In May 2026, Canon introduced something the news industry has wanted for years: a practical way to verify where a photograph came from. Its Authenticity Imaging System is built on the C2PA standard and designed to help news organizations preserve and verify image provenance from capture through editorial handling, distribution and publication.
The system also shows the limit of provenance. Canon can help prove where an image originated. It cannot show where the image goes next, who used it, or whether it shaped an AI-generated answer.
For photojournalism, origin covers a large part of the trust problem. For medical, legal, scientific and technical reference catalogues, it is only the starting point. The economic question comes later, when an AI system uses the content to answer a question, support a recommendation or generate value. That later stage is where content provenance and usage evidence diverge, and for reference publishers it is where prices are set.
In short
Content provenance records where a piece of content came from, who created it, and how it may have changed over time. It is attached to the asset and begins at creation.
Usage evidence records what happened after publication: which systems accessed the content, which entries were retrieved, which AI answers it shaped, how often, under which rights and in what context.
The two are complementary, and a serious AI content strategy needs both. Provenance protects trust in the asset, while usage evidence supports its commercial value once it is used.
At Citations, we use the term usage evidence to describe rights-aware records that make AI content use visible, attributable and commercially actionable. For AI licensing, pricing, renewal and audit depend on the second record, which is why AI usage evidence for publishers has to extend beyond provenance into retrieval, answer generation, licensing and pricing.
What Canon actually built
Canon’s Authenticity Imaging System is a C2PA-compliant solution for supported Canon cameras, including the EOS R1 and EOS R5 Mark II. Provenance data is embedded in the camera when the image is captured. At the service level, Canon manages certificates, applies trusted timestamps and supports verification of image history across professional editorial workflows.
The managed verification layer is the significant part. A signature at capture is useful, but the newsroom, agency, publisher or platform receiving the image needs to verify its history without relying on trust alone. Canon’s system preserves provenance information from capture onward, in line with professional editorial workflows, and that addresses the first half of the problem.
Where provenance stops being enough
Canon’s system answers one question well: whether an image came from the claimed source and whether its history has remained verifiable. That is provenance, and it protects an asset against manipulation, forgery and loss of trust. For a news image, it can be enough to preserve editorial credibility.
Provenance does not cover what happens after publication. Once content enters the open web, a licensed database, a retrieval system or an AI workflow, it can be read, embedded, summarized, retrieved, reformulated and used to generate an answer. The source may still be essential to that answer and yet invisible, with no citation, link, click, audit trail or pricing signal.
Provenance works like a birth certificate: it establishes origin and says nothing about what the content does afterward. For images, that limitation may be tolerable. For reference catalogues, it is the central problem, which is why machine-readable provenance for reference publishers has to connect to usage records. Without that connection, publishers can prove authenticity but not value.
Content provenance vs usage evidence
The two records differ on each point that matters commercially.
Question | Content provenance | Usage evidence |
|---|---|---|
What does it prove? | Origin, authenticity and content history | Downstream use, context and commercial value |
When does it begin? | At creation or publication | At access, retrieval or AI interaction |
What is it attached to? | The asset | The usage event |
What does it support? | Trust, verification, anti-forgery, editorial integrity | Licensing, pricing, audit, renewal, contributor compensation |
Main limitation | It does not show downstream value | It requires instrumentation at the point of use |
Provenance protects the asset a publisher owns, and usage evidence is what the publisher brings to a negotiation. When a publisher licenses content to an AI company, provenance helps establish that the catalogue is authentic, authoritative and trustworthy. The price, however, depends on usage: how often the catalogue is consulted, which entries are most valuable, which answers depend on it, which domains generate the most demand, and which parts of the catalogue do the economic work.
A negotiation based only on provenance treats the catalogue as a static asset. A negotiation based on usage evidence treats it as a source of recurring revenue, and that shifts the balance between the parties. A publisher cannot price a use it cannot observe.
Why reference catalogues are the hard case
A news photo is captured at a precise moment, with a device, a file and a clear point of origin. A reference catalogue is built differently. A medical reference may be written, reviewed, updated and validated over months or years. A legal commentary may combine statute, doctrine, case law and expert interpretation. A scientific corpus evolves through papers, datasets, updates, corrections and citations.
Canon’s model works best when there is a controlled point of capture, and reference content rarely has one. The more important issue, though, is downstream usage. A medical entry creates value when it informs a clinical answer, a legal commentary when it shapes reasoning about a statute, and a scientific paragraph when it grounds a generated explanation. That value appears each time the content is used, not at publication.
If that usage is invisible, the most valuable part of the catalogue becomes the hardest to measure. Reference publishers may be able to prove that their content is authoritative without being able to show how much value it creates inside AI systems, and that leaves them in a weak negotiating position.
The collapse of click-based value
For years, publishers relied on page views, search traffic, referral links, downloads, citations and subscriptions to measure demand. When an AI answer replaces a click, that analytics layer stops working. A user may receive the answer without visiting the publisher, a model may retrieve the relevant passage without exposing the source, and a platform may generate value from the catalogue without producing any of the signals publishers used to track.
This is a gap in commercial infrastructure, and usage evidence is what can fill it: records captured at the point of AI interaction, showing which content was accessed, which passages were retrieved, which answer used them and in which context value was created. Traffic analytics after publication, attribution after the fact and self-reported platform data cannot provide that on their own. For the same reason, usage reporting in AI licensing deals should be judged by whether it produces usable evidence or only activity summaries.
What usage evidence looks like in practice
Usage evidence has to be captured when a system interacts with the content, at the moment of use, instead of being inferred months later or summarized in general reporting. For a reference publisher, it should make it possible to establish:
which catalogue entries AI systems use most often;
which types of queries trigger retrieval from the catalogue;
which outputs rely on licensed reference material;
which parts of the catalogue are underused, overused or strategically critical;
which use cases justify premium pricing;
which rights governed the interaction;
whether attribution was required, and whether it was shown;
whether the record can support renewal, audit or contributor compensation.
The aim is to make trusted content observable enough to govern, attribute and price, without turning every sentence into a separate transaction. Flat-rate licensing tends to persist while use is invisible, and once use becomes measurable, the terms can be renegotiated on that basis.
Why this is not just a copyright issue
Many publishers still frame AI content use as a copyright problem, which is too narrow. Copyright asks whether content was used lawfully. Usage evidence asks how much value that use created, which is the commercial question.
Without usage evidence, licensing negotiations rely on assumptions. The publisher says the catalogue is valuable, the AI platform says it is one input among many, and the side without evidence usually concedes on price. For reference publishers, this is a real risk, because a trusted medical, legal, scientific or technical catalogue reduces uncertainty, improves answer quality, lowers risk and supports decisions. That value should not disappear into an AI system without a record.
The same reasoning applies to AI citation integrity in medical publishing, which cannot be reduced to correct source formatting. In high-stakes domains, what matters is whether source use can be verified.
The EU AI Act reinforces the documentation environment
General-purpose AI providers face transparency and documentation obligations under the EU AI Act, including summaries of training content. This does not turn publishers into AI providers, but it changes the documentation environment around AI licensing.
A public training-data summary may describe broad categories of content used in model development. It does not necessarily show whether a specific publisher asset was used in a specific AI interaction, under a specific licensing right, with a specific commercial consequence. Transparency at that level is useful without being transaction-level usage evidence, and EU AI Act content licensing comes back to the same operational question: whether publishers can show what happened after access was granted.
A provenance record establishes origin, a transparency summary provides broad disclosure, and a usage record supports licensing value. The same logic is reaching transparency law itself. Static disclosure rules risk becoming outdated with each new model generation, and a more durable requirement is a record of what a system did with protected content. That is the case for transparency law that regulates evidence surviving model change.
The objection: provenance should be enough for trust
If provenance proves that a source is authentic, it may seem sufficient. For some trust problems it is: where the main risk is manipulated media, forged images or false origin, provenance is essential.
AI licensing adds a second problem. An authentic source can still be used invisibly, a trusted catalogue can still generate value without attribution, and a verified work can still become an input nobody pays for. A provenance record can state that the content is real. It cannot state that the content was used in a given AI interaction, under a given right, with a given value. For AI licensing, the commercial question begins after a trusted source has been used.
The next phase of content licensing
Canon’s system shows that the market is building infrastructure for verifiable trust. The next phase concerns proving that content was used, in addition to proving that it is authentic. AI licensing will depend on three layers:
Layer | Core question | Commercial function |
|---|---|---|
Authenticity | Can the source be trusted? | Protects credibility |
Access | Was the system allowed to use the content? | Defines permission |
Usage | Did the content contribute to an answer, recommendation or product? | Supports pricing, renewal and audit |
Most of today’s debate focuses on the first two layers, while pricing will be decided on the third. This is especially true for reference catalogues, whose value comes from repeated use in high-value contexts rather than from publication alone. A catalogue used this way works as infrastructure, and infrastructure needs metering.
Where Citations fits
Citations is built for the layer after provenance: making the AI use of reference catalogues observable, governable and priceable. Each relevant AI use can become a record of which content was accessed, which passage was retrieved, which right applied, which AI workflow used it, whether attribution was required and given, and what commercial value the event can support.
With that record, a publisher can show where its catalogue created value, in addition to showing that it is authentic. Reference publishers need both the trust infrastructure and the licensing infrastructure.
The question reference publishers should ask now
Canon has made the origin of images easier to verify. For reference publishers, the strategic problem starts after origin: after publication, indexing, retrieval, AI processing and answer generation, once the source is no longer visible. The next licensing disputes will turn on usage evidence as much as on provenance or attribution.
Most reference publishers can show where their content came from. Fewer can show where it went after an AI system read it, and that is the gap to assess first.
Frequently asked questions
What is the difference between content provenance and usage evidence?
Content provenance records where a piece of content came from, who created it and how it may have changed over time. Usage evidence records what happened after publication: which systems accessed the content, which AI answers used it and how often.
Provenance protects the asset. Usage evidence supports licensing and pricing.
What is C2PA?
C2PA, the Coalition for Content Provenance and Authenticity, is an open technical standard for establishing the origin and edit history of digital content. It is used to create tamper-evident provenance records that can travel with an asset.
Why is provenance not enough for AI licensing?
Provenance proves that content is authentic. It does not show how often that content is used, which AI outputs rely on it, or which parts of a catalogue create the most value.
AI licensing needs usage evidence because pricing depends on use.
Why does this matter for medical, legal and scientific publishers?
Medical, legal and scientific catalogues create value when they inform answers, recommendations, interpretations and decisions. That value often arises downstream, inside AI systems.
Without usage evidence, publishers struggle to prove and price that value.
How does usage evidence change licensing negotiations?
Usage evidence turns licensing from a flat discussion about access into a measurable discussion about value. It helps publishers identify which content is used, how often, in which contexts and with what commercial importance.
Continue the evidence chain
Start with the broader framework: AI Usage Evidence for Publishers
Make authority machine-operable: Machine-Readable Provenance for Reference Publishers
Test who audits the number: Usage Reporting in AI Licensing Deals
See why disclosure rules age badly: AI Transparency Law Should Regulate Evidence That Survives Model Change
Verify the source before the citation: AI Citation Integrity in Medical Publishing
Separate disclosure from audit trail: AI Disclosure Is Not an Audit Trail for STM Publishers
Understand the regulatory frame: EU AI Act and Content Licensing
Ready to assess your evidence layer? Book an AI usage evidence assessment
Sources
Canon Global — “Canon Introduces C2PA-Compliant Authenticity Imaging System for News Organizations”
Canon Introduces C2PA-Compliant Authenticity Imaging System for News Organizations
Canon UK — “Canon introduces C2PA compliant Authenticity Imaging System for news organisations”
Canon introduces C2PA compliant Authenticity Imaging System for news organisations
C2PA — Technical Specification
C2PA Technical Specification
C2PA — Content Credentials 2.3 announcement
C2PA launches Content Credentials 2.3
Content Credentials — C2PA provenance information for digital content
Content Credentials
European Commission — Template for GPAI model providers to summarize training content
Template for GPAI model providers to summarize training content