C2PA content credentials are a digitally signed record attached to a file that states how it was created and edited and who signed that record; imperceptible watermarks embed a signal in the content itself that survives when that record is stripped. Together they are the main technical tools behind the AI Act’s machine-readable marking rules, which apply from 2 August 2026, and they give creators and brands evidence of where a piece of content came from. They prove provenance, not ownership. This guide is for brand owners, agencies, publishers and AI providers working across Europe and Latin America.
Key takeaways
- C2PA (Coalition for Content Provenance and Authenticity) publishes the open Content Credentials standard; version 2.4 dates from April 2026.
- Signed metadata is precise but easy to strip; watermarks and fingerprints help reconnect content with its record.
- The EU Code of Practice on AI-generated content asks signatories for two layers: signed, time-stamped metadata plus an imperceptible watermark.
- A content credential shows who signed and when. It does not show who is the author or owns the rights.
- Removing rights-management information can be actionable under Spanish copyright law.
What are C2PA content credentials?
C2PA is a project of the Joint Development Foundation that develops open technical standards for certifying the source and history of media content. Its specification calls the result Content Credentials. The current C2PA technical specification, version 2.4 (April 2026) works as follows:
- A tool (a camera, editing software or a generative AI system) writes statements called assertions: how the asset was created, which actions were applied, which earlier files were used as ingredients.
- Those assertions are listed in a claim, which is digitally signed with an X.509 certificate.
- A hard binding, a cryptographic hash of the content, ties the record to that exact file, so any later change becomes detectable.
- The specification recommends a trusted time-stamp from an RFC 3161 time-stamping authority, which proves the signature existed at a given date and time.
For AI content, the record can state that an asset was created with the digital source type “trainedAlgorithmicMedia”, and version 2.4 adds an AI disclosure assertion describing the model used and the level of human oversight. Since mid-2025 a C2PA conformance programme and official trust list identify conforming products and certification authorities.
The specification is explicit about its limits: it does not judge whether provenance data is “good” or “bad”, only whether the assertions can be validated as associated with the asset, correctly formed and free from tampering.
How do watermarks and fingerprints differ from metadata?
Metadata travels next to the content and disappears with a screenshot, a re-encode or a platform that strips it. C2PA therefore also defines soft bindings: a fingerprint computed from the content, or an invisible watermark embedded in it. If the manifest is removed but a copy remains in a repository, the soft binding lets the two be matched again. The specification calls this a durable content credential.
| Technique | What it shows | Main weakness | Role in the EU code |
|---|---|---|---|
| Signed metadata (for example C2PA) | Signer, creation history, edits, time | Easily stripped from the file | Required layer where the format supports metadata |
| Imperceptible watermark | That content came from a given system, or a reference to its record | Can degrade with heavy editing; detection depends on the provider | Required second layer (except very short text) |
| Fingerprint | That content matches or nearly matches a registered asset | Needs a database to compare against | Optional supplement |
| Logging | That a system generated a given output | Privacy and storage costs | Optional supplement, not sufficient alone |
Where do C2PA and watermarks fit in the AI Act?
Article 50(2) of the AI Act requires providers of generative AI systems to ensure that outputs are “marked in a machine-readable format and detectable as artificially generated or manipulated”, with solutions that are effective, interoperable, robust and reliable as far as technically feasible. Recital 133 lists watermarks, metadata identifications, cryptographic methods for proving provenance, logging and fingerprints as possible techniques.
The Code of Practice on Transparency of AI-Generated Content, published by the Commission on 10 June 2026, turns this into concrete measures for signatories:
- Digitally signed and time-stamped metadata recording whether content is AI-generated or manipulated, where the format supports it.
- An imperceptible watermark that is hard to separate from the content; for free-form text, where metadata cannot travel, the watermark alone is accepted.
- Robustness against cropping, compression, screenshots, paraphrasing and even print-and-scan.
- A detection tool free of charge, with a narrow exception for smaller providers facing substantial costs.
- Established standards for metadata from the outset and, because watermark standards are not yet mature, an interoperability solution for watermark detection by 2 February 2027.
- No intentional removal of existing metadata markings, and a ban on such removal in the provider’s terms of use.
The code does not name C2PA; it notes that, at publication, established interoperability standards existed only for digitally signed metadata. The Commission’s policy page (31 July 2026) says that the Commission and the AI Board have confirmed the code as an adequate voluntary tool and that about 190 organisations had signed it. Systems placed on the market before 2 August 2026 have until 2 December 2026 to comply, according to the Commission’s FAQ. Breaches of Article 50 can be fined up to EUR 15 million or 3% of worldwide annual turnover (Article 99(4)).
Can content credentials prove authorship?
Only in part. The signer of a C2PA claim is whoever holds the signing certificate: a camera maker, a software tool or an organisation. That is not necessarily the person who created the work. Under Spain’s Intellectual Property Act (TRLPI), the author is the natural person who creates the work, and the person who appears as author by name, signature or identifying sign is presumed to be the author unless proven otherwise (articles 5 and 6 TRLPI). Andean Decision 351 sets a similar presumption (article 8).
A credential that names the creator, for example through the identity assertion of the Creator Assertions Working Group (CAWG), can support that presumption, and the edit history can help document human contributions to AI-assisted work. On dates, the strongest tool is outside C2PA: under Regulation (EU) No 910/2014 (eIDAS), article 41, a qualified electronic time stamp is presumed accurate as to date and time and as to the integrity of the data. A C2PA time-stamp is not qualified by default, so for key assets choose a qualified provider.
Can you use content credentials to reserve rights?
CAWG’s training and data mining assertion (version 1.1, 16 May 2025) lets a rights holder state whether an asset may be used for data mining, AI inference, generative AI training or other AI training, marking each as allowed, not allowed or constrained. Whether this counts as an “appropriate” machine-readable reservation of rights under EU text and data mining rules has not been settled, so we recommend using it alongside robots.txt and your terms, not instead of them.
Removal also has legal consequences. Article 198 TRLPI lets rights holders act against anyone who knowingly removes or alters electronic rights-management information, meaning information that identifies the work, the author or the conditions of use, when they know it facilitates or conceals an infringement. A credential naming the author and licence terms can fall within that definition.
What this means for your business
- Enable content credentials in your capture and editing tools, and sign with an organisational certificate from a conforming provider.
- Add a qualified time stamp to campaign assets, designs and other works you may need to defend.
- Keep originals and manifests in a repository, so that stripped copies can be matched through soft bindings.
- Require agencies and platforms, by contract, to preserve credentials, and record your training and data mining choices.
- If you provide generative AI in the EU, plan both marking layers and the detection tool before your compliance deadline.
Our team for AI-generated content and digital asset protection can design this workflow with you, and our copyright and digital works team can align it with registration and licensing.
Where companies get content provenance wrong
- Treating a credential as a title deed. It shows who signed and when, not who owns the rights.
- Relying on metadata alone. Most platforms and screenshots strip it; without a watermark or fingerprint the record is lost.
- Signing with personal or test certificates. Validators may show the record as untrusted.
- Not preserving evidence. Without the original file, manifest and time-stamp, the record is hard to use in a dispute; our IP enforcement team can advise on what to keep.
Frequently asked questions
What are C2PA content credentials?
They are a signed, tamper-evident record attached to a digital file under the open C2PA standard. The record lists how the content was created and edited, which files were used, and who signed it, and it is bound to the content by a cryptographic hash. Any later change to the file can be detected when the credential is validated.
Is C2PA mandatory under the EU AI Act?
No. Article 50 requires machine-readable marking by providers of generative AI but does not prescribe a standard. The Commission’s Code of Practice asks signatories for signed, time-stamped metadata and an imperceptible watermark, and notes that only signed metadata had established standards at publication. C2PA is the best-known open specification for that metadata layer.
Do content credentials prove that I own the copyright?
No. They prove who signed a record and when, and that the file has not changed since. Ownership depends on authorship, employment and assignment rules, and licences. A credential naming you as creator can support the legal presumption of authorship, and a qualified time stamp strengthens the date, but neither replaces a clear chain of title.
Can IP Global Guard help us set up content provenance and enforce it?
Yes. We map which assets need credentials and time stamps, review agency and platform contracts so credentials are preserved, align provenance with copyright registration and licences, and act when marked content is copied or credentials are removed, coordinating qualified local correspondents across Europe, Latin America and Africa from a single point of contact.
How IP Global Guard supports content provenance
Content credentials and watermarks work best as part of a wider IP strategy: the right registrations, clear contracts and evidence you can produce when someone copies or manipulates your content. IP Global Guard, the IP services line of META Channel Corporation Limited, coordinates that strategy in more than 25 jurisdictions with one point of contact and one billing relationship; see our coverage across Europe, Latin America and Africa. Within the same group, our regulatory team advises AI providers on Article 50 compliance.
Tell us which content you produce, where you publish it and which tools you use. We will propose a provenance and evidence workflow that fits your markets. Contact our team to set it up.
This article is general information, not legal advice, and reflects the position on its publication date.
Sources
- C2PA, Content Credentials: C2PA Technical Specification, version 2.4 (April 2026)
- C2PA, Conformance Program and Trust List (2025)
- Creator Assertions Working Group, Training and Data Mining Assertion 1.1 (16 May 2025)
- European Commission, Code of Practice on Transparency of AI-Generated Content (June 2026)
- European Commission, publication of the Code of Practice (10 June 2026)
- European Commission, FAQ on the Code of Practice (10 June 2026)
- European Commission, Code of Practice on AI-generated content, policy page (31 July 2026)
- AI Act, Articles 50 and 99 and recital 133, Regulation (EU) 2024/1689 (2024)
- BOE, Spanish Intellectual Property Act, arts. 5, 6 and 198 (last updated 30 March 2022)
- Official Journal of the EU, Regulation (EU) No 910/2014 (eIDAS) (28 August 2014)
- Andean Community, Decision 351 (1993), via OAS SICE
