When could an AI chatbot's news answer infringe a publisher's rights?
In the EU Court of Justice's first generative-AI case, the Advocate General says live retrieval may infringe when a chatbot either accesses a publication unlawfully or produces a recognisable substitute. The Opinion is influential legal advice, not a judgment.
What changed · 3 November 2026 at 20:00 GMT
Updated at 20:00 GMT after opening the full Opinion and the official referral. The report now makes clear that unlawful access and recognisable substitution are alternative routes in the proposed reproduction test, adds the Court's case-specific guidance and expands the evidence record from two official sources to four. The original 15:00 release and 3 November source date are unchanged.
The short version
- What happened
- The First Advocate General advised that an AI chatbot's real-time retrieval and reuse of a press publication may infringe EU press-publisher rights when access was unlawful or the answer recognisably reuses enough protected expression to substitute for the article.
- Who it affects
- Readers may see more source links, shorter summaries or refusals where a service cannot verify permission, but the Opinion does not ban factual reporting by chatbots.
- What remains unproven
- The Advocate General's Opinion is not binding; the Court has not yet delivered its judgment.
Keep in mind
The Advocate General's Opinion is not binding; the Court has not yet delivered its judgment.
The full legal Opinion is now linked and checked, but it remains non-binding and does not settle all copyright exceptions, reservations or future factual disputes.
The training questions were treated as hypothetical because the case record did not establish use of Like Company's articles for training.
The Court of Justice interprets EU law but does not decide the underlying Hungarian dispute or determine damages.
Whether any output infringes depends on the actual access path, source expression and generated answer, not the mere use of an LLM or RAG.
The case concerns the EU press-publisher right and cannot be assumed to apply identically under other countries' copyright laws.
Multi-source analysis · 4 linked sources. Primary analysed source date: 3 November 2026. Earliest linked supporting source: 17 May 2019.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 11 min The evidence, methods, implications and limitations
The direct answer: two factual tests, followed by a court judgment that has not happened yet
The First Advocate General of the Court of Justice of the European Union has proposed two alternative routes for assessing when a chatbot's live use of a news article may infringe a press publisher's reproduction right. The first question is whether the chatbot accessed the publication unlawfully, including by circumventing an effective paywall or anti-scraping measure. The second is whether the generated answer merely reports facts or instead reuses the article's form of expression recognisably enough that the user no longer needs to visit the original publication. If either condition is present, Maciej Szpunar's 3 November Opinion says a legally relevant reproduction may occur.
That is not a final ruling. Advocates General provide independent legal solutions for selected cases; the judges may follow, narrow or reject the proposed reasoning. The Court has begun deliberating and will deliver a judgment later. It will then be for the Budapest Regional Court to decide the dispute between Hungarian portal operator Like Company and Google in line with the Court's interpretation. The immediate significance is therefore a detailed legal map for the first generative-AI case to reach the EU's highest court, not a new ban on retrieval-augmented generation and not an automatic victory for publishers.[1][2]
What the Hungarian case actually asks
Like Company operates Hungarian web portals and claims the related EU right granted to press publishers over the online use of their publications. It alleged that Google used its articles to train Gemini's language models and repeatedly reproduced or made the articles available through chatbot answers between 13 June 2023 and 31 May 2024. The Budapest court paused the national proceedings and referred four questions to the Court of Justice. Two concern model training; two concern the answers generated for users after the chatbot retrieves current web material.
The distinction matters because training and retrieval occur at different stages. Training processes large bodies of text to produce model parameters before a particular user asks a question. Retrieval-augmented generation, or grounding, selects current material at query time and supplies it to the model to help compose an answer. A finding about one stage does not automatically settle the other. The Advocate General treated the live-answer questions as relevant to the evidence before the Hungarian court, while advising that the two training questions were hypothetical on the case record.[1]
The distinction that matters
The proposed test for a live chatbot answer
The Advocate General separates model training from real-time retrieval and asks what happened to the specific publication and output.
Access
Did the chatbot bypass effective access or anti-scraping controls?
Unlawful retrieval would weigh toward infringement
Output
Does the answer report facts or recognisably reuse expression?
Facts alone are not protected; substitutive expression may be
Pattern
Is recognisable reuse systematic and tied to the service design?
Provider responsibility and safeguards become relevant
Status
Wait for the Court's judgment and national application
The Opinion is influential but not binding
Why the training questions were not answered on their facts
Szpunar advised the Court to declare the training questions inadmissible because the national case file contained no evidence that Like Company's articles had been used to train Gemini. He also reasoned that training an LLM is not an 'online use' of a press publication for the specific press-publisher right at issue. In EU preliminary-ruling procedure, the Court interprets law needed to resolve a real national dispute; it does not ordinarily issue an abstract opinion about facts that have not been established.
For completeness, the Advocate General nevertheless suggested that using copyright-protected works or related subject matter to train an LLM is, as a rule, permitted under the Directive's text-and-data-mining exception. That statement is part of a non-binding Opinion, and the press release does not resolve every condition, reservation or opt-out that can arise under EU copyright law. It should not be converted into the broader claim that all AI training on any material is lawful. The Court could decline to address the hypothetical questions, adopt different reasoning or leave important training disputes for later cases with a fuller evidential record.[1][2]
Facts are not the same as protected expression
The Opinion places a boundary at the heart of journalism and copyright. The press-publisher right protects the investment embodied in a press publication's particular expression; it does not give a publisher ownership of the underlying news facts. A chatbot answer that states a fact learned from an article is therefore not necessarily a reproduction of that publication. The legal concern becomes stronger when the answer preserves recognisable wording, structure or expressive choices from the source and substitutes for reading it.
This is a contextual test rather than a mechanical word count. The official summary says the relevant form of expression must be recognisable to the point that the user is entirely dispensed from accessing the original publication. Short factual answers, close paraphrases, lengthy summaries and copied passages may fall differently depending on what was taken and how it appears. The national court must examine the actual publications, access route, prompts and outputs. The Opinion does not supply a universal similarity score, a safe number of words or an automatic rule that every AI summary infringes.[1][2]
Paywalls and technical restrictions could change the analysis
The first limb asks whether the chatbot obtained the article unlawfully, particularly by circumventing effective technical measures designed to restrict access or scraping. That creates a practical difference between content deliberately available for ordinary public access and content obtained after bypassing a publisher's controls. It also makes evidence about robots instructions, account restrictions, paywall behaviour, licensing and the retrieval system's actual request path important. A publisher's assertion that scraping was unwanted will not by itself establish how the system accessed a page; a provider's claim that material was public will not answer whether an effective control was circumvented.
For developers, the Opinion points toward retrieval logs, source-level permissions and controls that can prevent recognisable reproduction. For publishers, it highlights the need to document access restrictions and the precise outputs said to replace an article. For auditors and courts, a generated answer without provenance may be insufficient: they may need to reconstruct the source request, the retrieved passages, the model's transformation and whether the behaviour was routine. The Court has not yet endorsed these requirements, but the proposed test makes technical system design part of the legal evidence.[1]
Making available and provider responsibility
The Advocate General also addressed the publisher's right of making a publication available to the public. The official summary says that right is, as a rule, infringed when the publication's expression is systematically recognisable in chatbot answers to related queries to the degree described above. The word 'systematically' matters: an isolated anomaly and a repeatable product behaviour may present different evidence about what the service makes available and whether the output substitutes for the source.
He proposed that the chatbot provider may be liable where the relevant use of press publications is inherent in its business model or where it failed to put credible and effective countermeasures in place. That does not mean every provider is strictly liable for every model output. It directs attention to product design, foreseeable patterns, safeguards and the provider's response to known risks. The full Opinion also indicates that Like Company's example may not satisfy the proposed test: subject to the national court's findings, effective technical restrictions were apparently absent, the article's expression did not appear recognisable in the answer, the acts did not appear inherent to Google's business model and Google seemed to have mitigation measures. The final judgment may define these concepts differently, and the Hungarian court will still need to apply any interpretation to evidence about Gemini and the disputed articles.[1]
What this could mean for readers, publishers and AI services
Readers could see more visible source links, shorter summaries, licensing signals or refusals when a service cannot verify permission. Those changes could help people distinguish a generated answer from the journalism it draws on, but they could also reduce access to convenient summaries. Publishers could gain a clearer route to challenge answer patterns that bypass controls and reproduce their work, while still lacking a monopoly over facts. Smaller outlets may find enforcement expensive unless the final judgment produces a test that can be applied without reconstructing every request manually.
AI providers may need to separate training governance from live-retrieval governance, because the Opinion treats them as different questions. A system could have a lawful basis for training yet still create risk by retrieving a restricted article and generating a recognisable substitute. Conversely, a service may retrieve publicly accessible material and produce an answer that conveys only facts in new language. Product teams would need source-aware controls, output testing and an escalation path for publishers rather than a single copyright setting applied to the entire model.[1][2]
What would change the assessment
The decisive next evidence is the Court's judgment. Confidence in the Advocate General's framework would rise if the judges adopt the unlawful-access and recognisable-expression test, clarify what counts as systematic substitution, and explain the provider safeguards expected in practice. The legal impact would then depend on how the Budapest court applies those principles to Like Company's evidence. A judgment could also narrow the dispute to procedural questions, reject parts of the proposed test or avoid the training issue entirely.
Practical assessment will require real output evidence: the disputed articles, the access controls in force, the chatbot's retrieved material and the answers delivered to users. Independent measurement of how often AI answers replace visits would inform the economic debate but would not itself decide copyright infringement. Until the judgment and national application arrive, publishers should not describe the Opinion as a binding win, and AI providers should not describe the suggested text-and-data-mining analysis as a blanket clearance. The development is consequential because it frames the questions the Court is now considering, not because it has already settled them.[1][2]
Continue the story
Read next
Continue exploring retrieval-augmented generation: compare the finding and its evidence limits.
Tlaxcala’s assistant invented no fee in 40 audited AI answers—but half of clarifying users left
New analysis today of an unreviewed 27 October preprint: 380 production traces suggest deterministic routing can contain numeric errors, while slow generative replies and a 53% clarification drop-off expose the human cost of friction.
Read the next reportExplore the full source and evidence record
What this means for people
- Readers may see more source links, shorter summaries or refusals where a service cannot verify permission, but the Opinion does not ban factual reporting by chatbots.
- Publishers gain a clearer proposed route for challenging outputs that bypass controls and substitute recognisable expression, although enforcement still requires evidence.
- AI product teams may need separate governance for model training and live retrieval, with source-level access checks and tests for repeatable substitutive outputs.
Global context
The case could influence AI retrieval products offered across the European Union, while courts elsewhere apply different copyright statutes and exceptions. It also arrives amid wider negotiations between publishers and AI companies over licensing, attribution and traffic. The Opinion preserves a boundary between facts and expression, so its proposed framework is narrower than a general right to payment whenever an AI service uses news information.
What the evidence does not yet show
- The Advocate General's Opinion is not binding; the Court has not yet delivered its judgment.
- The full legal Opinion is now linked and checked, but it remains non-binding and does not settle all copyright exceptions, reservations or future factual disputes.
- The training questions were treated as hypothetical because the case record did not establish use of Like Company's articles for training.
- The Court of Justice interprets EU law but does not decide the underlying Hungarian dispute or determine damages.
- Whether any output infringes depends on the actual access path, source expression and generated answer, not the mere use of an LLM or RAG.
- The case concerns the EU press-publisher right and cannot be assumed to apply identically under other countries' copyright laws.
What to watch next
- The Court of Justice's final judgment and whether it adopts the proposed two-part test
- The Budapest Regional Court's findings about access controls, retrieved passages and Gemini outputs
- How the Court defines systematic recognisability, substitution and credible countermeasures
- Provider changes to source permissions, citations, summary length and publisher complaint handling
- Later EU cases with direct evidence about model-training datasets and rights reservations
Living evidence record
Impact record IAI-1WBEPR2
Evidence stage
Announced
Confidence
Supported
Reporting basis
Multi-source analysis
Independent or research support
Not yet
Record status
Updated
Last checked
3 November 2026
Source trail
4 direct sources across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Evidence trail
Sources used for this report
Links checked 3 November 2026
Research question, topics and regions
Topics: Copyright · Press publishers · Retrieval-augmented generation · Google Gemini · EU law · Online news
Regions: European Union · Hungary · Europe
This report is labelled multi-source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.