Content Credentials (C2PA), watermarks and provenance

Video lecture

Content Credentials (C2PA), watermarks and provenance

13 chapters · about 9 min · full transcript

Coming soon

Chapter 1 of 13

Content Credentials and watermarks

From 'is it fake?' to 'where from?'
How C2PA works
Watermarks and fingerprints
Check your own files

The narrated lecture is in production

Every chapter is scripted and ready. Browse the chapters and read the full transcript now — the video will appear here when it’s published.

Chapters

Transcript of the narration, chapter by chapter.

0:00 Content Credentials and watermarks

Here's an uncomfortable truth. The best AI detectors of last year are already fooled by this year's models. If your plan for handling synthetic media is to look closely and spot the fake, that plan is running out of time. So the industry is asking a different question. Not is it fake, but where did it come from? In this lecture you'll learn how Content Credentials and the C2PA standard work, how watermarks fit in, what platforms do with them, and how to check your own files with a command-line tool and a short Python script.

0:42 Why it matters

Why should marketers care? Three reasons. Platforms read provenance to apply AI labels automatically, so it affects how your content appears. Regulators now expect generative tools to mark outputs: Article fifty, paragraph two of the EU AI Act, and California's AI Transparency Act, which became operative in August twenty twenty-six for large generative AI providers. And clients increasingly ask agencies to preserve provenance and prove it. If your export settings throw it away, you're throwing away a trust signal your client paid for.

1:18 The passport analogy

Here's the analogy. Think of a Content Credential like a passport with visa stamps. The passport is issued by a trusted authority and signed. Each border crossing adds a stamp. You can see where the traveler has been. But a passport doesn't tell you whether the traveler is a good person. It only tells you the journey is documented. That's C2PA. A signed record of who created a file, with what tool, and what happened to it since. It proves the history. It doesn't prove the content is true.

1:57 Inside a manifest

How does it actually work? A C2PA manifest has three parts. Assertions: statements like this was created, this was edited, and a digital source type. The key value is trained algorithmic media, which means a generative model made it. Then a claim that bundles the assertions with a cryptographic hash of the content, so if anyone changes a pixel, the hash no longer matches. And a signature from a certificate, which validators check against a trust list. Each edit in a C2PA-aware tool can add a new manifest that points to the old one as an ingredient, building a chain of history.

2:41 Three layers

Now the weakness. Metadata is fragile. A screenshot, a re-encode, or an upload to many platforms can strip it. That's where two other layers help. An invisible watermark, like Google's SynthID, is hidden in the pixels or audio and survives many edits, though usually only the vendor can detect it. And a fingerprint is a perceptual hash stored in a database, used to look up the original manifest after stripping. Put them together and you get what C2PA calls durable Content Credentials. Metadata for detail. Watermark or fingerprint so it can be recovered.

3:21 Two rules

Two rules to tattoo on your brain. First, no manifest proves nothing. Most files lose metadata somewhere, so absence isn't evidence that content is real, or fake. Second, a valid manifest proves who signed which claims, not that the content is true. A real camera can capture a staged scene. A signed AI image is still AI. Provenance is powerful evidence about history. It isn't a truth machine.

3:51 Example 1: Manchester photographer

First example, simple. A Manchester photographer shoots a product with a camera that signs captures, then edits in a C2PA-aware editor, and exports with Content Credentials switched on. When the client posts on a platform that reads provenance, the post can show that it was captured with a camera. When a competitor later claims the photos are AI fakes, the photographer opens the file in a verify tool and shows the signed capture and the edit history. Argument over in a minute.

4:27 Example 2 (illustrative): Dubai agency pipeline

Second example, a realistic business scenario with illustrative details. A Dubai agency produces AI-assisted product visuals for a GCC retailer that also sells into the EU. It sets a pipeline rule. Generate only in tools that attach Content Credentials. Edit in a C2PA-aware editor. Export with credentials on. Run a Python check on the delivery folder. Record results in the asset register. Illustratively, the first check found that one designer's export preset was stripping metadata from every file. Fixed in five minutes. And when a platform later auto-labeled a visual as AI-generated, the client wasn't surprised, because the register already said so.

5:11 Watch me do it: check exports

Watch me do it. I open a terminal in my exports folder. I run c2patool on the hero image. It prints the manifest as JSON. I look for the actions assertion and find created, with the digital source type trained algorithmic media. So the file declares AI generation. I check the validation section: no errors. Now I run it on the square crop for Instagram. It says no manifest found. That's my export preset stripping metadata. I change the setting, re-export, run it again, and the manifest is there. Finally I run the Python script from the lesson on the whole folder, and it prints one line per file: credentials true or false, and AI declared true or false.

6:03 Common mistakes

Common mistakes. Treating a missing manifest as proof something is real. Treating a valid manifest as proof something is true. Delivering assets by screenshot, which throws away all provenance. Stripping metadata to save file size, without realizing you've removed a trust signal. And assuming every platform keeps the data. Many read it on upload to apply a label and then serve a file without it. Your own copy, with credentials intact, is your evidence.

6:35 Signing your own content

Should an agency sign its own content? Sometimes, yes. If you publish original photography or official brand announcements that are likely to be faked, signing them with Content Credentials gives your audience a way to check what's genuine. Tools like Adobe's Content Authenticity app, and open-source libraries like the C2PA SDKs, can attach credentials. One caution. Files signed with a test certificate, the kind the command-line tool uses by default, will show as untrusted in verify tools. For public use, you need a certificate from a provider on the C2PA trust list, or you sign through a tool that already has one. Start small: your founder's official photos and major announcements.

7:23 Measure it

How do you measure success here? Track the share of delivered AI-assisted files that still carry valid Content Credentials. The target is all of them, unless a client explicitly asks otherwise. Track mismatches between your asset register, the manifest, and the platform label. The target is zero. And keep your vendor list current: for each generative tool, does it mark with metadata, a watermark, or both, and when did you last check? That list is also what an EU client will ask for under Article fifty.

8:00 Recap + try this now

Recap. Detection is a losing race; provenance is the durable answer. C2PA manifests are signed records of history, strengthened by watermarks and fingerprints. No manifest proves nothing, and a valid one proves history, not truth. Keep credentials through your pipeline and check your deliveries. Try this now: run c2patool or the Python checker on five of your recent AI-assisted exports, find any that lost their credentials, and fix the export setting responsible. Next, we look at AI and media rules in the UAE and Saudi Arabia.

From "is it fake?" to "where did it come from?"

Detecting AI content by looking at it is a losing game: models improve faster than detectors. The industry's answer is provenance: attaching a verifiable record of where a file came from and how it was changed. For marketers this matters three ways: platforms read provenance to apply AI labels automatically, regulators (the EU AI Act's Article 50(2), California's AI Transparency Act) expect generative tools to mark outputs, and clients increasingly ask agencies to preserve it.

The three layers of provenance

Layer	What it is	Strength	Weakness
Metadata manifest (C2PA Content Credentials)	A cryptographically signed record attached to the file: who made it, with what tool, what actions (created, edited, AI-generated)	Rich, verifiable, tamper-evident	Easily stripped by screenshots, re-encoding or some platforms
Invisible watermark	A signal hidden in the pixels or audio (for example Google's SynthID)	Survives many edits and re-uploads	Usually only the vendor can detect it; says less about history
Fingerprint	A perceptual hash stored in a database, used to look up a file's manifest	Recovers provenance after stripping	Needs a lookup service

The C2PA approach combines them into durable Content Credentials: a manifest for detail, plus a watermark or fingerprint that points back to it if the metadata is stripped.

How C2PA works, briefly

C2PA (the Coalition for Content Provenance and Authenticity) publishes an open technical standard. Its specification is on the 2.x series (version 2.4 was released in April 2026). A manifest contains:

Assertions: statements about the asset, such as actions taken (c2pa.created, c2pa.edited), the tool used, and the digital source type. The IPTC value trainedAlgorithmicMedia means the content was created by a generative model; compositeWithTrainedAlgorithmicMedia means AI elements were combined with other content.
A claim bundling those assertions, with a hard binding (a cryptographic hash of the content, so any change breaks the match).
A signature from a certificate. Validators check the certificate against a trust list.

Each edit in a C2PA-aware tool can add a new manifest that references the earlier ones as ingredients, building a chain of history.

Crucial limitation: a valid manifest proves who signed what claims, not that the content is true. A real photo can be staged; a signed AI image is still AI. And no manifest proves nothing: most files lose metadata somewhere along the way.

Who supports it (check current status)

Creation: many generative tools attach Content Credentials to outputs (Adobe Firefly and OpenAI's image generation, for example). Some cameras and phones sign captures, including models from Leica, Sony and Nikon and Google's Pixel 10 series; Samsung marks some AI-edited images.
Platforms: YouTube, Meta, TikTok and LinkedIn read C2PA signals in different ways (see Module 3). Many platforms still strip the metadata from the file they serve, even when they use it to apply a label.
Watermarks: Google's SynthID marks content from Google's models and has a detection portal; other vendors use their own schemes.

Hands-on 1: inspect a file from the command line

# c2patool: open-source CLI from the Content Authenticity Initiative (contentauth/c2pa-rs, cli folder).
c2patool campaign_hero.jpg              # manifest store as JSON, or an error if none is present
c2patool campaign_hero.jpg --detailed   # full detail including validation results
c2patool campaign_hero.jpg --info       # short summary of the file's manifest info

Or drag the file into a public Content Credentials verify tool (linked from contentcredentials.org) for a visual history.

Hands-on 2: check a folder of exports in Python

This script flags which exported assets still carry Content Credentials and whether any manifest declares AI generation. Use it in your delivery checklist before files go to a client or platform.

# pip install c2pa-python
import json
import sys
from pathlib import Path

import c2pa

AI_TYPES = ("trainedAlgorithmicMedia", "compositeWithTrainedAlgorithmicMedia")


def inspect(path: Path) -> dict:
    try:
        with c2pa.Reader(str(path)) as reader:
            store = json.loads(reader.json())
    except c2pa.C2paError.ManifestNotFound:
        return {"file": path.name, "credentials": False, "ai_declared": None}
    except c2pa.C2paError as err:
        return {"file": path.name, "credentials": "error", "detail": str(err)}

    text = json.dumps(store)
    return {
        "file": path.name,
        "credentials": True,
        "ai_declared": any(t in text for t in AI_TYPES),
        # newer SDKs report "validation_state" (e.g. Valid, Trusted, Invalid); older ones list "validation_status" issues
        "validation_state": store.get("validation_state"),
        "validation_issues": len(store.get("validation_status") or []),
    }


if __name__ == "__main__":
    folder = Path(sys.argv[1] if len(sys.argv) > 1 else "exports")
    for f in sorted(folder.iterdir()):
        if f.suffix.lower() in {".jpg", ".jpeg", ".png", ".webp", ".mp4", ".mov", ".wav", ".mp3"}:
            print(inspect(f))

What to do with the results:

credentials: False on a file that had them at generation: your export settings are stripping metadata. Turn on "include Content Credentials" or "keep metadata" in your editor and re-export.
ai_declared: True: make sure the asset register and the platform disclosure match.
validation_state of Invalid, or validation_issues greater than zero: the file changed after signing or the certificate is not trusted; investigate before relying on it.

Worked example: an agency's provenance-preserving pipeline

A Dubai agency produces AI-assisted product visuals for a GCC retailer that also sells into the EU. It sets a rule: generate in tools that attach Content Credentials, edit in a C2PA-aware editor, export with credentials on, run the Python check on the delivery folder, and record results in the asset register. When a platform later auto-labels one visual as AI-generated, the client is not surprised: the register already says so, and the agency can show the chain of edits.

Measuring success

Share of delivered AI-assisted files that still carry valid Content Credentials (target: all, unless a platform or client explicitly requires otherwise).
Mismatches between asset register, manifest and platform label (target: zero).
Vendor list up to date with each tool's marking method (metadata, watermark or both).

Pitfalls

Treating a missing manifest as proof that content is real.
Treating a valid manifest as proof that content is true.
Screenshotting assets for delivery, which throws away all provenance.
Stripping metadata "for file size" without realizing it removes the trust signal.

Provenance asks where content came from; it is more durable than trying to spot fakes by eye.
C2PA Content Credentials are signed manifests; watermarks and fingerprints help them survive stripping.
A valid manifest proves who signed which claims, not that content is true; a missing one proves nothing.
Keep credentials through your export pipeline and check deliveries with c2patool or c2pa-python.

Run c2patool or the Python checker on five of your recent AI-assisted exports, note which lost their Content Credentials, and fix the export setting responsible.

Enrol for free to save your progress

Reading is always free. Enrol to keep your place, take the final assessment and earn a verifiable certificate.

Content Credentials (C2PA), watermarks and provenance