AI Publishers’ Missing Proof: Why Legal Presumptions Need Usage Records
The Darcos bill exposed the right problem.
It also exposed the missing layer.
A legal presumption can shift who must prove AI content use. It can change the litigation posture. It can make the burden of proof less brutal for publishers, authors and rights holders.
But it cannot, by itself, produce the record that makes proof possible.
That is the deeper lesson.
The next phase of AI copyright will not be decided only by legal principles. It will be decided by evidence systems.
Rights need records.
Licensing needs measurement.
Compensation needs usage evidence.
In short
The Darcos bill proposed to create a presumption of use of cultural content by AI providers when credible indicators make that use plausible. In practice, this would shift part of the burden of proof: the rights holder would no longer have to prove everything from the outside, while the AI provider could be required to rebut the presumption.
That is a meaningful legal move.
But it does not remove the need for evidence.
A presumption still needs indicators to be triggered. A rebuttal still needs records to be credible. A licensing negotiation still needs measurable usage to support pricing, audit and contributor compensation.
At Citations Logic, we use the term usage evidence to describe rights-aware records that make AI content use visible, attributable and commercially actionable.
A law can decide who must prove.
It cannot create the operational record.
That is the missing layer for publishers: not just stronger rights, but records that show which content was accessed, under what conditions, by which system, and with what evidentiary value.
No record, no proof.
No proof, no enforceable market.
What the Darcos bill tried to fix
On April 8, 2026, the French Senate adopted a bill on the presumption of use of cultural content by AI providers.
The bill was then transmitted to the French National Assembly on April 9. A committee report was filed on June 2. The text was expected to be examined in public session on June 11.
Then the political process stalled.
The bill was not rejected after a full debate. It was pushed to the end of a crowded agenda, slowed by amendments, and ultimately not examined that day.
That procedural outcome matters less than the problem the bill exposed.
The Darcos bill tried to address one of the hardest asymmetries in AI copyright.
Publishers and rights holders do not control AI training data. They do not control model internals. They do not control retrieval logs, generated outputs at scale, or the operational records of AI providers.
Yet they are often expected to prove that their content was used.
That is the impossible demand.
The publisher sees the harm.
The platform controls the evidence.
That is why AI usage evidence for publishers is becoming a central commercial and legal issue.
Presumption is not proof
The key idea in the Darcos bill was simple: if credible indicators make the use of protected content plausible, the AI provider should have to prove otherwise.
The text adopted by the Senate introduced a presumption “sauf preuve contraire” — unless proven otherwise — when an indicator related to the development, deployment or output of an AI system makes the use of a protected work plausible.
That is powerful.
But it is not magic.
A rebuttable presumption does not eliminate proof.
It relocates the proof problem.
The publisher still needs enough evidence to trigger the presumption.
The AI provider then needs evidence to rebut it.
The court needs evidence to evaluate both sides.
So the law does not remove the operational question. It makes it unavoidable.
What counts as an indicator?
What records should the AI provider keep?
What can the publisher show?
How can a specific work be connected to a specific system, output, dataset, retrieval event or use case?
A legal presumption can change the pressure.
It cannot manufacture the trail.
That is where proof of AI content usage becomes the missing commercial asset.
The Mistral levy proposal has the same blind spot
The debate around Mistral AI points to the same structural gap from another direction.
Arthur Mensch argued for a mandatory contribution on commercial AI providers operating in Europe, based on revenue, to help fund the cultural sector. Public reporting described the proposal as a contribution of between 1% and 5% of revenue from AI model providers in Europe.
That idea responds to a real problem.
AI systems may benefit from large volumes of cultural, journalistic, scientific and publishing content. If that content contributes to AI products, the creative and publishing ecosystem should not be economically erased.
But a levy has a limit.
A levy may redistribute value.
It does not show where value came from.
It does not show which catalog was used.
It does not distinguish between a news archive, a legal commentary, a medical reference work, a textbook, a scientific journal or a cultural image library.
It may compensate the system broadly.
It does not create usage evidence.
That is the same blind spot as the presumption debate.
The market is asking the right question: how should value return to content owners?
But the harder question comes first: how do we know what was used?
The AI Act increases transparency, not transaction-level evidence
The EU AI Act adds another layer.
Under Article 53, providers of general-purpose AI models must publish a sufficiently detailed summary of the content used for training. The European Commission has also published a template for those summaries.
That is important.
It creates a baseline of transparency.
It helps rights holders understand broad categories of training data. It gives regulators a stronger disclosure framework. It may help publishers assess whether certain modalities, sources or types of content were used.
But it is not the same as usage evidence.
A training data summary can say what kind of content was used.
It does not necessarily show whether a specific work was used by a specific model in a specific way.
It does not necessarily connect use to a commercial answer, a professional workflow, a paid feature, a retrieval event or a licensing term.
It does not necessarily answer the question that matters in a dispute:
Was this content used?
Nor the question that matters at renewal:
How much value did this content create?
That distinction matters.
The AI Act pushes disclosure forward.
But disclosure is not attribution.
And attribution is not yet transaction-level evidence.
That is why EU AI Act content licensing still needs a measurement layer.
Legal tools, market tools and evidence tools are different
Publishers should not confuse three categories.
Tool | What it does | What it does not do |
|---|---|---|
Legal presumption | Shifts the burden of proof when use appears plausible | Does not create the factual record |
Levy or contribution | Redistributes value across a sector | Does not attribute value to specific content use |
Transparency summary | Discloses broad training data information | Does not prove transaction-level usage |
Usage evidence | Connects content, rights, access and use records | Requires operational instrumentation |
This distinction is not academic.
It decides whether AI copyright becomes enforceable.
A right without records is fragile.
A licensing deal without measurement is blunt.
A compensation system without usage data risks becoming political allocation.
Publishers do not only need to be recognized.
They need to be counted.
Why compensation without proof remains fragile
The publishing sector is not only asking for money.
It is asking for leverage.
A statutory contribution may help. A collective fund may help. A legal presumption may help. Transparency summaries may help.
But none of those mechanisms solve the core pricing problem alone.
A breaking news article, a legal database, a scientific journal, a medical reference library, an educational textbook and a financial research archive do not have the same AI value.
They do not carry the same authority.
They do not serve the same use cases.
They should not necessarily be licensed, measured or compensated in the same way.
A flat redistribution mechanism may be politically convenient.
It may also flatten value.
That becomes a problem for high-authority catalogs.
If a specialist source improves the quality of professional AI answers, that usage should not disappear into a generic pool.
If a legal commentary grounds a response in an enterprise workflow, that event matters.
If a medical reference source supports a clinical information answer, that use carries a different evidentiary and commercial weight.
Compensation without proof is better than invisibility.
But it remains fragile.
Usage evidence is what allows compensation to move from assumption to allocation.
This is why AI copyright collective management distribution keys will need evidence inputs, not only policy formulas.
No traceability, no market
The current AI-content economy is stuck in a contradiction.
Content feeds AI systems.
Answers are generated elsewhere.
Value flows through products, subscriptions, APIs, copilots, enterprise tools and user interfaces that may never send traffic back to the original source.
The publisher sees loss of visibility.
The AI provider sees a general-purpose model.
The user sees an answer.
But the actual contribution of the source disappears in between.
That is not a healthy market.
A real market requires observability.
It must be possible to know which content was accessed, under what conditions, for which type of use, and with what level of attribution or commercial consequence.
Without that, there is no scalable negotiation.
Only suspicion.
Only litigation pressure.
Only broad redistribution.
Only reporting controlled by the party holding the logs.
A market cannot mature if the unit of value remains invisible.
The objection: law should come first
There is a fair objection.
Without stronger law, platforms will not create the records publishers need.
That is partly true.
Legal pressure matters. Regulatory obligations matter. Presumptions matter. Audit rights matter. They can force the market to move.
But law alone cannot operate the market.
A legal right still needs a record.
An audit right still needs something to audit.
A collective distribution mechanism still needs inputs.
A licensing renewal still needs evidence of what happened.
This is the point publishers should not miss.
The law can open the door.
Evidence walks through it.
From rights to records
The next phase of the AI-content economy will not be decided only by whether publishers obtain stronger rights.
It will be decided by whether those rights can be operationalized.
Rights need records.
Licensing needs measurement.
Attribution needs traceability.
Compensation needs a connection between use and value.
This is the missing layer between AI systems and trusted content.
Without it, lawmakers can reverse the burden of proof, but proof will remain hard to produce.
AI providers can propose contributions, but value will remain disconnected from actual usage.
Regulators can require summaries, but disputes will still lack granular evidence.
Publishers can demand compensation, but pricing will still depend on assumptions.
No record, no proof.
No proof, no enforceable right.
No measurable use, no real market.
The Darcos bill asked the right question.
The answer will not be legal alone.
It will be evidentiary.
And that is the problem Citations Logic is built to address: helping authoritative content owners create rights-aware AI usage records that make content use visible, attributable and commercially actionable.
Frequently asked questions
What was the Darcos bill about?
The Darcos bill proposed to introduce a legal presumption of use of cultural content by AI providers when credible indicators make that use plausible. Its purpose was to rebalance the burden of proof between rights holders and AI providers.
Why is a presumption of use not enough?
A presumption can shift who must prove or disprove use. But it does not create the records needed to establish what content was accessed, used, generated from, retrieved or attributed.
What is the difference between transparency and usage evidence?
Transparency provides broad information, such as summaries of training data categories. Usage evidence connects specific content, rights and usage events in a way that can support licensing, audit, renewal or dispute resolution.
Why does this matter for AI licensing?
AI licensing depends on knowing what was used, how often, under which conditions and with what commercial value. Without usage evidence, pricing and compensation remain based on assumptions.
What should publishers require?
Publishers should require rights-aware usage records, auditability, attribution logic, access controls, usage reporting definitions and the ability to connect usage events to assets, contributors, rights categories and licensing terms.
Continue the evidence chain
AI Usage Evidence for Publishers
AI Copyright Collective Management Distribution Key
EU AI Act and Content Licensing
AI Reuse Rights and Usage Signals
Book an AI usage evidence assessment
Sources
French Senate — Legislative dossier on the presumption of use of cultural content by AI providers
https://www.senat.fr/dossier-legislatif/ppl25-220.html
French Senate — Text adopted by the Senate on April 8, 2026
https://www.senat.fr/leg/tas25-085.html
French National Assembly — Legislative dossier and committee report filed on June 2, 2026
https://www.assemblee-nationale.fr/dyn/17/dossiers/DLR5L17N53359
French National Assembly — Report No. 2864 on the Darcos bill
https://www.assemblee-nationale.fr/dyn/17/rapports/cion-cedu/l17b2864_rapport-fond
Conseil d’État — Opinion on the proposed presumption of use of cultural content by AI providers
https://www.conseil-etat.fr/avis-consultatifs/derniers-avis-rendus/a-l-assemblee-nationale-et-au-senat/avis-sur-une-proposition-de-loi-relative-a-l-instauration-d-une-presomption-d-exploitation-des-contenus-culturels-par-les-fournisseurs-d-intelligen
ActuaLitté — Coverage of the June 11, 2026 parliamentary agenda and non-examination of the bill
https://actualitte.com/article/131870/politique-publique/presomption-d-utilisation-des-contenus-culturels-pour-le-groupe-d-attal-ia-pas-moyen
Reporters Without Borders — Call to unblock and adopt the Darcos bill
https://rsf.org/fr/soutenabilit%C3%A9-des-m%C3%A9dias-en-france-rsf-appelle-les-d%C3%A9put%C3%A9s-%C3%A0-d%C3%A9bloquer-le-d%C3%A9bat-et-%C3%A0-adopter-la-loi
European Commission — Template for GPAI model providers to summarize training content
https://digital-strategy.ec.europa.eu/en/faqs/template-general-purpose-ai-model-providers-summarise-their-training-content
European Commission — Guidelines for providers of general-purpose AI models under the AI Act
https://digital-strategy.ec.europa.eu/en/policies/guidelines-gpai-providers
Maddyness / AFP — Arthur Mensch’s proposal for a mandatory AI contribution in Europe
https://www.maddyness.com/2026/03/20/droit-dauteur-arthur-mensch-mistral-ai-plaide-pour-une-contribution-obligatoire-des-fournisseurs-de-modeles-dia-en-europe/