EU chatbot copyright opinion puts access and reused wording under scrutiny
An Advocate General says some chatbot practices can infringe publishers' rights. The opinion is not the final judgment, but it sketches two tests that could shape retrieval and answer design.
A chatbot can deliver a factually accurate answer and still create copyright exposure. Accuracy is only one issue. The legal questions concern how the service obtained a publisher's material, whether its answer reproduces protected expression and what responsibility the provider bears for that process.
On November 3, an Advocate General of the Court of Justice of the European Union issued an opinion. The dispute involved publisher Like Company and an artificial intelligence company.
The Court's case page describes a dispute involving retrieval and reuse of press publications by a large language model. The technology includes retrieval augmented generation, or RAG, which fetches external material before generating an answer.
The Court's press release outlines the Advocate General's approach. The opinion is nonbinding. The judges may accept, narrow or reject its reasoning.
The commercial question is whether chatbot providers can retain the usefulness of publisher-sourced answers without incurring greater licensing, compliance or infringement costs.
Access can be part of the infringement
The first proposed test concerns how material enters the system.
Under the Advocate General's analysis, unlawful access to publications, including circumvention of a paywall, may contribute to infringement. A publication's technical accessibility does not necessarily establish permission for its subsequent use.
The mechanism has four stages: obtaining source material, storing or indexing it, retrieving relevant passages and generating an answer.
Those stages matter because they create separate questions about authorization and reproduction. Permission to access a document does not necessarily resolve whether a particular output lawfully reuses its protected expression.
Two otherwise similar retrieval systems illustrate the distinction. One draws on a collection with documented permissions; another retrieves material without comparable authorization records. Their technical functions may be identical, but the evidence supporting their access practices differs.
This is an operational comparison, not a judicial finding that one architecture is categorically lawful.
For providers, provenance records, access controls and licensing documentation could become commercially important. The opinion leaves open whether every questionable scraping practice constitutes infringement.
The answer can matter as much as the request
The second proposed test concerns the answer itself.
According to the press release, the Advocate General's analysis addresses recognizable reuse of protected expression extensive enough for users to dispense with the original publication.
That is different from treating every factual answer as infringement. Facts and protected expression are not interchangeable legal categories.
Compare two chatbots. One states a reported fact and links to its source. The other reproduces an article's distinctive wording, selection or arrangement so comprehensively that reading the original adds little value.
The latter presents the stronger substitution concern. But neither answer length nor the presence of a citation determines legality by itself.
A further distinction concerns who supplied the material. An answer based on text submitted by a user involves a different access pattern from a service that systematically retrieves publishers' reporting from its own collection.
Neither arrangement automatically resolves the output question.
Providers therefore face a specific design challenge: preserve useful factual responses while reducing unauthorized reproduction of protected expression.
Liability can reach the business model
The Advocate General also considers whether infringement is inherent in a provider's business model or whether reasonable preventive measures are lacking.
That is not a finding of automatic liability for every generated response. It raises questions about recurring conduct, system design and the provider's ability to mitigate infringement.
The costs could affect both providers and publishers.
For chatbot providers, tighter source permissions, licensing arrangements, retrieval controls and output safeguards could increase operating costs or restrict available content.
For publishers, answers that replace visits to original reporting could weaken opportunities to earn advertising, subscription or licensing revenue.
The supplied record does not quantify either effect. It establishes no measured traffic loss, licensing price or provider cost.
There is also a counterargument: useful summaries and citations may help readers discover publishers rather than replace them. The balance between referral and substitution cannot be determined from this opinion alone.
The business question is whether providers can preserve answer quality while limiting unauthorized reuse, and whether publishers can capture economic value when their reporting informs those answers.
Possible safeguards include permission-controlled collections, access logs, licensing agreements and restrictions on reproducing protected passages. These are potential product responses, not measures already ordered by the Court.
The case is about architecture as well as law
RAG can connect chatbot answers to recent information and identifiable sources. That can improve traceability, but it also makes retrieval practices relevant to legal scrutiny.
A permission-controlled retrieval system may be better positioned to document authorized access than one lacking reliable source records. Yet both systems could still generate outputs that raise copyright questions.
This distinction separates two compliance problems: obtaining content lawfully and using protected expression lawfully.
The InfoCuria record provides the formal case materials as proceedings continue.
The public record provides a legal analysis and a pending case, not a comparable series of infringement rates, licensing costs or chatbot output measurements.
Possible outcomes
The most likely outcome is cautious adaptation. Providers strengthen source controls, preserve access records and reduce potentially substitutive outputs while awaiting judicial clarity. Evidence would include published sourcing policies or documented product changes.
A broader outcome is stronger publisher leverage. If the Court substantially adopts the Advocate General's reasoning on unlawful access, protected expression and provider responsibility, publishers could gain leverage in licensing negotiations. The evidence to watch is the judgment's reasoning, followed by material changes in provider practices.
A narrower outcome is more limited exposure. If the Court confines liability to particular access violations or clearly substitutive reproduction, ordinary factual answers may retain greater flexibility. The evidence to watch is a judgment expressly narrowing or rejecting the broader analysis.
CTOL expects major European chatbot providers to publish material source-access or publisher-licensing disclosures by June 30, 2027.
The forecast fails if no qualifying disclosure is observable by that deadline. Because the supplied record does not define a fixed provider universe, this remains a directional forecast rather than a fully specified industry-wide statistical test.
If the judgment has not arrived by June 30, the disclosure forecast remains assessable, but any claim that the disclosures resulted from the final ruling would be premature.
Evidence to watch
The decisive legal evidence is the Court's final judgment: its treatment of access, recognizable protected expression and provider responsibility.
The subsequent commercial evidence is whether providers materially change sourcing, licensing or output controls, and whether publishers can demonstrate changes in compensation or substitution.
Until then, the Advocate General's opinion is a legal signal, not binding law. The central question is whether a chatbot merely uses reporting to inform an answer or obtains and reproduces protected material in ways that require authorization.