Images generated with ChatGPT, Gemini, or Adobe Firefly can carry a C2PA manifest known as Content Credentials. The open standard is backed by a coalition that includes OpenAI, Google, Microsoft, Adobe, and major camera manufacturers. Its purpose is to document where media came from.
There is a practical paradox: the manifest is easy to remove, and platforms often remove it for you.
How C2PA works
A C2PA manifest is a cryptographically signed metadata container:
- JPEG: an APP1 segment identified by
jumb, containing a superbox with the CBOR manifest and signature. - PNG: a
jumbchunk. - WebP: a RIFF
JUMBchunk. - MP4, MOV, AVIF, and HEIC: an ISO-BMFF
jumbbox.
The manifest can include the producer, claim_generator, date, instance ID, and assertions such as c2pa.created. A trusted implementation signs the structure, so changing it breaks the signature.
The current specification is 2.4 (c2pa.org, accessed August 2026). It also defines durable Content Credentials: a cryptographic hard binding paired with a watermark or fingerprint soft binding, so provenance can be recovered after metadata disappears.
The acknowledged weakness: manifests can be removed
The C2PA specification explicitly says provenance metadata can be removed. The manifest is separable from the content; re-encoding, taking a screenshot, or stripping metadata can delete it.
Platforms do this at scale through upload re-encoding, storage optimization, and privacy policies. Social networks have regularly stripped C2PA metadata, a behavior documented by standards observers since January 2025.
Why generators now use two signals
Providers add a latent watermark to compensate for removable metadata.
- Google embeds SynthID in Gemini and Imagen images, Veo video, and Lyria audio.
- OpenAI combines C2PA + SynthID on images generated with ChatGPT, Codex, and the API, and uses SynthID on audio (source: help.openai.com, accessed August 2026).
If metadata is removed by you or a platform, the embedded signal can still indicate origin. That is why C2PA removal and latent-signal treatment are separate jobs.
What this means for you
- Platforms can remove C2PA without removing a latent watermark.
- Missing C2PA metadata does not prove that a file is not AI-generated.
- Cleaning your own content may require both surgical metadata removal and documented latent-signal purification.
The EU AI Act response
Since August 2, 2026, Article 50 of Regulation (EU) 2024/1689 requires general-purpose model providers to make synthetic outputs detectable with machine-readable marking. It also requires deployers of deepfakes to disclose the artificial nature of the content.
Summary
- C2PA is signed, detailed metadata that can be separated from a file.
- SynthID is a durable latent signal with less context.
- OpenAI and Google combine both layers in 2026.
- Removing metadata from your own files is lossless; use remains subject to applicable law.