← All posts

The Archibal Team

Watermarking vs metadata vs fingerprinting: which AI provenance actually survives

Watermarks, embedded metadata, and perceptual fingerprints are the three ways to mark AI content - and each fails differently. Here is how they compare, why no single one is enough, and how to layer them so provenance survives the real world.

There are three real ways to mark AI-generated content so you can show where it came from later: embedded metadata, watermarking, and perceptual fingerprinting. Each one protects against a different kind of loss. Each one also has a specific way it breaks. No single one is enough, so a sound setup layers them - and the layers come from different hands. Watermarks are usually applied by the AI provider at generation (Google's SynthID, for example). Signed metadata, a record kept outside the file and a fingerprint index are what a studio can add to its own deliveries. That is the part Archibal handles; it does not watermark.

Mechanism Where it lives Survives re-encoding? Survives metadata stripping? Survives cropping / edits? Human-readable?
Embedded metadata (C2PA) In the file's metadata Sometimes No Often no Via a viewer
Watermarking In the pixels or audio Often Yes Sometimes No
Perceptual fingerprinting In an external index n/a - recomputed Yes Light edits only No

No row is all yes. That is the entire problem.

Embedded metadata: rich, signed, and easy to lose

C2PA Content Credentials write a signed record directly into the file. The manifest describes which model made the asset, the production steps, the prompts, and who signed off. Anyone can read it with a Content-Credentials-aware viewer - verify.contentauthenticity.org is public and requires no login.

This is the most useful layer when it survives. It is structured and cryptographically signed, so a change to the signed content breaks the signature. It is also one of the techniques Recital 133 of the EU AI Act names for the machine-readable marking that Article 50 asks AI providers to apply, and real platforms now ship it - Google Pixel cameras, LinkedIn, TikTok, Sony.

The problem is that metadata is the first thing that disappears. Social platforms strip it during upload and transcoding. A screenshot drops it entirely. The moment an asset leaves a controlled environment, the manifest may simply be gone. The law splits the duties: Article 50(2) puts machine-readable marking on the AI providers, and Article 50(4) puts the human-readable label on whoever publishes. Neither duty moves when a platform strips a file - but once the manifest is gone, the file itself can no longer say how it was made.

Watermarking: resilient signal, thin record

Watermarking works differently. Instead of tagging the file container, it embeds a signal into the content itself - subtle perturbations in the pixels of an image or the waveform of audio. Because the mark lives in the perceptual data rather than the metadata, it can survive being re-saved, format-converted, or stripped of all headers.

That resilience is real and useful. As a basic "this is AI" flag, a watermark is more durable than anything in the metadata.

But a watermark carries almost no information. It can tell you that content is likely AI-generated. It cannot tell you which model produced it, what prompts shaped it, or who approved it. Detection is probabilistic, not a clean cryptographic yes or no. And a watermark can be weakened or removed by heavy cropping, regeneration, or adversarial processing - it is a signal, not a chain of custody.

Perceptual fingerprinting: durable, but only as good as your index

Fingerprinting takes a different approach entirely: it does not put anything into the file. Instead, it computes a content-based identifier from the file and stores that in an external index. A perceptual hash - a code derived from the content itself - is designed so that visually or audibly similar content produces similar codes. Compress the video, re-encode it at a lower bitrate, post it twice on different platforms: the fingerprint still matches the registered asset.

When someone hands you an asset later and asks where it came from, you compute the fingerprint and look it up. A match returns the project, date, and approval record you registered against it. Nothing in the file to strip.

The catch: this only works if you control the index and registered the asset in the first place. A stray heavily-processed derivative might fall below the match threshold. And on its own, the fingerprint tells you that an asset matches your record - all the descriptive detail lives in your system, not the file.

Why this is a layering problem, not a picking problem

Lay the failure modes side by side and the gap is obvious.

Metadata alone: the first platform upload can erase the record. Watermark alone: you get a weak probabilistic flag with no production detail. Fingerprint alone: you have a durable match mechanism but no portable evidence a third party can read off the file without access to your index.

Each one covers a gap the others leave. Metadata is rich but fragile. Watermarking is physically resilient but informationally thin. Fingerprinting is the most durable but entirely dependent on your own infrastructure. Overlap is what makes provenance holdable under pressure.

How the layers work together

The watermark, where there is one, comes from the AI provider at generation. What a studio delivering the work can stack on top is straightforward:

  1. C2PA manifest in the file - signed, portable, readable by any C2PA-aware tool, and one of the marking techniques Recital 133 of the AI Act names. (The Article 50(2) marking duty itself sits with the AI providers, not with the studio delivering the work.)
  2. A signed project record outside the file - the human-readable document that describes what shipped, what models and prompts were involved, who approved it, and when. Exists independently of whether any platform strips the metadata.
  3. A perceptual fingerprint in an external index - so even a re-encoded, metadata-stripped copy surfacing somewhere later can be traced back to the project. It finds copies, not edited versions: a recut or heavy crop may no longer match.

Decision logic: if the manifest is there, read it. If it isn't, the project record still states everything. If the copy in your hands has lost its metadata, the fingerprint can still lead back to that record. For how this fits into a production workflow, see what is AI provenance in media, and for the legal context behind why any of this matters right now, the field guide to AI disclosure laws in 2026.

Frequently asked questions

Is a watermark the same as C2PA? No. A watermark lives in the pixels or audio data. C2PA is a signed metadata manifest in the file container. The European Commission's draft Code of Practice on AI labeling actually contemplates both - complementary layers, not alternatives.

If C2PA gets stripped by social platforms, why bother with it? Because while it survives, it carries more verifiable detail than any other layer. The point of the fingerprint and the external record is that stripping the manifest is not fatal - you have other answers. Use C2PA for what it does well; don't rely on it alone.

What does the EU AI Act actually require? Article 50(2) requires the providers of generative AI systems - not the agencies using them - to mark outputs in a machine-readable way, detectable as AI-generated. It does not mandate a specific technology: Recital 133 names techniques such as watermarks, metadata, cryptographic provenance methods and fingerprints, and the Commission's draft guidance treats secured metadata and watermarking as complementary.

Which layer is most resilient? For answering "is this asset one of ours?" after it has been through the internet, perceptual fingerprinting wins - there's nothing in the file to remove. But it's also the least descriptive on its own. Resilience and richness are in tension, which is why you need both.

The bottom line

These are not competing options. They are layers with complementary weaknesses, and treating them as alternatives is how you end up with a single point of failure. Let the providers' watermarks do their part, build your own stack - manifest, external record, fingerprint - and the record of how the work was made survives most of what the real world does to files. Which is quite a lot.

Curious what this stack means for your own output? See which rules apply to your work.