SynthID vs C2PA Content Credentials: How Watermarks and Signed Metadata Differ

SynthID and C2PA Content Credentials both help people tell where a piece of media came from, but they work in opposite ways, answer different questions and fail in different conditions, so treating them as interchangeable leads to wrong conclusions about a file. SynthID is Google DeepMind's invisible watermarking technology for AI-generated images, video, audio and text; C2PA Content Credentials are an open technical standard from the Coalition for Content Provenance and Authenticity for attaching signed, tamper-evident records of how any media was made and edited. This page compares the two on identical fields, explains how each works, where each is supported, what each can and cannot prove, and how they combine. It is written for trust and safety teams, newsrooms, compliance staff preparing for AI labelling rules, and engineers building media pipelines. In tables and lists the two are given in alphabetical order (C2PA first); order is not a ranking.

Last checked: 2026-11-15.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.

Key takeaways

• Signed record versus hidden signal. C2PA attaches a cryptographically signed manifest describing origin and edits; SynthID hides an imperceptible signal inside the content itself.
• Different questions. C2PA answers "who signed this record of how the file was made, and has the file changed since?"; SynthID answers "was this generated or altered with a tool that applies SynthID?"
• Different weak points. C2PA metadata can be stripped from a file; SynthID detection is probabilistic and its text detector loses confidence when text is heavily rewritten or translated.
• Designed to combine. The C2PA specification defines watermarks as "soft bindings" that can point back to a stored manifest, and Google describes the two as separate layers it uses together.
• Both are spreading in 2026. Google added C2PA verification to the Gemini app in May 2026 and SynthID verification to Search on 20 May 2026, with OpenAI, Kakao and ElevenLabs named as bringing SynthID to their content.

What are SynthID and C2PA?

C2PA Content Credentials: the C2PA describes itself as providing an open technical standard for establishing where digital content came from and how it was edited, and uses the public name Content Credentials, likened to a nutrition label for digital content. The coalition is organised as a Joint Development Foundation Projects series under the Linux Foundation, and its steering committee lists Adobe, Amazon, BBC, Google, Meta, Microsoft, OpenAI, Publicis Groupe, Sony Group, TikTok and Truepic. The C2PA site links specification version 2.3, dated December 2025, as the specification to adopt.

SynthID: Google DeepMind describes SynthID as watermarking for AI-generated content that embeds digital watermarks imperceptible to humans but detectable by its technology. It covers images, video, audio and text. Google says it has applied SynthID to over 100 billion images and videos and 60,000 years of audio (May 2026), up from the 10 billion pieces of content it reported in May 2025.

SynthID vs C2PA at a glance

How do C2PA Content Credentials work?

Under the C2PA 2.3 specification, a C2PA Manifest is built from one or more assertions (statements about the asset, such as the capture device, the edit actions or the use of AI), a single claim that references those assertions and the data needed to bind them to the asset, and a claim signature made with the signer's private key. Claims are CBOR-encoded with the label c2pa.claim.v2, and signatures use COSE (RFC 8152).

• Hard binding: one or more cryptographic hashes that uniquely identify the asset or part of it, proving that the manifest belongs to that asset and that it has not changed.
• Soft binding: a content identifier that is either a fingerprint (not statistically unique) or a watermark. The specification says a soft binding shall not be used as a hard binding.
• Certificates: only X.509 certificates may be used for signing since version 2.0. The C2PA Trust List holds trust anchors for hardware and software signers and, since version 2.2, is limited to certificates with the c2pa-kp-claimSigning extended key usage. A separate list covers time stamping authorities.
• Trust model: the specification states that trust decisions rest on the identity of the signer associated with the signing key. That makes key protection an identity problem, covered in depth in the AI Content Provenance and C2PA Guide.
• AI disclosure: the C2PA explainer notes that AI-performed actions are identified through the digitalSourceType field.
• Redaction: an assertion may be removed when an asset becomes an ingredient of another, and the removal is recorded; c2pa.actions assertions and hard bindings may not be redacted.

A C2PA-aware application checks the manifest's integrity and signatures and whether the asset changed since signing, then shows an indicator such as a badge, letting the viewer see who signed and decide how far to trust it. The explainer is explicit that Content Credentials do not establish whether content is true, and that edits made in tools that are not Content Credential-aware may go unrecorded.

How does SynthID work?

SynthID adds its signal at generation time, so it is present from the first moment a file exists. Google DeepMind describes the behaviour per modality:

SynthID Text in detail: Google has open sourced SynthID Text, with a production implementation in Hugging Face Transformers from version 4.46.0 and a reference implementation on GitHub. It runs as a logits processor after top-k and top-p sampling and uses a pseudorandom g-function keyed by a private list of integers; ngram_len trades robustness against detectability (5 is a common default). Detection uses a Bayesian detector that returns watermarked, not watermarked or uncertain, with thresholds tuned for target false positive and false negative rates. Google says the watermark configuration must be kept private, that SynthID Text is less effective on factual responses, and that it is not designed to stop determined adversaries.

Detection surfaces: the SynthID Detector portal was announced on 20 May 2025 for early testers and, per Google DeepMind's SynthID page, is now open to everyone for image, video and audio files made with Google AI or partner tools. Users can also ask "Is this made with AI?" through Lens, AI Mode, Circle to Search or Gemini in Chrome. Google reported 50 million uses of SynthID verification in the Gemini app at I/O 2026.

What can each prove, and what can it not?

How do watermarks and signed metadata work together?

The C2PA specification defines an invisible watermark as information embedded in a way that is substantially imperceptible to humans, which can identify the asset or store a reference to a C2PA Manifest. Version 2.3 also introduced fine-grained watermarking actions. A Durable Content Credential is a Content Credential with one or more soft bindings that allow it to be found in a manifest repository: if a platform strips the embedded manifest, a watermark or fingerprint can still lead a verifier back to the stored record. In this model the watermark carries the link and the manifest carries the signed detail.

Google presents SynthID and C2PA as two layers: SynthID flags AI-generated content, while C2PA records the creation and editing history of any media, whether or not AI was involved. Its September 2025 Pixel post says Content Credentials "can be complemented by watermarking technologies such as SynthID". Google's May 2026 post does not describe a technical integration in which SynthID acts as a C2PA soft binding, so that should not be assumed.

Where are SynthID and C2PA supported?

The table lists support stated on the official pages checked. It is not a complete industry list; the AI-Generated Media Labelling Tracker follows product-by-product status.

The Pixel 10 implementation shows what strong C2PA signing looks like in practice: claim signing keys are generated and held in Android StrongBox on the Titan M2 chip, Android Key Attestation lets Google's C2PA certification authorities confirm a genuine device and app before issuing certificates, and a trusted on-device clock timestamps captures offline so credentials stay verifiable after certificate expiry.

How to choose between SynthID and C2PA

For most organisations this is not a choice between rival products, because one is a standard any vendor can implement and the other is a watermark applied by the tools that adopt it. The factual criteria below decide which to rely on for a given job.

• Proving your own content is authentic: C2PA, because you sign it with your own certificate and any C2PA-aware viewer can validate it. Watermark detection does not establish who published a file.
• Checking whether incoming media came from a SynthID-enabled generator: SynthID verification, because it can survive the re-encoding that usually removes metadata.
• Meeting marking duties for AI output you generate: Article 50(2) of the EU AI Act requires providers of systems generating synthetic content to mark outputs in a machine-readable format, applying from 2 August 2026 (Article 50 text); the EU Code of Practice on Marking and Labelling AI-Generated Content Guide covers how multi-layer marking is described there, and the EU AI Act Article 50 Transparency Obligations Guide covers the obligation itself.
• Running your own text model: SynthID Text is an open option you can deploy yourself, with a private key configuration.
• Keeping provenance through social platforms: combine an embedded manifest with a soft binding and a manifest repository (Durable Content Credentials).
• Small teams: verifying is cheaper than signing; start by checking inbound media with C2PA-aware tools and Google's SynthID surfaces before investing in signing infrastructure.

Where identity fits

C2PA turns provenance into a certificate problem: whoever controls a claim signing key can make content appear to come from your organisation, so signing keys are non-human credentials that need hardware protection, rotation and revocation. SynthID moves the secret to the watermark configuration, which Google says must stay private because exposure could let others replicate the watermark. For key handling around model and content signing see the AI Model Provenance and Watermarking Guide. Neither technology verifies a live person on a call; for impersonation defence see the Deepfake and AI Impersonation Guide.

Common failure patterns

• Reading absence as authenticity: no manifest and no watermark proves nothing, because both are opt-in and most tools apply neither.
• Treating a SynthID result as universal AI detection: it only recognises content from tools that apply SynthID.
• Stripping your own credentials: image pipelines, CDNs and CMS resizing that drop metadata remove the manifest before publication.
• Signing with an unprotected key: a software key on a shared server lets anyone with access sign content as the organisation.
• Exposing the text watermark configuration: publishing SynthID Text keys lets others forge or test against the watermark.
• Over-trusting text detection: heavily rewritten, translated or factual text gives low-confidence results.

Detection and measurement

• Share of published images and videos that still carry a valid manifest after passing through your delivery pipeline.
• Share of inbound media reviewed by trust and safety that was checked for both C2PA and SynthID.
• Number of signing certificates, their storage (hardware or software) and days to revoke after a suspected compromise.
• For SynthID Text deployments, the uncertain-result rate and measured false positive rate against your thresholds.
• Number of AI-generated assets published without either marking, against your labelling policy.

Practitioner checklist

• Decide, per content type, whether the goal is proving your own authenticity (C2PA signing) or identifying AI output (watermark checks), or both.
• Test that your CMS, CDN and social publishing tools keep C2PA manifests; fix any step that strips them.
• Hold claim signing keys in hardware and register them in your non-human identity inventory with an owner and rotation date.
• Add C2PA validation and SynthID checks to the triage steps for suspicious inbound media, and record both results.
• Never treat a missing credential or watermark as evidence that content is genuine.
• If you deploy SynthID Text, keep the keys private, tune detector thresholds and document the uncertain state.
• Map your marking approach to EU AI Act Article 50 and any other labelling rule that applies to you.
• Re-check vendor support each quarter; Google stated Chrome and Search C2PA verification as coming in 2026.

Standards and references

• EU AI Act, Regulation (EU) 2024/1689, Article 50: Transparency obligations (text)
• C2PA: Technical Specification 2.3
• C2PA: Explainer 2.3
• C2PA: Coalition for Content Provenance and Authenticity
• Google DeepMind: SynthID
• Google AI for Developers: SynthID Text
• Google: Making it easier to understand how content was created and edited
• Google: I/O 2026 announcements
• Google: SynthID Detector
• Google Security Blog: Pixel and Android trusted images with C2PA Content Credentials

Related NHI Mgmt Group resources: AI Content Provenance and C2PA Guide · AI Model Provenance and Watermarking Guide · AI-Generated Media Labelling Tracker · AI Image, Video, Voice and Music Models Hub