Scholarly, medical, legal, technical and reference publishers have spent decades building authority through peer review, editorial control, versioning, corrections, retractions, attribution, citation and rights management. These processes are what readers rely on when they trust a source.
Generative AI changes the setting in which that trust has to work. The question is no longer only whether a publisher’s content is accurate and valuable to human readers. It is whether that authority survives once the content enters machine workflows.
In short
Machine-readable provenance is the structured encoding of a content object’s origin, version, rights, attribution and status in a form that AI systems can retrieve, preserve, display and audit.
For reference publishers, much of the value sits in the signals around the text: peer review, the Version of Record, corrections, retractions, citation rules, attribution requirements and rights conditions. Those signals do not survive inside AI workflows by default. A model can retrieve a fragment without knowing whether it came from the authoritative version, summarize a claim without surfacing a correction, or cite a source while losing the attribution chain behind it.
At Citations, we use the term usage evidence to describe rights-aware records that make AI content use visible, attributable and commercially actionable. That evidence depends on an earlier step: the catalogue must be able to prove what each object is, which version is authoritative, what rights apply, and whether the source was still trustworthy at the point of use.
From content authority to machine authority
In March 2026, STM published Toward Responsible Use of Research Content in Generative AI, a discussion document on how research content should be handled in GenAI tools. Its core concern is that research content carries properties that general AI systems were not designed to preserve: peer review, the Version of Record, corrections, retractions, attribution, citation, provenance and verifiability.
For reference publishers, this is a business issue as much as a policy one. Whether content was reviewed, which version is authoritative, whether it has been corrected or retracted, what rights apply, who should be credited and how it should be cited are routine questions in publishing. AI systems do not always ask them.
For this reason, AI usage evidence for publishers starts with identity, rights and source status. A system that cannot identify the authoritative source cannot produce a meaningful measure of its use.
What breaks when authoritative content enters AI workflows
When a generative model ingests, retrieves, summarizes or displays trusted content, the answer can reach the user without the signals that made the content reliable. The model may retrieve a fragment without knowing whether it came from the Version of Record, summarize a claim without the published correction, or repeat an outdated statement without detecting that the record has changed. It can also cite a source while losing the attribution chain behind it, or mix validated research with preprints, commentary, outdated records and low-quality web text, and present the result with the same confidence.
There is evidence of this. C&EN reported on research testing whether GPT-4o-mini recognized retracted or otherwise problematic scholarly papers. The analysis examined 217 papers, each evaluated multiple times, and found that the system did not reliably surface retraction or validity concerns.
In such cases the content is still used, but its status is lost. Publishers then lose more than referral traffic: they lose control over how their content is understood, attributed, licensed and trusted.
What is machine-readable provenance?
Machine-readable provenance is the encoding of a content object’s origin, version, rights, attribution and status in a form that AI systems can retrieve, preserve, display and audit. It goes further than web-page metadata, a DOI or a citation. It allows a content object to establish, inside a machine workflow:
what it is;
where it came from;
which version is authoritative;
whether the record has changed;
what rights apply;
who should be credited;
how it may be used;
how it was used.
In AI workflows, authority has to be encoded, retrieved, preserved, displayed and audited. If those signals are missing, the catalogue still exists on the web but risks dropping out of the machine workflows where discovery, research assistance, professional decision support and knowledge work increasingly take place.
Metadata and machine-readable provenance
Machine-readable provenance is different from having metadata. Many reference publishers already have rich metadata. The question is whether that metadata survives inside AI workflows and stays connected to usage.
Question | Traditional metadata | Machine-readable provenance |
|---|---|---|
What does it describe? | The content object | The content object, its status, rights and trust conditions |
Where does it operate? | Publisher platforms, catalogues, indexes | AI retrieval, generation, display and audit workflows |
What does it support? | Discovery and management | Trust preservation, rights control, attribution and usage evidence |
What is the risk? | Metadata stays outside the AI workflow | Authority travels with the content |
Commercial value | Helps content be found | Helps content be trusted, governed and licensed |
Metadata helps a catalogue get discovered, and provenance keeps it authoritative after discovery. The article on content provenance vs usage evidence develops the next distinction: provenance establishes what the source is, usage evidence records what happened to it, and reference publishers need both.
Why attribution disappears in AI retrieval
Generative AI changes the unit of value. In traditional publishing, value attaches to the article, chapter, entry, database, journal, platform or subscription. In AI environments, it may surface as a paragraph, a claim, a citation, a generated summary, a recommendation or a professional answer.
The publisher’s investment is used upstream, while the user sees only the downstream answer. The source and the editorial work behind it can matter a great deal and stay hidden, and the rights that apply become harder to enforce. AI systems can draw on trusted content while dropping the signals that explain why it should be trusted. A catalogue that cannot carry attribution, versioning and rights into the AI workflow may still be used, but that use will be harder to count, price and defend.
What a reference catalogue must be able to prove
A reference publisher’s catalogue has to answer the following questions in a form machines can read.
What each content object is
AI systems need to distinguish an article, chapter, entry, dataset, abstract, correction notice, retraction notice, preprint, commentary and Version of Record. A system that cannot identify the object cannot preserve its authority.
Which version is authoritative
In scholarly, medical, legal and technical publishing, using the wrong version creates real risk. A machine workflow must be able to tell a stale copy from the current Version of Record.
Whether the record has changed
Corrections and retractions need to be exposed in machine-readable form. An AI system that cannot detect that a claim has been corrected or withdrawn can repeat it with confidence, and in high-stakes domains that can cause harm.
What rights apply
Rights must be expressed in a form machines can interpret. Training, retrieval, summarization, display, citation, redistribution and professional use are different uses, and the rights infrastructure has to support those distinctions.
How the content was used
Licensing depends on knowing whether the content was used in training, retrieved at inference time, summarized, quoted, cited or used to support a generated recommendation. Without that evidence, publishers negotiate from a weak position.
Whether attribution survived
Attribution has to travel through retrieval, generation, display and audit instead of stopping at ingestion. If it disappears before the final user experience, the publisher’s contribution becomes invisible and harder to defend.
The same evidence chain underlies proof of AI content usage. A publisher cannot prove valuable use if the content object, version, rights and attribution state are not preserved.
Corrections and retractions are the stress test
Corrections and retractions show whether machine-readable provenance works in practice. Crossref’s Crossmark service gives readers the current status of a content item, including corrections, retractions and updates, and Crossref provides guidance for reflecting retraction status in metadata and on landing pages.
AI workflows add a further requirement. A retraction notice on a landing page helps a human reader, but a machine workflow needs the status signal before it summarizes, recommends or cites the content. The record status therefore has to become part of the content’s machine-readable identity instead of staying at the edge of the publisher’s platform, or withdrawn claims will keep being repeated.
The same applies to AI citation integrity in medical publishing, which depends on more than citation formatting. A correctly formatted citation can still point to stale or retracted content.
Where authority becomes fragile in AI systems
Authority weakens inside AI systems in several distinct ways:
if a system cannot reliably identify the authoritative source, it may rely on a weaker substitute;
if the publisher’s role does not appear in the answer, the editorial investment behind it is hidden;
if the system cannot distinguish an outdated record from the current Version of Record, the answer may rest on the wrong version;
if corrections and retractions are not reflected, stale information can mislead professionals at the point of decision;
if publishers cannot define, monitor and evidence how their content is used, licensing becomes harder to defend.
Catalogue size alone will not decide which publishers do well. Their position will depend on whether their catalogues can establish their authority inside AI systems.
AI content licensing depends on provenance
The debate about AI and publishing is often reduced to copyright. Copyright sets the legal frame, while the practical difficulty is content governance in machine environments. A publisher cannot build a strong AI licensing model if it cannot define what is being licensed, how it may be used, how it must be attributed and how compliance can be evidenced.
A license without machine-readable provenance is weak, and one without usage evidence or attribution persistence is harder still to enforce. AI systems reorganize value as well as consuming content: the publisher’s content may support the answer, its editorial process may improve the output, its citation may justify a recommendation and its database may feed the workflow. If the publisher’s role is invisible, that value is harder to measure, negotiate and monetize. The commercial risk is less disappearance from the web than disappearance from the workflow.
The EU AI Act reinforces the documentation environment
The EU AI Act does not make publishers providers of general-purpose AI models, but it changes the documentation environment around AI content licensing. General-purpose AI providers face transparency and documentation obligations, including summaries of training content, and that pressure will shape the market for licensing, rights and evidence.
Publishers can treat this as a documentation question for the whole content supply chain. They will need to define what was licensed, express rights in machine-readable form, identify the authoritative version, preserve attribution and evidence usage later. This is the commercial side of EU AI Act content licensing: the regulatory burden sits mainly with AI providers, while the negotiating advantage goes to the parties that can produce credible records.
The objection: publishers already have authority
Reference publishers already have authority. Their brands are known, their journals, databases and catalogues are trusted, and their editorial processes are mature.
AI workflows, however, do not automatically preserve institutional authority. A trusted source can be retrieved without attribution, a corrected article summarized without the correction, a retracted paper treated as valid, a licensed version confused with an outdated copy, and a citation kept while the source status disappears. The question is whether that authority remains usable by machines once the content leaves the publisher’s platform.
What a provenance audit should test
Reference publishers can start with a practical machine-readiness audit. It should test whether:
each content object can be uniquely identified;
the authoritative version can be distinguished;
corrections and retractions are machine-readable;
rights are expressed in a way AI systems can interpret;
attribution requirements travel with the content;
source status can be checked at the point of AI use;
retrieval and usage events can be logged;
usage evidence can support licensing, audit and renewal;
downstream users can distinguish trusted records from stale copies;
the publisher can show what happened after access was granted.
The audit defines what the publisher can credibly license and shows where its catalogue’s authority weakens inside AI systems.
Where Citations fits
Citations helps reference publishers make provenance, attribution, rights, versioning and usage evidence operational inside AI systems. For each use, the aim is to establish which source and version were involved, which rights and attribution requirements applied, which correction or retraction status mattered, which usage event occurred, and which record can support audit, pricing and renewal.
Publishers need this layer to build AI licensing models, because a license is only as strong as their ability to show what was licensed, how it was used, how it should be attributed and whether compliance can be verified.
The question every reference publisher should ask now
For STM, medical, legal, technical and reference publishers, the practical test is whether the catalogue can show what each object is, where it came from, what rights apply, which version is authoritative, whether the record has changed, and how it was used inside an AI system. A publisher that cannot answer has a strategic gap as well as a technical one, because in AI-mediated discovery, trust depends on signals machines can read as well as on reputation.
Frequently asked questions
What is machine-readable provenance?
Machine-readable provenance is the structured encoding of a content object’s origin, version, rights, attribution and status so AI systems can retrieve, preserve, display and audit those signals.
Why does machine-readable provenance matter for reference publishers?
Reference publishing value depends on trust signals. If AI systems retrieve content without preserving versioning, attribution, corrections, retractions and rights, the publisher’s authority becomes invisible inside the workflow.
Why do corrections and retractions break inside AI systems?
Many AI systems retrieve and summarize text without reliably detecting whether a record has been corrected, updated or retracted. Without machine-readable correction and retraction signals, outdated or withdrawn claims can be repeated with confidence.
Is this only a copyright issue?
No. Copyright sets the legal frame. The operational issue is governance: defining what is licensed, how it may be used, how it must be attributed, and how compliance can be evidenced inside AI systems.
How does provenance support AI content licensing?
Provenance helps publishers prove what content is being licensed, which version is authoritative, what rights apply, whether attribution should persist, and whether the content remained trustworthy at the point of use.
What should a reference publisher do first?
Start with a provenance audit. Test whether your catalogue can prove version, rights, attribution, corrections, retractions and usage in machine-readable form. That audit defines what you can credibly license to AI platforms.
Continue the evidence chain
AI Usage Evidence for Publishers
Content Provenance vs Usage Evidence
AI Citation Integrity in Medical Publishing
EU AI Act and Content Licensing
Book an AI usage evidence assessment
Sources
STM Association — “Toward Responsible Use of Research Content in Generative AI”
Toward Responsible Use of Research Content in Generative AI
STM Association — Consultation page for “Toward Responsible Use of Research Content in Generative AI”
Consultation page
C&EN — “ChatGPT tends to ignore retractions on scientific papers”
ChatGPT tends to ignore retractions on scientific papers
Learned Publishing — “Does ChatGPT Ignore Article Retractions and Other Reliability Concerns?”
Does ChatGPT Ignore Article Retractions and Other Reliability Concerns?
Crossref — Crossmark
Crossmark
Crossref — Version control, corrections and retractions
Version control, corrections and retractions
NISO — “Workshops Highlight Standards Needed for Tracking Provenance, Attribution, and Usage in the AI Ecosystem”
Workshops Highlight Standards Needed for Tracking Provenance, Attribution, and Usage
The Scholarly Kitchen — “Attribution, Provenance, Reference, Citation, and AI for Research Applications: Understanding the Differences”
Attribution, Provenance, Reference, Citation, and AI for Research Applications
European Commission — General-purpose AI obligations under the AI Act
General-purpose AI obligations under the AI Act