early supporting service, and is progressing through the official C2PA Conformance Program.
• Cloudinary is a member of ISO/JPEG and participates in the JPEG Trust work on ISO/IEC 21617, which sets standards for trustworthy media.
• Cloudinary holds U.S. Patent US 11,550,879 B2, "Providing Authenticity Mechanism for Media Content".

Scope

C2PA in Cloudinary is currently available only to customers who request it. To use the capabilities described on this page, contact our support team to have them enabled for your product environment.

It's implemented for images only (specifically, these output formats: avif, heic, heif, jpg, jpeg, png, svg, tif, tiff, and webp).

Cloudinary adds its signature in a new manifest on top of any existing manifests. If a previous signature is invalid, Cloudinary also marks the previous manifest as invalid.

Alterations made by Cloudinary are classified as transcoded or edited. The transcoded actions are defined by a closed allowlist and include: c_fit, c_mfit, c_pad, c_lpad, c_mpad, f_*, q_*, and c_scale with a single dimension, w or h.

All other actions are classified as edited.

The metadata attached to each asset contains the certificate, a private key, and a timestamp.

How Cloudinary adds C2PA manifests

Cloudinary adds C2PA manifests in two automatic cases and one on-demand case.

Preserving an existing manifest

When an asset that already carries C2PA credentials is uploaded, Cloudinary preserves the existing manifest and carries it forward at delivery, so a chain that began in a camera or a creative tool isn't broken by transformation and optimization.

When Cloudinary delivers a derived asset, it doesn't replace or start over from the original manifest. It adds a new manifest, signed by Cloudinary, on top of the existing ones. That new manifest records Cloudinary's own actions, such as transcoding or editing, and the original manifest stays in the chain.

Starting a chain for generative AI transformations

Any transformation that uses generative AI produces an asset that has been materially altered by AI. For those transformations, Cloudinary dynamically injects a manifest recording that AI altered the asset. This means a derived asset carries a machine-readable record of the AI involvement without you needing to add anything to the URL.

The following transformations start a provenance chain:

• Background removal (e_background_removal)
• Content extraction (e_extract)
• Drop shadow (e_dropshadow)
• Generative background replace (e_gen_background_replace)
• Generative fill (b_gen_fill with c_pad)
• Generative recolor (e_gen_recolor)
• Generative remove (e_gen_remove)
• Generative replace (e_gen_replace)

Requesting a manifest on demand

When delivering images that you want to be signed by Cloudinary, include the c2pa flag (fl_c2pa in URLs). For example:

The c2pa flag creates a manifest for assets that don't already have one. Assets that already carry a manifest get an updated manifest on transformed derived assets without the flag, for example when you resize an asset, even though no AI was involved. The examples below use fl_c2pa.

For images that are transcoded but not edited, for example c_scale,w_550/fl_c2pa, a validation tool would show something similar to the following (view in validation tool):

For images that are edited, whether or not they're also transcoded, for example e_cartoonify/fl_c2pa, a validation tool would show something similar to the following (view in validation tool):

Provenance and AI fields

Alongside the manifests themselves, Cloudinary tracks two asset properties: whether AI was involved, and whether the asset carries a manifest at all. A manifest on its own doesn't mean AI was used, so the two are independent. They also work differently from each other:

Both fields are available as supported expression fields for the Search API and Media Library search.

Field names

The name you use depends on where you're using the field:

The conditional transformation variable carries a cld_ prefix because it's read from contextual metadata, where it could otherwise clash with a contextual metadata key of the same name that you've defined yourself. The prefix keeps the two apart, so the result is never ambiguous.

How the AI-generated field gets set

The AI-generated field can be set in several ways:

• By you, on upload: pass the ai_generated upload parameter in your upload call, directly or via an upload preset.
• By Cloudinary, from an existing manifest: when an uploaded asset contains a C2PA manifest indicating AI generation, Cloudinary sets the field.
• By Cloudinary, on AI generation: when you generate an image with Cloudinary's generative AI transformations, Cloudinary sets the field on the result.

You can't update this field for existing assets.

Setting ai_generated on upload is useful for bulk imports from a generation pipeline, where you know the assets are AI-generated but the source files may carry no manifest. For example:

Finding assets by provenance

You can search for assets by provenance using the ai_generated and has_c2pa search expression parameters. Use the expression with the Search API, or enter it in the Media Library search box.

Find AI-generated assets

Use the following expression to find assets identified as AI-generated:

If your Media Library Preferences administrator sets AI-generated as the primary attribute, AI-generated assets are labeled as such in all Media Library views. This option requires an Assets Enterprise plan.

Find assets without a C2PA manifest

Use the has_c2pa field to find assets based on whether they have a C2PA manifest. For example, use the following expression to find assets without a C2PA manifest:

For more information, see Query Builder and Advanced Search.

Include provenance fields in Search API responses

To include the ai_generated and has_c2pa fields in a Search API response, add provenance to the with_field parameter. See the sample response.

Disclosing AI-generated content

Regulations such as the EU AI Act require transparency when AI-generated media is published. Article 50 of the EU AI Act sets out two distinct obligations, and meeting them calls for two different mechanisms:

Adding a visible AI disclosure

To disclose AI generation to people viewing your assets, use a conditional transformation that overlays a logo only when the asset's AI-generated field is true. Assets that aren't AI-generated are delivered unchanged, so you can apply the same transformation across your whole library.

The following transformation overlays the image with public ID docs/icons/LABEL_AI_black_transparent, sized to 10% of the asset width, in the bottom-right corner, but only when cld_ai_generated is true:

For example, the following image was uploaded with the ai_generated parameter set to true, so the logo appears in the bottom-right corner:

Because the overlay is applied at delivery, you control its size, placement, and appearance per delivery URL, and you can combine it with any other transformation. For example, cropping to a 9:16 aspect ratio and then adding the disclosure:

Verifying a manifest

Anyone can verify the signature and provenance of a delivered asset using standard C2PA verification tools:

• The Content Credentials verification tool shows the chain of manifests in a browser. Pass the delivery URL as the source query parameter, for example this signed image.
• The c2patool command-line tool inspects an image file's manifests in detail.

A verification tool shows each step in the chain in order. For an asset created by a generative AI tool and then transformed by Cloudinary, the chain shows the original AI generation, any edits made in other tools, and the actions Cloudinary performed.