The internet is approaching an uncomfortable point: generating convincing text, images, audio and video is becoming cheaper than verifying where any of it came from.
The obvious response has been AI detection.
Look at a piece of content and ask whether a model produced it.
That approach is inherently fragile. Generation techniques change. Detectors make errors. Human-edited AI content sits between categories. Compression and screenshots remove clues.
A more durable idea is provenance.
Instead of guessing what produced a file from its appearance, record its history.
From detection to receipts
Content provenance asks different questions:
Who or what created this asset? Which tool captured or generated it? What edits were made? Can the history be verified cryptographically?
The Coalition for Content Provenance and Authenticity, or C2PA, develops an open technical standard for attaching this information to digital media through Content Credentials.
A credential can contain signed assertions describing origin and transformations.
Conceptually, it is closer to a chain of custody than an AI detector.
Why provenance is attractive
Detection tries to infer the past from the final output.
Provenance records the past while it happens.
That has several advantages.
A camera can record that an image originated from a capture device. Editing software can append transformations. A generative system can record that an image was created algorithmically.
Each participant adds verifiable information to the history.
The user does not need a classifier to guess.
What Content Credentials can prove
A valid cryptographically signed credential can provide evidence that particular assertions were made by identified software, devices or organisations and that the manifest has not been silently altered.
That can be valuable.
But provenance does not prove that the depicted event is true.
A genuine photograph can be staged. A real recording can be taken out of context. A truthful provenance chain can document the creation of propaganda.
Content Credentials are evidence about origin and editing history, not a universal truth machine.
Missing credentials do not mean fake
This is one of the most important design principles.
If provenance systems become common, users may start treating a missing credential as evidence that content is synthetic.
That would be dangerous.
Legacy cameras, unsupported tools, stripped metadata, screenshots and many distribution platforms can all produce content without credentials.
The correct interpretation is asymmetric:
A valid credential can add useful information. No credential means the history is unknown.
Unknown is not fake.
Screenshots are a hard problem
Digital content is constantly transformed.
A signed image can be screenshotted. A video can be re-encoded. A social platform can strip metadata. A small section can be cropped out and republished.
C2PA's evolving work includes mechanisms intended to improve resilience and reconnect content to provenance after transformations, but no system can guarantee perfect persistence across every medium.
This is why provenance needs ecosystem adoption rather than one app.
Identity is the next layer
A credential becomes more useful when the signer matters.
"Created by a verified news organisation" conveys more information than "created by unknown software".
Future provenance systems therefore intersect with digital identity.
The challenge is to support meaningful identity without forcing every creator into a centralised registry or exposing unnecessary personal information.
Different contexts need different levels of identity:
- a photojournalist;
- a government agency;
- an anonymous dissident;
- an AI model provider;
- an advertising studio;
- an ordinary phone user.
One identity design will not fit all of them.
The economic problem
Standards succeed when incentives align.
Creators need tools that preserve credentials automatically. Platforms need reasons to display them. Users need interfaces that make the information understandable. Publishers need workflows that do not create unreasonable friction.
If adding provenance requires extra manual work while removing it is effortless, adoption will remain patchy.
The strongest path is invisible infrastructure: credentials attached and preserved by default, with users seeing them only when needed.
News and evidence may adopt first
Some contexts have unusually high value for chain-of-custody information:
- journalism;
- conflict documentation;
- insurance claims;
- legal evidence;
- public statements;
- political advertising;
- scientific imagery.
In these domains, provenance can become part of professional workflow rather than a consumer feature.
That may create trust from the top down.
Generative AI makes authentic capture more valuable
A surprising effect of synthetic media is that verified non-synthetic capture may become a premium category.
If anyone can generate a plausible image, evidence that an image came directly from a particular device at a particular time becomes more valuable.
The future media economy could therefore have two parallel signals:
content quality and content provenance.
The first asks whether something is compelling. The second asks whether its origin can be traced.
Three futures for the synthetic web
Provenance becomes normal
Phones, creative tools, AI systems and major platforms adopt interoperable credentials. Users can inspect media history when authenticity matters.
Fragmented standards
Different ecosystems use incompatible provenance systems, limiting network effects and confusing users.
Detection remains dominant
Platforms continue relying mainly on classifiers and watermarks because provenance adoption is incomplete. Authenticity remains probabilistic.
The first is technically plausible but requires coordination.
What to watch
Signals of real adoption include:
- cameras signing at capture;
- major platforms preserving credentials through upload and transformation;
- generative systems marking outputs by default;
- browsers displaying provenance consistently;
- newsrooms adopting credentials in ordinary workflow;
- credentials surviving common editing and sharing paths.
Conclusion
The AI-generated internet does not need a magical detector that can look at any file and announce truth.
It needs better evidence about origin.
Content provenance cannot tell us what to believe. It can tell us more about how a piece of media came to exist.
In an environment where appearances are increasingly cheap to manufacture, that history becomes valuable.
The future of digital trust may depend less on spotting the fake and more on preserving the receipt.
Sources
- C2PA — Specifications 2.4
- C2PA — New implementation guide for Content Credentials
- C2PA — Implementation guidance
Frequently asked questions
What is content provenance?
Content provenance records evidence about where digital media came from and how it changed, rather than trying to infer origin only from the final appearance.
Do Content Credentials prove that media is true?
No. They can provide evidence about origin and editing history, but authentic media can still be staged, miscaptioned or misleading.
Does missing provenance mean content is fake?
No. Missing credentials mean the provenance is unknown; legacy tools, screenshots and unsupported platforms can all strip or omit credentials.