The Signature After the Image

Content Credentials, provenance standards and invisible watermarks promise a chain of trust. Yet no technical label can establish meaning, ownership or truth on its own.

A signature on a painting occupies the same physical surface as the work. It may be forged, added later or misread, but it remains visible as part of the object. Digital art has never possessed an equivalent that is both durable and widely legible. A file can be duplicated without apparent loss. Metadata can disappear during export. A screenshot can detach an image from its history. A platform can compress a work, strip information and circulate the result at a scale its maker never intended.

Generative artificial intelligence has made this old weakness newly urgent. When a technically convincing image can be produced without a camera, when a voice can be replicated without the speaker and when an artwork can be altered in ways that are difficult to detect by sight, provenance becomes more than a concern for collectors. It becomes a public question about how digital culture establishes trust.

The emerging answer is not a universal detector that separates truth from falsehood. Detection remains an adversarial and imperfect field. The more durable approach is to record what happened to an asset as it moves through a creative and publishing process. The Coalition for Content Provenance and Authenticity, known as C2PA, has developed an open technical specification for this purpose. Its Content Credentials can attach cryptographically signed claims about creation, editing and publication to digital media.

The promise is significant. The limitation is equally important.

A provenance record can tell an audience that a particular party made a signed assertion, that certain tools were used and that a file changed after a recorded step. It cannot decide whether the artwork is good, whether the artist acted ethically, whether every source was licensed or whether the scene represented by an image ever occurred. Provenance is evidence about a chain of custody. It is not a machine for producing truth.

The central answer

Digital art can build a credible account of origin when technical provenance, human identity, legal rights, process documentation and institutional preservation reinforce one another. Content Credentials can make parts of that account portable and tamper-evident. They cannot independently prove authorship, copyright ownership, consent or artistic authenticity. The strongest record therefore combines machine-readable credentials with source files, contracts, edition information, an artist statement and a preservation dossier that explains which elements of the work are essential.

Why a file needs a biography

A born-digital artwork may pass through an unusually complex life before anyone encounters it. It can begin as code, a scan, a photograph, a model, an animation, a game-engine scene or a data set. It may be altered by several programs, collaborators and automated systems. The final public version might be projected, streamed, printed, minted, exhibited through a browser or installed as an interactive environment.

Each of those forms creates different evidence.

The source code may reveal how behavior is generated. A project file may preserve layers and timing. A cryptographic signature may identify the party who issued a claim. A contract may establish who owns which rights. An edition certificate may define the number of authorized copies. A museum record may document the configuration that was approved for acquisition. A video of an installation may show how the work behaved when its original hardware was still available.

Reducing this history to a single label such as "digital artwork" conceals the decisions that make the work intelligible. A file needs a biography because the file alone may be only one temporary manifestation of the artwork.

This is especially important when the work changes over time. A generative piece can create new outputs at each execution. A networked work may depend on live data. An interactive installation may adapt to visitors. A machine-learning artwork may rely on a model, a data set and a software environment that will not remain stable indefinitely. In such cases, provenance must describe a system rather than merely identify an image.

What Content Credentials actually record

The C2PA specification is designed to preserve a sequence of signed manifests. When an asset changes, a new manifest can describe an action while referring back to earlier ingredients and claims. A compatible verifier can inspect whether signatures are valid, which assertions were attached and whether the asset still corresponds to the signed record.

Version 2.4 of the specification, published in April 2026, added new asset-format support, a JSON-LD representation called crJSON for evaluation and validation reporting, a repository-receipt assertion and a field for machine-generated environmental sustainability metrics. The development illustrates how provenance is expanding beyond a simple camera-to-publisher pathway. It is becoming an infrastructure capable of describing more varied creative systems.

The C2PA conformance program adds another layer. Products that generate or validate credentials can be assessed against technical and security requirements, and conforming products appear on public lists. This matters because a standard is only as dependable as its implementations. A theoretically secure specification offers little protection when software applies it incorrectly.

For audiences, the visible entry point is usually the Content Credentials icon, a small "CR" mark that can reveal a digital label. Depending on the implementation, the label may show the creator or publisher, the tools involved, editing actions and whether generative AI was used. Identity information remains optional, which is essential for privacy and for artists working under pseudonyms or in unsafe political conditions.

The system is opt-in. That is a strength because it avoids turning identity disclosure into a universal requirement. It is also a limitation because the absence of a credential tells the viewer almost nothing. A file without Content Credentials may be an unaltered photograph, an older artwork, an image exported through unsupported software or a synthetic asset produced without provenance. Missing evidence is not evidence of fabrication.

The difference between provenance and detection

Two technical strategies are often confused.

Detection analyzes an asset and estimates whether it contains signs associated with manipulation or machine generation. Provenance carries information forward from participating tools and parties. Detection asks, "What does this file appear to be?" Provenance asks, "Which claims accompanied this file, and can their integrity be verified?"

The National Institute of Standards and Technology has examined watermarking, provenance, labeling and detection as complementary approaches to synthetic content. Its ongoing GenAI evaluations also demonstrate why a single detector should not become the foundation of public trust. Generation and detection evolve against each other. A result that performs well against one group of systems or one data set can fail when methods change.

For art, the distinction is decisive. An abstract image may resemble synthetic output even when it was painted, scanned and composited by a person. A generative work may openly declare its process and be more transparent than a heavily retouched photograph with no record. Trust should not depend on whether a classifier finds the surface statistically suspicious.

A provenance system begins earlier. It can record capture, creation, editing and publication. This does not eliminate deception, but it gives a responsible creator a way to carry evidence with the work rather than asking a later detector to reconstruct history from pixels alone.

Durable credentials and the problem of separation

Digital platforms frequently remove or rewrite metadata. C2PA therefore distinguishes between hard binding, where credentials are embedded in or directly attached to an asset, and forms of soft binding that can help recover a record after separation.

Adobe's Content Authenticity application illustrates one implementation. It combines embedded Content Credentials with an invisible watermark and a digital fingerprint linked to a public credentials cloud. Adobe states that this combination can allow credentials to remain recoverable even when metadata is removed or a screenshot is made.

This durability is useful, but it should be described carefully. Invisible watermarking is not magical permanence. It depends on technical conditions, transformations and access to an appropriate recovery system. Fingerprinting involves matching an asset or derivative to a stored representation. A creator and an institution should test the actual workflow through which a work will be exported, uploaded, compressed, displayed and archived.

The strongest approach is redundancy. The work should not depend on one vendor, one metadata block or one cloud repository. A museum or collector should preserve the original credentialed asset, validation results, source files and human-readable records outside the distribution platform.

A signed claim is not the same as a verified right

Cryptography can verify that a credential was signed with a particular key and that its contents were not altered after signing. It cannot establish that the signer possessed every legal authority implied by the claim.

A person can sign a file that includes material belonging to someone else. A verified name does not automatically prove copyright ownership. A studio may own a production file while a performer retains rights in likeness or voice. An artist may have permission to exhibit a data set but not to train a commercial model on it. A publisher may accurately describe an edit while remaining unaware of an earlier unauthorized use.

This is why provenance must be connected to legal records.

For a digital artwork, the relevant documents may include:

• contributor agreements;
• model and performer releases;
• licenses for images, sound, software and data;
• terms governing machine-learning models;
• commissioning contracts;
• edition and resale conditions;
• institutional acquisition agreements;
• permissions from communities whose cultural materials are represented.

Content Credentials can carry or reference useful assertions, but the C2PA clarified in January 2026 that its core specification does not contain a standard assertion for text and data mining permission and does not provide digital-rights-management enforcement. The standard is extensible, yet extensibility should not be confused with a built-in ability to stop training, copying or redistribution.

Artists should therefore avoid language such as "Content Credentials protect my work from AI." They may communicate a preference, preserve attribution and strengthen evidence. They do not create an enforceable technological barrier around every use.

The European transparency threshold

Article 50 of the European Union AI Act began to apply on 2 August 2026. The rules require providers of certain AI systems to support machine-readable marking of AI-generated or manipulated content. Deployers also face disclosure obligations in defined contexts, including deepfakes and certain publications on matters of public interest without human review or editorial control. A limited transition until 2 December 2026 applies to certain systems placed on the market before the main date.

For artists and institutions, the cultural significance extends beyond compliance. Machine-readable marking is moving from a voluntary experiment toward regulated information infrastructure. Exhibition labels, collection records and publishing systems will increasingly need to distinguish among several questions:

• Was AI used at any stage?
• Was the final visible or audible content substantially generated or manipulated?
• Was the work subject to human editorial control?
• Is a realistic person, event or statement represented synthetically?
• Which public disclosure is necessary in this jurisdiction and context?

A single universal phrase will not answer all five.

Artistic and satirical contexts receive particular treatment under the European framework, but artistic status is not a license to obscure material facts. The most credible institutions will build a disclosure vocabulary more precise than the minimum legal label.

The five layers of a credible digital-art record

A serious provenance system for art can be organized in five layers.

1. Technical origin

This layer records the asset, source files, software, model versions, significant actions, cryptographic credentials and checksums. It answers how the digital object came into its present state.

2. Human authorship and responsibility

This layer identifies the artist, studio and collaborators, explains their roles and records the decisions that establish the work's final form. It distinguishes direct execution from delegation, automation and assistance.

3. Rights and consent

This layer holds contracts, licenses, releases, data permissions and restrictions. It answers not merely who touched the file, but who authorized the relevant uses.

4. Edition, ownership and transaction

This layer describes authorized editions, certificates, acquisition history, transfers, display rights and any relationship between the artwork and a token or registry. A token may record a transaction, but it should not be treated as the artwork's full provenance unless the legal and technical relationship is explicit.

5. Preservation and interpretation

This layer explains how the work should be installed, what may change, which dependencies can be substituted, which behaviors are essential and how audiences should understand the process. It turns provenance from a retrospective certificate into a future-oriented conservation tool.

These layers can reinforce one another without being collapsed into a single database. The purpose is not administrative excess. It is to prevent the visible asset from becoming detached from the human, legal and technical conditions that give it meaning.

What provenance cannot solve

Trust infrastructure often attracts unrealistic expectations.

It cannot prevent every forgery. A copied image can still circulate without its record. A malicious party can make a new asset and sign false claims under an identity it controls. A legitimate work can be misrepresented through a deceptive caption. A technically valid photograph can document a staged event. A credential can preserve a history of edits while saying nothing about whether those edits were fair.

Nor can provenance replace criticism. An artwork may have perfect documentation and remain conceptually empty. Another may emerge from an informal, collective or endangered practice in which rigid identity systems would create risk. Cultural authority cannot always be reduced to an individual verified name.

The goal is not to make every work fully knowable. It is to make consequential claims inspectable.

A viewer should be able to tell whether a statement comes from the artist, a platform, a camera, an institution or an automated process. A conservator should be able to distinguish an original dependency from a later replacement. A collector should know whether ownership of a file includes display rights. A community should be able to challenge the use of its cultural material even when the file's cryptographic chain is intact.

A new form of connoisseurship

Traditional connoisseurship trained the eye to recognize a hand, a material and a historical context. Digital connoisseurship requires an expanded literacy.

It must look at the work and around the work. It asks how code behaves, how a model was trained, which data enters the system, what the interface conceals, how the file travels and what will remain when the current platform disappears.

This does not diminish visual experience. It protects it from false simplicity.

The future signature may not be a name placed in a corner. It may be a distributed record connecting the artist's identity, decisions, rights, software and preservation instructions. Its elegance will depend on how little of that complexity must interrupt the encounter while remaining available when trust requires examination.

The best provenance system will not tell viewers what to think. It will allow them to know what kind of claim they are being asked to believe.

Six questions readers may ask

1. Do Content Credentials prove that a digital artwork is genuine?

They provide evidence about provenance, not a final judgment of artistic authenticity. A valid credential can show that a signed claim accompanied an asset, identify participating tools or parties and reveal whether the asset changed after a recorded step. It cannot independently prove that the signer owns all relevant rights, that every collaborator consented or that the represented event is true. Museums and collectors should treat Content Credentials as one layer in a broader record that includes source files, contracts, artist testimony, edition information and acquisition documentation.

2. Does the absence of Content Credentials mean that an image is fake or AI-generated?

No. The system is opt-in, and many legitimate creative tools, platforms and archives do not preserve or display credentials. Older works may predate the standard. Metadata may have been removed during an unsupported export. A genuine work may circulate as a screenshot or derivative. The absence of a record means only that a compatible record was not found. It should not be converted into an accusation. Provenance is most useful as positive evidence when present, not as a universal test of suspicion when absent.

3. Can an artist use Content Credentials to stop AI companies from training on a work?

Not by themselves. The C2PA has stated that the core specification does not contain a standard text-and-data-mining assertion or digital-rights-management mechanism. Some applications allow creators to express preferences, and third parties can build additional assertions, but a preference is not the same as technical prevention or legal enforcement. Artists still need clear licensing terms, platform choices, contracts and applicable legal rights. Credentials can strengthen attribution and evidence, but they do not function as an impenetrable access control.

4. What should a collector receive when acquiring born-digital art?

The exact package depends on the work, but it should usually include the authorized asset or access method, source or preservation files where appropriate, checksums, edition and ownership documentation, rights granted to the collector, installation instructions, hardware and software dependencies, a record of acceptable substitutions and a process statement from the artist. If Content Credentials are used, the collector should receive the credentialed original and a validation record. Interactive or generative works also need guidance on data, networks, updates and future behavior.

5. Can blockchain serve as a complete provenance system for digital art?

A blockchain can record transactions or registry events, but it does not automatically establish the truth of the information entered or preserve the artwork itself. The link between a token and a file may break, and ownership of a token does not necessarily transfer copyright or exhibition rights. Blockchain can be one component of an edition or transaction history. It should be connected to the asset, legal agreement, creator identity and preservation record rather than presented as a substitute for them.

6. How much process information should an artist disclose publicly?

Enough to make material facts understandable without exposing private data, security-sensitive details or every experimental step. A public record may identify the principal tools, the role of AI, the nature of human intervention, major collaborators and relevant rights. A fuller technical dossier can remain with the artist, gallery, collector or institution. Disclosure should be proportionate to the work. The aim is accountable interpretation, not compulsory surveillance of the creative process.

1. C2PA, Content Credentials Technical Specification 2.4, April 2026.
2. C2PA, Conformance Program and Trust Lists, accessed August 2026.
3. C2PA, clarification on text and data mining assertions and DRM, January 2026.
4. C2PA, Content Credentials 2.3 and five years of adoption, February 2026.
5. C2PA, TikTok joins the Steering Committee, July 2026.
6. C2PA, official Content Credentials icon and user experience, 2023.
7. NIST, Reducing Risks Posed by Synthetic Content, November 2024.
8. NIST, GenAI Evaluation Program, accessed August 2026.
9. European Commission, Guidelines on Transparency of AI-Generated Content, July 2026.
10. European Commission, questions and answers on Article 50, July 2026.
11. Adobe, Content Authenticity public beta overview, updated September 2025.
12. Adobe, impact of applying Content Credentials, September 2025.
13. Adobe, Content Authenticity public beta and creator attribution, April 2025.
14. Content Authenticity Initiative, open tools and membership, accessed August 2026.