To prove where a piece of media came from, record its origin when it is made instead of guessing afterward. The open C2PA standard, shown to people as Content Credentials, attaches a signed record of origin and edits to the file. Watermarks add a hidden mark. Detectors only estimate, and none of these proves the content is true.
- Detection guesses origin after the fact and returns a probability. Provenance declares origin at creation and signs it.
- A C2PA manifest bundles statements about the file, a signed claim and cryptographic hashes that bind the record to the exact bytes.
- Screenshots, re-encoding and platforms that strip metadata detach the record. A missing record proves nothing either way.
- Watermarks survive more edits but carry less information, and the C2PA standard uses them to reconnect a stripped file to its record.
- A valid signature proves who made the claim, not that the scene is real.
- In the EU, the AI Act's transparency obligations for AI-generated content apply from 2 August 2026.
Detection vs provenance: guessing origin or declaring it
There are two ways to answer "where did this image come from?". Detection looks at the finished file and infers its origin from statistical traces: texture patterns, lighting inconsistencies, artefacts typical of a particular generator. Provenance takes the opposite approach: origin is recorded and signed when the file is made, so nobody has to infer it later.
Detection degrades as generators improve. Every trace a detector learns to read is a trace the next generation of tools may stop leaving, so the method chases a moving target. It also produces an uncomfortable kind of output: a probability, not a fact. A score saying a photo is "likely AI-generated" is a weak basis for a consequential decision, such as rejecting a news photo or accusing a student, because it is wrong often enough to be unfair.
Provenance avoids guessing but pays a different price. It only works for files whose origin was recorded at creation, so it says nothing about older media or anything made by a tool that does not sign. And the absence of a record is not evidence of anything: a plain file is not a guilty one. Provenance can confirm the origin it captured; silence stays silence.
What a Content Credential contains
The most widely backed provenance standard comes from the Coalition for Content Provenance and Authenticity (C2PA), whose members include major technology, camera and media companies. Its specification calls Content Credentials "the preferred non-technical term" for a C2PA manifest. Version 2.2 of the specification, dated May 2025, defines the building blocks (C2PA technical specification):
- Assertion
- A single statement made by the signer about the file, for example that it was captured by a camera, generated by an AI model or edited in a particular way.
- Claim
- A digitally signed, tamper-evident structure that references the set of assertions about an asset.
- Claim signature
- The digital signature on the claim, made with the signer's private key, which lets anyone confirm who signed and that nothing has changed since.
- Hard binding
- One or more cryptographic hashes that identify the exact bytes of the asset, so the record cannot be moved onto a different file unnoticed.
- Soft binding
- A content identifier, such as a fingerprint or an invisible watermark, that can help reconnect a file to its record after the two have been separated.
- Manifest
- The package that holds this provenance information. A file can carry a series of manifests, recording its creation and later edits.
The difference from ordinary metadata is the seal. Ordinary metadata is a sticky note: easy to write, easy to change, trusted only by habit. A signed manifest is a sealed envelope whose seal names its maker and breaks visibly when tampered with. Trust moves from the claim to the cryptography underneath it.
What breaks the chain of custody
A chain of custody is the unbroken, signed history linking a file's current state to its origin. It is strong while every handoff is recorded and weak the moment one is not. Most breaks are not attacks. They are ordinary ways media moves around:
- Re-encoding. Compressing an image, transcoding a video or passing a file through a system that drops unfamiliar data can remove the manifest while the pixels survive.
- Screenshots and screen recordings. These create a brand-new file that looks identical and carries no history at all.
- Uploads to platforms. Many sites resize or optimise uploaded images, and that processing may or may not keep the record.
- Edits in tools that don't support the standard. The edit is not recorded, and the chain stops there.
The specification acknowledges this directly: "An asset can become separated from its C2PA Manifest due to removal or corruption of asset metadata." Its answer is the soft binding. If a manifest was stored elsewhere, a watermark or fingerprint in the content can be used to find it again.
Fragility is partly the point. A record that survived any transformation could be transplanted onto a forgery. The design accepts that a broken chain produces no claim rather than a false one. The cost is that much media in circulation arrives stripped, and a recipient must treat a missing record as missing information, never as proof.
Watermarks vs signed provenance
Watermarks and provenance records are often mentioned together, but they do different jobs. A watermark is a signal embedded in the content itself. Google DeepMind's SynthID, for example, adds imperceptible watermarks to images, video, audio and text produced with Google's AI tools, and the company says its image and video watermark is designed to withstand cropping, filters, frame-rate changes and lossy compression (SynthID). A provenance manifest travels alongside the content and is checked against a signature.
| Dimension | Embedded watermark | Signed provenance manifest |
|---|---|---|
| Where it lives | Inside the image, audio or text itself | Attached to the file, bound by hashes and a signature |
| Survives re-encoding and screenshots | Often, by design | Usually not |
| Information carried | A short identifier | A full history of origin and edits |
| Who can check it | Typically a detector from the watermark's provider | Anyone with a compatible validator |
| Main job | Persistence through transformation | Attribution and tamper evidence |
The mature view treats them as layers, and the C2PA standard does exactly that: it counts watermarks and fingerprints as soft bindings, a thread that can lead back to the envelope when the envelope is lost. Asking which one is right is like asking whether you need a thread or an envelope. It depends on what the media is about to go through.
How to check where an image or video came from
Verification is an ordered process. Skipping a step lets a confident-looking file pass on appearance alone.
- Look for Content Credentials. Open the file in a C2PA-conforming validator. The Content Authenticity Initiative documents Adobe's Inspect tool, which reads credentials from common image, video, audio and PDF formats (how to inspect Content Credentials).
- Treat absence as unknown. No credential means no information, not a fake.
- Check the signature and the signer. A valid signature tells you the record is intact; the signer's identity tells you whose claim you are trusting.
- Confirm the binding. The record must describe these exact bytes, not a different file it was copied from.
- Read the history in order. Note where the asset was captured or generated, and every edit since.
- Look outside the file. Search for earlier copies, and check who first published it and whether the context matches.
- Match the evidence to the claim. Trust only what the signed steps actually cover.
What this cannot do matters as much as what it can. A valid signature proves the record is authentic. It does not prove that what the media depicts is real or fairly presented. A clip can be honestly labelled as AI-generated and still mislead about its subject. Verification establishes provenance, not truth.
What the EU AI Act requires for AI-generated content
Law is starting to require part of this. Under Article 50 of the EU AI Act, as summarised by the European Commission, providers of generative AI systems must ensure outputs are "marked in a machine-readable format and detectable as artificially generated or manipulated". Deployers must disclose deepfakes, meaning images, audio or video that resemble real people, objects, places or events and could falsely appear authentic. They must also disclose AI-generated text published to inform the public on matters of public interest, unless a person has reviewed it and an editor takes responsibility (Code of Practice on AI-generated content).
These obligations apply from 2 August 2026. A voluntary Code of Practice, published in final form on 10 June 2026, gives providers and deployers a recognised way to show compliance. Rules elsewhere differ, so treat this as one significant benchmark rather than a global standard, and take legal advice for your own situation.
Where disclosure becomes a social contract
Technical provenance ends where human agreement begins. A signature can prove a file declared its origin. It cannot make anyone attach the declaration, preserve it or read it honestly.
The gaps are predictable. A creator can decline to sign. A platform can strip records on upload. A viewer can ignore a clear label out of haste or motivated belief. Provenance raises the cost of casual deception without closing the door on deliberate fraud. That is a real gain, as long as nobody oversells it.
What tends to work is expectation. When a missing record becomes conspicuous in settings that demand one, such as news, evidence or official records, absence starts to carry meaning by convention even though it carries none by logic. Questions about who owns generated work run in parallel with questions about where it came from; ownership in AI music is one example.
How to publish media with its origin intact
If you publish photos, illustrations or video on your own site, a few habits keep the evidence attached:
- Keep the originals. Store camera or export originals with their credentials, separate from web-optimised copies.
- Sign at export where your tools support it. Some cameras and editing tools can attach Content Credentials when you save or export a file.
- Test your publishing pipeline. Upload a file with credentials, download it from the live page and inspect it. If the record is gone, your host or a processing step strips it.
- Publish a disclosure page. Explain in plain words which images are photographs, which are illustrations and whether any are AI-generated, and how people can ask about a specific image.
- Record how generated work was made. If you use AI tools, keep the prompt, the tool and the settings with the output. A detailed description of the kind covered in the craft of prompt-to-image doubles as a record, and the same habit makes results repeatable, as same prompt, different output explains.
Publish your work with WisyLink, origin included
WisyLink builds and hosts websites from a chat, which makes it a quick way to give your work a home where its origin is explained rather than implied. A portfolio, a press page or a project archive can be live at its own address after one conversation.
- Every asset has a known source. WisyLink does not generate images, video or music. A site it builds uses your own files, stock photos linked by address or hand-coded vector art, so you always know where each asset came from.
- Your files, your originals. Upload your own photos, clips and audio to the chat (each file can be up to 25 MB), while the credentialed originals stay in your own archive.
- Disclosure built in. Ask for an "About our images" page and a contact form for questions about a specific image in the same message as the rest of the site.
- Test before launch. Run the upload, download and inspect test above on the live site, as you would with any host.
For a worked example, read how to build a portfolio website with AI; the website launch checklist covers the legal and privacy pages to add, and use cases shows what else people build. When you're ready, start free and describe the site your work deserves.
F.A.Q.