ia-generative3 November 2026DILAIG

AI, RAG and copyright: what Advocate General Szpunar proposes

In Like Company v Google, the Advocate General distinguishes training, RAG and chatbot outputs. What businesses can prepare before the Court’s judgment.

Article written with the help of artificial intelligence — learn more about Article 50

Last updated: 3 November 2026 · Reading time: 8 minutes

In brief. On 3 November 2026, Advocate General Maciej Szpunar delivered his Opinion in Case C‑250/25, Like Company v Google Ireland. He proposes an approach to press publishers’ rights when chatbots retrieve online publications to answer user queries. This is a non-binding Opinion, not a judgment of the Court of Justice. Its practical value for businesses lies in separating model training, the sources consulted for a query and the content actually returned.

Does your assistant consult articles, summarise documents or search a vector database? The Opinion is relevant to your review of content rights. It does not establish that all RAG is unlawful, or that paraphrasing is enough to respect copyright.

What has happened before the Court?

Like Company operates news portals in Hungary. It challenges the use of its publications by Google Gemini and invokes the press publishers’ related right under Directive (EU) 2019/790. The Hungarian court referred questions on the interpretation of EU law to the Court of Justice.

To illustrate its claim, Like Company produced a Gemini answer to a query requesting a summary of an article about singer Kozsó’s plan to introduce dolphins into an aquarium at Lake Balaton. The publisher considers this evidence that its publication was used to train the model and that the answer infringes its rights. The Advocate General distinguishes these questions: retrieving a publication to answer a query does not establish that it was used for training. He examines access to the content and the form of the resulting answer separately.

The Advocate General provides an independent proposal to assist the Court. The Court may follow or depart from that reasoning. After its interpretative judgment, the national court will still need to establish the facts and resolve the dispute in accordance with that interpretation.

CURIA describes this as the first generative AI case to reach the Court. Saying that “the CJEU has just ruled on AI for the first time” would nevertheless be inaccurate. At the date of this article, the document is a proposed interpretation concerning a particular use, primarily under copyright and related-rights law.

Three operations to examine separately

Operation	What the system does	What needs documenting
Training or fine-tuning	Processes a corpus to adjust a model’s parameters	What content was copied, with what access and rights?
Document retrieval or RAG	Retrieves content to ground an answer	Which sources, copies, restrictions and retention periods?
Answering the user	Produces text, a summary or excerpts	Does the answer recognisably reuse the source’s expression?

This distinction also matters for responsibility. The model provider, application provider and organisation using the application do not necessarily perform the same operations. Using an API does not mean having trained the model. Choosing and indexing the application’s sources is nevertheless an activity that warrants its own examination.

Training: a supplementary analysis, not blanket permission

The Advocate General proposes that the questions about training should be declared inadmissible because there is insufficient evidence that the articles at issue were used to train the Gemini models. A chatbot can retrieve a publication during a search without having memorised it during training.

He nevertheless provides a supplementary analysis, parts of which are expressly provisional. Under that analysis, reproductions necessary for training can, in principle, fall within the text and data mining exception, or TDM, subject to conditions. Lawful access and appropriate reservation of rights are among the issues examined.

This does not mean that all training on internet content is permitted, that a rights reservation has no effect, or that the Court has already approved a general practice. The training questions are precisely those that the Advocate General proposes excluding from the operative answer in this case.

RAG: searching to answer a query is not necessarily TDM

In the dynamic RAG operation examined, the chatbot retrieves a particular publication to enrich a particular answer. At point 179, the Advocate General considers the TDM exception irrelevant to that operation: its purpose differs from extracting patterns or correlations from a dataset.

He then examines the technical reproductions made during processing and the conditions of the temporary-copy exception under Directive 2001/29/EC. His analysis requires, among other things, lawful access and lawful end use. The scenario described involves copies automatically deleted no later than after the answer is provided.

Persistent document databases require separate analysis. A retained corpus, lasting cache or vector index does not automatically match that scenario. You need to examine what is actually stored, for how long, what reproductions take place and the justification for each use. Conversely, deleting a copy after an answer is not, by itself, enough to satisfy every condition of an exception.

The technical label “RAG” therefore provides no universal legal answer. Two assistants can have very different architectures and permissions.

Paraphrasing is not always enough

The Opinion distinguishes facts from their expression. Factual information does not receive the same protection as its presentation in a publication. The reasoning concerns the form of expression, which may include arrangement, structure and wording.

In his final proposal concerning the press publishers’ related right, the Advocate General refers to an answer in which the expression of all or part of the publication is recognisable to such an extent that the user can dispense entirely with consulting the original publication. Paraphrasing therefore does not necessarily rule out reproduction; conversely, mere resemblance does not fully describe the criterion he proposes.

He distinguishes reproduction from making available to the public: for the latter, he requires expression to be recognisable in that way systematically in answers to related user queries. Unlawful access to the source, particularly through circumvention of effective technical protections, is a separate scenario in his reasoning. These criteria, proposed at points 229 and 240, have not yet been adopted in a judgment of the Court and are not a universal test for all works. No similarity percentage or word count provides a general threshold of lawfulness.

Avoid three shortcuts:

“The page is accessible.” Absence of a technical block does not, on its own, establish general consent to reuse the publication to ground a chatbot.
“We include a link.” Attribution helps readers find the source; it does not replace the rights needed for the actual use.
“The model paraphrases.” A paraphrase can preserve recognisable elements. Controls need to consider the actual answer.

This is not a prohibition on all summaries. Subject to the national court’s factual findings, the Advocate General even considers that the facts presented in this case appear not to satisfy the conditions for infringement. The Opinion is therefore not a finding against Google.

Who should act within the business?

The Advocate General does not propose automatic provider liability for every problematic answer. To attribute the acts of reproduction or making available examined to the provider, he identifies two scenarios: the acts are inherent to the provider’s business model, or the provider has failed to implement the appropriate measures expected of a reasonably diligent operator to counter those infringements credibly and effectively. The service’s design and the measures actually taken therefore require examination, without presuming provider liability or excluding it through a simple reference to the user or model provider.

An application provider that selects sources, organises retrieval and determines the format of summaries should examine those operations separately. An organisation using a third-party tool should start by identifying its uses, the documents it supplies and the contractual protections actually available.

The AI Act’s obligations for general-purpose AI model providers remain a distinct issue. Integrating a third-party model does not automatically make an organisation subject to every obligation of that model’s provider. Equally, it does not settle the rights over the organisation’s own corpus or outputs.

Six useful checks to start now

These are preparation recommendations, not six new legal obligations created by the Opinion.

Describe the operations. Training, fine-tuning, live retrieval, retained corpora and analysis of supplied documents: establish who does what.
Inventory the sources. Distinguish internal content, customer documents, official texts, licensed material, legal commentary and press publications. Check rights even when a document is already in your workspace.
Document permitted uses. Retain references to licences, permissions or exceptions relied upon, their scope and identified restrictions. A GDPR lawful basis answers a different question.
Examine access and retention. Identify technical protections, rights reservations, copies, caches and indexes. Plan appropriate controls and a procedure to exclude problematic sources.
Test outputs. Review ordinary requests for summaries and reproductions. Similarity checks can help identify issues; they do not provide a legal verdict.
Assign responsibility for review. Identify who handles uncertain cases, complaints and the future judgment. Missing evidence calls for verification and does not, by itself, establish infringement.

FAQ

Has the CJEU banned RAG?

No. This is a non-binding Opinion about a particular operation involving press publications. It proposes a conditional analysis, not a general prohibition.

Can I summarise an article if I name the publisher?

Naming the publisher does not answer the whole question. Source access, applicable rights and exceptions, copies made and the content returned all need examination.

Our assistant only uses internal documents. Is this relevant?

That differs from the dispute, but “internal” does not establish ownership of every right. Your corpus may contain purchased articles, standards, studies or partner documents subject to particular conditions.

Should we change our AI Act risk classification?

The Opinion creates no new high-risk category for RAG systems. AI Act classification and content-rights analysis follow different criteria. Document both without conflating them.

What now?

Start by understanding your systems, their uses and their sources. The free DILAIG audit helps structure your situation under the AI Act. That process provides neither permission to use third-party content nor certification of copyright compliance. For sensitive corpora or output practices, have the rights and supporting evidence specific to your case examined.

Editorial analysis as of 3 November 2026. Discussion of points 179 to 240 refers to the Advocate General’s Opinion and should be reviewed in light of the Court’s judgment. This article does not replace legal advice tailored to a particular situation.