Standard
C2PA Content Credentials for AI Provenance
A Content Credential is a signed manifest, not a watermark. It bundles machine-readable assertions (c2pa.actions for edit history, c2pa.ingredient for source assets, c2pa.hash.data for a tamper-evident hash) into a claim signed through an X.509 certificate chain, the same trust model that secures HTTPS. OpenAI attaches one to ChatGPT, Codex, and API images alongside an invisible SynthID watermark, except gpt-image-1 and gpt-image-1.5 images through the API, which OpenAI's help page excludes, precisely because the two fail in different ways. Coverage elsewhere is uneven: Amazon's Titan and Nova Canvas images carry C2PA plus a watermark by default, Microsoft's marks are off unless someone turns them on, Meta has not shipped video marking, and OpenAI lists the discontinued Sora's video as Content Credentials only.
Correction,
The page said every image from the OpenAI API carries a Content Credential plus SynthID. OpenAI's help page, read 2026-10-06, excludes gpt-image-1 and gpt-image-1.5 from API coverage, so API images from those models should not be assumed to carry either signal. It also said the November 2025 Sora investigation used OpenAI's own Verify tool; this site's records agree only that one of its two tools was the open-source c2pa-rs CLI, and OpenAI's Verify tool does not accept video. Both passages are corrected.
Correction,
Carried from the C2PA vs SynthID page, now part of this one, which had four claims wrong. It said Anthropic uses the same two-layer pattern as OpenAI; Anthropic's watermark is in text and its C2PA metadata is on generated files, so no single Claude output carries both. It presented Stable Signature as Meta's deployed watermark; Stable Signature is 2023 research, and Meta's shipped image mark is Content Seal, launched 2026-07-07. It said the open-source SynthID Text detector lets anyone check text; it only scores text marked with keys you hold, and Google's production keys have never been published. And it said OpenAI's API images carry both layers without noting the gpt-image-1 and gpt-image-1.5 exclusion.
Correction,
Carried from the video AI watermarking page, now part of this one, which had three claims wrong. It said Sora videos carry SynthID alongside C2PA and described Sora as a running product; OpenAI's help page, read 2026-10-06, lists video as Content Credentials only and says Sora is discontinued. It said OpenAI's Verify tool was one of the two tools the November 2025 investigation used; OpenAI's Verify tool does not accept video. And it said Microsoft's Copilot image watermark is always on; Microsoft documents it as a per-user setting, so image, video and audio marks are all off unless somebody turns them on.
Key takeaways
- A Content Credential bundles typed assertions (c2pa.actions for edit history, c2pa.ingredient for source assets, c2pa.hash.data for the tamper-evident hash, plus EXIF/IPTC/XMP-mapped metadata) into one signed claim. It is not a tag and not a watermark.
- Hard binding works like a tamper-evident seal. A SHA-256, SHA-384, or SHA-512 hash is computed over the file's bytes at signing time and locked into the manifest, so a verifier recomputing it catches any later edit, however small.
- The signing chain mirrors HTTPS: a private key signs the manifest, an X.509 certificate proves who signed it, and that certificate is checked against a C2PA Trust List. That proves who signed and that nothing changed, not that the depicted content is true.
- OpenAI attaches C2PA Content Credentials to images from ChatGPT, Codex, and the API alongside an invisible SynthID watermark built with Google DeepMind, except gpt-image-1 and gpt-image-1.5 images through the API, which OpenAI's help page excludes. As of the July 31, 2026 audio rollout, its audio outputs carry only SynthID, with no C2PA layer yet.
- Content Credentials can be removed. ExifTool lists the C2PA/JUMBF box as a format it can read and delete, and a screenshot, a recompressing upload, a format conversion or a re-save in most editors also drops the manifest. None of that touches an invisible watermark such as SynthID, which is embedded in the content itself.
- SynthID is checked by a statistical detector that can answer "uncertain", and only approved testers can run Google's detector on images, video and audio. A C2PA manifest is a yes-or-no lookup anyone can run with a free browser tool.
- Coverage differs by generator. Amazon's Titan Image Generator and Nova Canvas add C2PA plus an invisible watermark by default, Microsoft's image, video and audio marks are off unless somebody turns them on, Meta has promised but not shipped video marking, and OpenAI lists Sora video, now discontinued, as Content Credentials only.
Provenance stack
Where the signal lives
Metadata
C2PA and file manifests carry signed origin and edit history.
Embedded signal
SynthID or media watermarks live inside pixels, audio, video, or text.
Detector
Verification tools check whether a supported signal is present.
Figure 2
What survives what: four signals against eight ordinary things people do to content
The single most useful comparison in provenance, because it explains why a file that was genuinely AI-generated so often arrives with nothing left to find.
| What happens to the content | C2PA metadata | Embedded media watermark | Statistical text watermark | Visible label |
|---|---|---|---|---|
| Screenshot | Destroyed | Usually survives | Not applicable | Survives |
| Re-encode or compress | Usually destroyed | Degrades | Not applicable | Survives |
| Crop | May survive | Degrades | Not applicable | Often removed |
| Upload to a social platform | Commonly stripped | Usually survives | Not applicable | Survives |
| Strip metadata deliberately | Destroyed | Untouched | Not applicable | Untouched |
| Paraphrase the text | Not applicable | Not applicable | Degrades with edit strength | Removed |
| Translate the text | Not applicable | Not applicable | Usually lost | Removed |
| No detector available | Public verifiers exist | Provider-gated | Gated preview only | Human-readable |
Method C2PA behaviour follows from the specification's binding of a manifest to file bytes. Media and text watermark rows follow provider documentation and the published robustness literature. Cells describe expected behaviour, not a bench test by this site.
Checked 2026-08-13
Full comparisonWhat is inside a C2PA Content Credential?
Confirmedc2pa.actions, c2pa.ingredient, c2pa.hash.data: the named parts of a manifest, plus the split recording which assertions the signer will actually stand behind.
A Content Credential is not a single tag. C2PA's explainer defines a manifest as "the set of information about the provenance of an asset consisting of one or more assertions that are digitally signed."
In practice, a handful of assertion types do most of the work.
- c2pa.actions: a log of what was done to the file (created, opened, cropped, color-corrected, composited), where each entry can carry a digitalSourceType field stating whether that step came from a camera, a human edit, or an AI model.
- c2pa.ingredient: a pointer to any source asset folded into this one, such as the photo a generative-fill edit started from.
- c2pa.hash.data: the cryptographic hash that hard-binds the manifest to the file's actual bytes.
- c2pa.thumbnail: a small preview image.
- Standard metadata assertions that map onto existing EXIF, IPTC, and XMP fields rather than reinventing camera or caption metadata from scratch.
All of that is organized by a "claim," the structure that gathers together everything the signer is vouching for. The claim also records the claim generator (the specific product and version that did the signing) plus a unique instance ID for that version of the file.
When assets get combined, each ingredient carries its own provenance forward. C2PA describes the result as "a tree of provenance, much like a family tree, that can stretch all the way back to each ingredient's creation."
How does C2PA hard binding detect tampering?
ConfirmedHere's the mechanism that makes tampering visible without anyone having to spot it by eye.
Hard binding locks a hash of the file's bytes into the signed manifest, so any later edit makes the hash a verifier recomputes stop matching. The c2pa.hash.data assertion records a cryptographic hash (SHA-256, SHA-384, or SHA-512) computed over specific byte ranges of the real file, deliberately excluding the manifest itself. That hash is then rolled up into the rest of the signed data.
C2PA's explainer describes the technique directly: "standard cryptographic hashes, such as SHA2-256, with a Merkle tree-like approach of having one hash included in a secondary hash (etc.) and then finally digitally signed."
The practical effect is simple. Change a single pixel after signing and the hash a verifier recomputes from the file no longer matches the one locked into the manifest, so tampering surfaces automatically, with no cooperation from the editor required.
The weakness is just as simple. Strip the manifest out of the file and there is nothing left to hash-check.
C2PA's answer is to pair hard binding with a "soft binding" (watermarking or fingerprinting) to create what the spec calls durable Content Credentials, so a manifest can still be found after it's been physically removed from the file.
OpenAI cites the same tradeoff for its own stack. Its help center notes that metadata "can sometimes be removed by platforms, editing tools, or file conversions," while a watermark "is part of the content itself and may persist through some edits or transformations."
How do you remove C2PA Content Credentials, and what does that not remove?
ConfirmedWhat strips a manifest, what does not, and why a stripped file can still carry an invisible watermark.
A C2PA manifest is stored as a metadata box in the file, so a metadata tool can delete it. ExifTool's documentation lists C2PA/JUMBF, the box format the manifest is stored in, as a format it can read and delete, and calls it "not yet writable."
That is why a C2PA remover is a one-command job, while putting back a manifest that still validates is not. A new manifest needs a signing certificate chaining to a Certification Authority on the C2PA Trust List, which a metadata editor cannot produce.
These also strip or lose the manifest, without any intent to remove it.
- A screenshot. An OS-level screen capture produces a new file from rendered pixels, with no path back to the original file's manifest.
- Recompression or resizing during upload.
- Format conversion.
- Most editing tools that re-save the file.
Re-encoding breaks the hard binding even if the metadata box happens to survive, because the manifest depends on a hash of the exact file bytes.
Meta says "it's not yet possible to identify all AI-generated content, and there are ways that people can strip out invisible markers." OpenAI says C2PA metadata "can be stripped, lost through uploads and downloads, or broken by transformations."
Anthropic's help page scopes signed provenance metadata to "where Claude supports processing files", and on AWS, Google Cloud and Microsoft Foundry says it "applies only where a platform offers Claude's file generation features". Coverage is uneven before anything gets stripped.
A missing manifest means no metadata survived. It is not evidence that the file was not AI-generated, which is why the spec also defines a soft-binding fallback of watermarking and fingerprinting.
How does C2PA compare with SynthID?
ConfirmedA signed wrapper around the file against a signal inside the content: what each survives, how each is checked, and why providers ship both.
C2PA wraps a signed manifest around the file. SynthID, developed by Google DeepMind, embeds a signal in the generated pixels, audio or video during creation and never touches file metadata.
For text, SynthID runs as a logits processor that nudges the model's token probabilities with a pseudorandom function. It leaves a statistical pattern across word choices, not a file-level marker.
| Question | C2PA Content Credentials | SynthID |
|---|---|---|
| Where does the signal live? | In a signed metadata manifest bound to the file's exact bytes | In the pixels, audio samples or token choices |
| What does it record? | Issuer, generating tool, edit history and source assets | Much less: whether the content carries the mark |
| Does it survive a screenshot or re-encode? | Usually not; the manifest is lost or its hash stops matching | Designed to survive cropping, filters, changed frame rates and lossy compression |
| How is it checked? | A manifest lookup: present and valid, or not | A statistical detector that returns watermarked, not watermarked or uncertain, against a threshold the operator sets |
| Who can check it? | Anyone, with a free reader such as Content Credentials Verify | Images, video and audio: approved testers only. Gemini text: nobody outside Google, because the open-source text detector only scores text marked with keys you hold |
Google says the audio version of SynthID is meant to resist added noise, MP3 compression and changes to playback speed.
That durability is trained in. A Google-assigned patent for an "End-to-End Watermarking System" (US 12,238,322 B2, filed 2022, granted 2025) describes training a watermark encoder and decoder alongside a network that distorts the image, so the decoder learns to recover the signal anyway. The filing is never branded SynthID.
None of these claims is unconditional. Google says SynthID Text's detector confidence can be greatly reduced when text is thoroughly rewritten or translated.
Published research has removed the broader class of low-perturbation pixel watermarks with diffusion purification, and with UnMarker, a black-box attack presented at IEEE S&P 2025. No source found reproduces either attack against SynthID-Image, Content Seal or Amazon's watermark by name.
Products do not choose between the two. OpenAI pairs C2PA with SynthID on the same images because, in its words, C2PA helps content carry detailed context while SynthID helps preserve a signal when metadata doesn't survive. That is the soft binding the C2PA spec defines as its recovery path.
Anthropic uses both on different outputs: a version of Google DeepMind's SynthID-Text approach in generated text, and signed C2PA metadata on supported Claude-generated files. No single Claude output carries both, so neither layer backs up the other.
How can I check whether a file still has Content Credentials?
ConfirmedA free browser tool reports whether a manifest is present and whether it still validates. Checking for a watermark is a separate and more restricted job.
The Content Credentials Verify tool at verify.contentauthenticity.org is maintained by the contentauth organization on GitHub. Its repository describes it as reading and validating C2PA data "on the client side using WebAssembly," so the file is parsed in your browser rather than uploaded for inspection.
- Whether a manifest is present at all.
- Whether the hash and signature still validate.
- What the manifest claims about origin and AI involvement.
ExifTool, Phil Harvey's command-line metadata utility, reads the raw C2PA/JUMBF data directly. To see the effect of removal, run an image through the Verify tool, then screenshot it and run the screenshot.
Checking for a watermark needs the provider's own detector. OpenAI's tool at openai.com/research/verify/ reads both its C2PA manifest and its SynthID watermark, but accepts images and audio only, so no OpenAI tool checks video.
Google's Gemini app documents checks for both SynthID and Content Credentials, capped at 100MB per file, video under 90 seconds, audio under an hour and roughly 10 checks per media type per 24 hours.
Google's standalone SynthID Detector portal, launched at I/O on May 20, 2025, is limited to approved journalists, media professionals and researchers, with no public API. Journalist Henk van Ess found in November 2025 that its video and audio detection were still hit-or-miss, with only image detection reliable.
What does a verified Content Credential actually prove?
ConfirmedA verified credential proves the signer is who the certificate says and that nothing changed since. C2PA is blunt about everything it does not prove.
A verified Content Credential proves only who signed the file and that nothing was altered after signing. Once the assertions and claim are assembled, the bundle is signed by whoever, or whatever software, performed the operation. C2PA's explainer says the manifest is "signed with the private key of the software or hardware that performed the operations, thus ensuring its authenticity and integrity. The corresponding public key is made available for verification."
It compares the design to "SSL/TLS and PDF signatures" using X.509 certificates, the same technology behind the padlock in your browser. A verifier checks the signature against the signer's public key, then checks that certificate against a trust anchor.
For C2PA, that anchor is the C2PA Trust List, which the explainer says works "similar to how a web browser or PDF viewer includes their own set of Trust Lists (e.g., CA/Browser Forum, EUTL, etc.)." Getting onto it runs through C2PA's Conformance Program, which independently tests three separate things:
- Generator products, the tools that create and sign manifests.
- Validator products, the tools that check them.
- Certificate authorities, who issue the signing certificates in the first place.
None of this is a truth guarantee, and C2PA says so plainly: "the trust in the contents of a Content Credential is based on the trust relationship between the creator of the Content Credential and the consumer." Verification proves the signer is who the certificate says and that nothing was altered after signing. It does not prove the signer told the truth.
OpenAI states on its provenance page that it has become "a C2PA Conforming Generator Product." That is OpenAI's own announcement, not something this site could check against a public conformance registry, so treat it as self-reported rather than third-party-verified.
Which AI image and video generators add Content Credentials or a watermark?
ConfirmedOpenAI, Google, Meta, Amazon and Microsoft by images and video, including the entries that describe a capability rather than a default.
Every major image generator now embeds something. Video coverage is thinner, and Microsoft's marks are off until somebody turns them on.
| Provider | Images | Video |
|---|---|---|
| OpenAI | C2PA plus SynthID from ChatGPT, Codex and the API since 2026-05-19, except gpt-image-1 and gpt-image-1.5 through the API | Sora: Content Credentials only, no SynthID. OpenAI says Sora is discontinued and excludes sora-2 from API coverage |
| SynthID across Google's services | SynthID, the same system Google applies to images, audio and text | |
| Meta | Content Seal on Muse Image output since 2026-07-07; photorealistic AI images also carry visible markers, invisible watermarks and metadata that Meta does not specify as C2PA | None shipped. Meta says it plans to extend Content Seal to video soon |
| Amazon | Titan Image Generator and Nova Canvas: invisible watermark plus C2PA metadata by default, at no extra charge | Nova Reel: an invisible watermark in every frame, designed to survive H.264 compression |
| Microsoft | Watermark is a per-user setting, off unless turned on; provenance information goes into image metadata | Clipchamp invisible watermark and AI-generated notice, off until a tenant admin enables them; no provenance metadata yet |
Google's SynthID-Image paper says the system "has been used to watermark over ten billion images and video frames across Google's services." That is Google's own figure, not a third-party audit.
A separate AWS post reportedly adds C2PA Content Credentials to Nova Reel 1.1. This site has not re-verified it.
Amazon also ships a detector. Bedrock's Watermark Detection scores whether an image carries a Titan or Nova Canvas mark; it is in public preview and runs only in us-east-1 and us-west-2.
Microsoft documents video and audio marks behind a tenant admin Cloud Policy, and says provenance information is added only to the metadata of images so far, with no time frame for video and audio. Three details of the video policy matter:
- The policy defaults to Disabled or Not Configured, so nothing is marked until an administrator changes it.
- It is not available for US Government GCC, GCC High or DoD tenants.
- The notice wording and placement cannot be customized once it is on.
A provider confirming that it can mark an image or video is a different claim from a specific file in front of you carrying a detectable mark.
What does an OpenAI image's Content Credential carry?
What does OpenAI actually attach to a generated image? Confirmed coverage for images, audio and video first, then a field-by-field sketch that is explicitly inference.
An OpenAI image carries a C2PA Content Credential plus an invisible SynthID watermark, whether it came from ChatGPT, Codex, or the OpenAI API, except gpt-image-1 and gpt-image-1.5 images through the API, which OpenAI's help page excludes. Coverage varies by content type, and OpenAI's own materials are specific about it.
- Images from ChatGPT, Codex, or the OpenAI API: a C2PA Content Credential plus an invisible SynthID watermark, a joint capability built with Google DeepMind and expanded through 2026. API images from gpt-image-1 and gpt-image-1.5, which OpenAI's help page lists as "planned for imminent deprecation", are excluded and should not be assumed to carry either signal.
- Audio, as of a July 31, 2026 update: a SynthID watermark only, with no C2PA metadata layer yet.
- Video from Sora, which OpenAI's help page, read 2026-10-06, says is discontinued: Content Credentials only, with no SynthID. The same page excludes sora-2 from API coverage.
- A visible OpenAI watermark, if you ask a model for one: a separate feature that is explicitly not machine-readable, and not the same thing as either embedded signal.
OpenAI's public tool at openai.com/research/verify/ checks an uploaded file for a SynthID watermark or a trusted C2PA manifest associated with OpenAI. A positive match means a supported signal was found.
A negative result proves far less. The file could predate the rollout, come from an unsupported product or export path, or have had its signals stripped or degraded on the way to you.
What OpenAI has not published is the literal field-by-field manifest JSON for a specific image. Combining its own description with the C2PA assertion vocabulary and IPTC's digitalSourceType codes, a plausible (explicitly inferred, not confirmed) sketch of a pure text-to-image generation would include:
- A c2pa.actions assertion with a "created" action whose digitalSourceType points to IPTC's trainedAlgorithmicMedia code ("digital media created algorithmically using an Artificial Intelligence model trained on captured content").
- A c2pa.hash.data assertion hard-binding the manifest to the image bytes.
- A claim generator string identifying the OpenAI product and version.
- A claim signature chaining to an OpenAI-controlled certificate.
For an edit rather than a generation, say you upload a photo and ask for an object to be removed, you would also expect a c2pa.ingredient assertion referencing that uploaded source file. The relevant action's digitalSourceType would more plausibly use IPTC's compositeWithTrainedAlgorithmicMedia code instead. Treat that whole sketch as informed inference from verified building blocks, not a captured real-world manifest.
Have OpenAI files been found without provenance signals?
Community discussionTwo independent checks, both finding OpenAI provenance signals absent from files that should have carried them. Neither has a documented response from OpenAI.
Two independent checks have found gaps between the public claims and the actual files. Neither has a documented response from OpenAI, and neither has been replicated by this site.
In November 2025, a named author cross-posted an investigation to LessWrong and the EA Forum. Sora 2 videos carrying the visible watermark showed no detectable C2PA metadata through either of two verification tools, one of them the open-source c2pa-rs CLI. OpenAI's own Verify tool does not accept video.
OpenAI's help page, read 2026-10-06, now lists video as Content Credentials only and says Sora is discontinued. It does not address the November 2025 finding, and with Sora discontinued no new output can be generated to settle it.
The second is a June 28, 2026 thread on OpenAI's own developer forum, asking whether C2PA metadata chunks are being stripped from images uploaded to ChatGPT on purpose. The poster's OpenAI-generated PNG verified cleanly at openai.com/research/verify/ with Content Credentials and SynthID both present, then came back from a round trip through a ChatGPT conversation stripped down to bare PNG chunks.
Both sit at community-discussion confidence, and both are worth knowing because they set expectations. A provenance signal is only as good as every pipeline the file passes through afterward.
What does the law require for AI-generated video?
Official announcementTwo separate EU duties land on video, and New York requires disclosure of a fabricated face in ads while exempting audio-only ads.
EU AI Act Article 50 covers synthetic video on both sides of its scope. Article 50(2) is the provider's duty to mark synthetic output in a machine-readable, detectable format; Article 50(4) is the separate deployer duty to label deepfakes.
The deepfake duty applies to realistic AI-generated or manipulated video resembling real people, objects, places or events, regardless of any intent to deceive. Which duty lands on you is covered in EU AI Act and AI watermarking.
New York's Synthetic Performer disclosure law, effective June 9, 2026, requires conspicuous disclosure on any visual or audiovisual ad shown to New York consumers that includes an AI-generated human likeness not depicting a real person.
Penalties are $1,000 for a first violation and $5,000 for each one after that. Audio-only ads are explicitly exempt.
FAQ
Can Content Credentials be stripped or removed from an image?
Yes, and both C2PA's explainer and OpenAI's help center say so plainly: metadata is fragile. Uploading to a platform that re-encodes files, converting formats, taking a screenshot, or running an image through most editing tools can drop the manifest entirely, and a verifier can't tell "never had one" from "had one and lost it." That's exactly why OpenAI pairs C2PA metadata with an invisible SynthID watermark on generated images instead of shipping metadata alone.
Does a Content Credential prove an image is authentic or "real"?
No, and C2PA is explicit about this in its own explainer. A verified Content Credential shows only that the manifest data is "well-formed and free from tampering" and was signed by an identity tied to a known trust list. A photo can carry a perfectly valid, unmodified Content Credential that truthfully documents a heavily staged scene, because verification checks the chain of custody, not the honesty of what's depicted.
What's the actual difference between a Content Credential and a watermark like SynthID?
A Content Credential is metadata, a signed record riding alongside or inside the file, describing who created it, what tools touched it, and in what order, which is why it can carry far more detail. A watermark like SynthID is embedded in the pixels or audio samples themselves, carries much less information, and tends to survive transformations that destroy metadata. OpenAI frames this as a deliberate tradeoff: metadata carries more context but can be removed, watermarks may be more durable but say less about what happened.
Is C2PA only used for labeling AI-generated content?
No. Its steering committee (Adobe, Amazon, the BBC, Google, Meta, Microsoft, OpenAI, Publicis Groupe, Sony, TikTok, and Truepic) spans camera makers, publishers, and social platforms, not just AI labs. Camera manufacturers and news organizations use C2PA to certify the source and edit history of ordinary photos and video that never touched a generative model, using the same assertion types that flag an AI edit to record a conventional crop or color correction.
Can Content Credentials be deleted from a file with a metadata tool?
Yes. ExifTool's own documentation lists the C2PA/JUMBF box as a format it can read and delete, alongside standard EXIF, XMP, and IPTC fields. What it cannot do is forge a new, validating manifest in its place, since that requires a signing certificate chaining to a Certification Authority on the C2PA Trust List.
Does removing C2PA metadata also remove an invisible watermark like SynthID?
No. A Content Credential is metadata stored in the file, and deleting it does not change a watermark embedded in the pixels or audio samples. That is why OpenAI pairs C2PA metadata with SynthID on generated images, and why C2PA's own soft binding relies on watermarking and fingerprinting to look a manifest up again after it has been stripped.
If a photo or video has no Content Credentials, does that mean it was not made or edited with AI?
No. Read a missing manifest as "no metadata survived," not as evidence about origin. Screenshots produce a new file with nothing to carry a manifest, and most re-encoding on upload or re-save changes the bytes the hard-binding hash was computed over, breaking validation even if a manifest was originally attached.
Can I check whether a file has valid Content Credentials without installing anything?
Yes. The Content Credentials Verify tool at verify.contentauthenticity.org runs in the browser and validates the manifest client-side using WebAssembly, per its maintainers' own repository. It reports whether the hash still matches the file, whether the signature validates, and what the manifest claims. No account or third-party upload is required for the check itself.
Which is harder to remove: C2PA or SynthID?
In practice a SynthID-style embedded watermark is harder to remove than C2PA metadata, which is the reason OpenAI gives for shipping both on the same images. OpenAI says C2PA metadata can be stripped, lost through uploads and downloads, or broken by transformations like resizing or screenshots, while watermarking tends to survive those better. Harder is not impossible: Google says heavy rewriting or translation can greatly reduce SynthID Text's detector confidence, and defeating an image or audio watermark under determined adversarial editing is an open research problem.
If a screenshot strips the C2PA metadata, is there any way to tell an image came from an AI tool?
Only if a separate embedded watermark was applied too, and you have access to its detector. A screenshot destroys the manifest, but a SynthID mark lives in the pixels and can still be detected. If no watermark was applied, or you cannot reach that vendor's detector, there is currently no reliable way to establish origin after a screenshot.
Are C2PA and SynthID competitors, or are they used together?
Used together. OpenAI pairs C2PA metadata with SynthID on the same ChatGPT, Codex and API images, except gpt-image-1 and gpt-image-1.5 images through the API. Anthropic uses a text watermark and C2PA metadata on files, so on different outputs. Meta runs its Content Seal image watermark alongside separate AI-content metadata labeling.
How do I check whether a file has a C2PA credential or a SynthID watermark?
Different tools for different signals. For C2PA, use a manifest reader such as Content Credentials Verify, or openai.com/research/verify/ for OpenAI images and audio. For SynthID on images, video and audio, Google's checker is the SynthID Detector, through the Gemini app or a portal limited to approved testers. Google's open-source text detector only works on text marked with your own keys, so it cannot check Gemini output. No single tool checks both, and no OpenAI tool accepts video.
Does every AI video carry a watermark or Content Credentials?
No. It depends on the provider, the product and sometimes a tenant's settings. Microsoft's Copilot video watermarking is off unless an IT administrator turns it on, Meta has not shipped video marking, and an independent check found Sora 2 videos with the visible watermark but no detectable C2PA metadata.
Is OpenAI's Sora watermarking claim fully verified?
Not consistently. An independent November 2025 investigation found videos carrying Sora's visible watermark with no detectable C2PA metadata, and videos without the visible watermark carrying it, the opposite of OpenAI's claim that every Sora video includes both signals. No OpenAI response has surfaced. OpenAI's help page, read 2026-10-06, lists video as Content Credentials only and says Sora is discontinued, so no new output can be generated to settle it.
Does Meta watermark AI video?
Not yet, as far as any announcement shows. Content Seal covers images generated with Muse Image in the Meta AI app and meta.ai, and Meta says it plans to extend Content Seal to video soon. As of this site's 2026-09-25 recheck no video launch has been announced.
Is there a published way to remove SynthID, Content Seal or Amazon's image watermark?
Not against those specific systems. Research has shown diffusion purification can defeat the broader category of low-perturbation invisible watermarks, that detection can be fooled into flagging unwatermarked images, and that a fully black-box attack (UnMarker) exists. No source found reproduces any of those attacks against SynthID-Image, Content Seal or Amazon's watermark by name.
Next steps
- Drop one of your own files into the Content Credentials Verify tool and see what actually survived its last upload. Content Credentials Verify
- Read what independent red teams found when they attacked SynthID across text, images, audio and video. SynthID watermark
- If your organization runs Microsoft 365, check whether the AI-content watermark Cloud Policy is enabled in your tenant before assuming your video output is marked. Microsoft Learn: Copilot watermarks
- Read what a missing manifest does and doesn't tell you about a file's origin before you draw conclusions from one. AI content provenance
- See how removal works across every layer, not just the metadata one, before assuming a stripped file is untraceable. Can AI watermarks be removed?
Sources and citation status
- OfficialC2PA Content Credentials explainer
- OfficialC2PA specifications
- OfficialOpenAI: content provenance
- OfficialC2PA: Steering Committee & Overview
- OfficialC2PA Technical Specification 2.3
- OfficialIPTC Photo Metadata: digitalSourceType controlled vocabulary
- OfficialOpenAI Help: provenance signals in OpenAI-generated content
- OfficialC2PA Conformance Program
- CommunityOpenAI Developer Community: are C2PA metadata chunks stripped from images uploaded to ChatGPT? (2026-06-28)
- CommunityLessWrong: OpenAI does not appear to be applying watermarks honestly (2025-11-06)
- OfficialC2PA FAQ
- OfficialContent Credentials Verify tool
- OfficialContent Credentials Verify: source (contentauth/verify-site, GitHub)
- OfficialExifTool documentation
- OfficialMeta: Labeling AI-generated images on Facebook, Instagram and Threads
- OfficialClaude Support: how Claude marks AI-generated content
- OfficialGoogle AI Developers: SynthID
- OfficialGoogle DeepMind: SynthID technology overview
- ResearchSynthID-Image paper (arXiv)
- OfficialGoogle LLC: End-to-End Watermarking System (US Patent 12,238,322 B2)
- ResearchFernandez, Couairon, Jégou, Douze, Furon: The Stable Signature (arXiv)
- OfficialMeta: introducing Muse Image and Muse Video (Content Seal)
- ResearchSaberi, Sadasivan, Rezaei, Kumar, Chegini, Wang, Feizi: Robustness of AI-Image Detectors (arXiv)
- ResearchKassis, Hengartner: UnMarker, a universal attack on defensive image watermarking (arXiv)
- OfficialAWS Bedrock: Titan image models documentation
- OfficialAmazon Science: Nova and our commitment to responsible AI
- OfficialMicrosoft Learn: watermarks for AI-generated content in Copilot
- OfficialGoogle: SynthID Detector portal launch
- OfficialGemini Apps Help: verify AI-generated images, videos, and audio
- ReportingDigital Digging (Henk van Ess): Google's SynthID, three tools, three results (2025-11-20)
- ReportingLead Stories via Yahoo Tech: Gemini's SynthID checker returned verdicts for the wrong upload (2026-07-13, updated 2026-07-16)
- RegulatoryEU AI Act Article 50
- RegulatoryNew York Governor: first-in-the-nation law requiring AI disclosure in advertisements