Takeaway: On 3 November 2026, Advocate General Maciej Szpunar delivered his Opinion in Case C-250/25 Like Company — the first generative-AI dispute to reach the Court of Justice of the European Union. He advises the Court to treat LLM training questions as inadmissible on these facts (and, for completeness, broadly covered by the TDM exception in Article 4 CDSMD), while holding that real-time RAG / chatbot reuse of online press publications may infringe press publishers’ related right under Article 15 CDSMD when access is unlawful or answers recognisably substitute the original article. The Opinion is not binding; judgment follows later.
Primary sources: Court of Justice press release No 155/26 (Luxembourg, 3 Nov 2026); analysis by João Pedro Quintais on the Kluwer Copyright Blog (3 Nov 2026); summary by CEA Digital Law (4 Nov 2026). This is a different angle from today’s product-news posts on Google K-12 Gemini/Search defaults, ChatGPT search-like volume vs Google, and Anthropic FlashProve — here the story is publisher rights and EU copyright, not a new product launch.
What the Like Company case is about
Like Company, a Hungarian operator of web portals, sued Google Ireland Ltd. before the Budapest Regional Court over Google’s Gemini chatbot. It relies on the press publishers’ related right in Article 15(1) of Directive (EU) 2019/790 (the Copyright in the Digital Single Market Directive, or CDSMD) — protection for the “online use” of press publications. The claim: Google used Like Company’s articles to train Gemini’s large language models and, between 13 June 2023 and 31 May 2024, reproduced those articles and made them available through the chatbot.
Gemini uses Retrieval-Augmented Generation (RAG, also called “grounding”): the system can collect information in real time from the open web to improve answer accuracy. That means a news query may trigger retrieval (or scraping) of press articles, then reuse of that material when generating the reply. The Hungarian court stayed proceedings and referred four preliminary questions: two on training / text-and-data mining, and two on whether chatbot answers engage reproduction and making-available rights.
Training questions: inadmissible — and TDM for completeness
Advocate General Szpunar advises the Court to declare the two training questions inadmissible. First, there is no evidence that Like Company’s articles were used to train the Gemini models; the AG emphasises that the main proceedings are not really about LLM training. Second, he treats LLM training as outside “online use” of press publications under Article 15, so the right relied on in the case does not reach training copies in the same way.
For completeness, he still sketches the copyright picture for training. Reproductions of protected material made to train an LLM are, as a rule, covered by the text and data mining exception in Article 4 CDSMD, if its conditions are met: lawful access, no applicable reservation of rights (the familiar “opt-out”), and compliance with the three-step test. He reads TDM’s definition broadly enough to cover extracting linguistic patterns for model training; later generative use of those patterns does not, in his view, strip the analysis out of TDM. Exceptional three-step-test failures remain possible case by case — but the baseline is a permissive training frame when Article 4’s gates are satisfied.
That training-vs-output split matters for publishers who have been collapsing every AI use into one licensing conversation. Machine-learning crawls, robots.txt training tokens, and product answer surfaces are different legal and commercial problems — a distinction that also shows up in practical tooling we cover for AI crawler allow/block rules and Google’s shared AI-training robots.txt policy token.
Where the Opinion focuses: chatbot answers and RAG
The AG urges the Court to concentrate on the chatbot-answer questions, which are relevant on the facts. When an LLM-driven chatbot retrieves a press publication from a publisher’s site, reuses it to generate an answer, and shows that answer to the user, the press publisher’s right may be infringed in certain circumstances.
Two guardrails structure the analysis:
- No news monopoly. Article 15 protects the specific form of expression of the press publication and the investment that went into it — not mere facts or a proprietary claim over “the news.”
- RAG is “online use.” Unlike training (on the AG’s view), retrieving publications to generate answers can constitute online use under Article 15. Technical copies in RAG are not Article 4 TDM (they fetch specific content to improve an answer, not to mine patterns), though temporary-reproduction rules may still matter for intermediate copies.
Infringement factors: unlawful access or substitutive reuse
For the reproduction right, the Opinion turns on two alternative factors:
- Unlawful access — for example, circumventing effective technical measures that restrict access or scraping (paywalls and comparable anti-scrape controls).
- Recognisable form of expression — the answer reuses the publication’s protected expression so that the user is entirely dispensed from reading the original.
These are alternatives, not a checklist that both must fail. Lawful access to an article does not automatically make a recognisable, substitutive paraphrase in the chatbot answer lawful. Conversely, a factual summary that does not carry recognisable structure, arrangement, or phrasing may divert traffic without reproducing the protected publication under the AG’s proposed test.
For making available to the public, infringement arises, as a rule, where the form of expression is systematically recognisable in answers to related user queries, to that same substitution extent — users can skip the original because the chatbot already supplies the expressive content.
| Issue | AG’s proposed direction | Publisher / SEO takeaway |
|---|---|---|
| LLM training (Qs 2–3) | Inadmissible here; training not “online use” under Art.15; Art.4 TDM generally covers qualifying training copies | Keep training opt-outs and lawful-access hygiene separate from chatbot licensing |
| Chatbot RAG answers | May infringe Art.15 when access is unlawful or answers recognisably substitute the article | Paywalls, anti-scrape measures, and answer design become central evidence |
| Facts vs expression | Art.15 protects form / investment, not a monopoly over news facts | Pure factual restatements are weaker Article 15 claims |
| Provider liability | Liable if acts inherent to business model OR failed appropriate / credible countermeasures | Diligent technical and product safeguards matter even when a user typed the query |
| Status of Opinion | Not binding; Court judgment later | Monitor final judgment + related Hungarian referrals C-806/26, C-829/26 |
Who is liable: the chatbot provider
Attribution is not automatically pushed onto the end user who typed the prompt. The AG proposes that relevant acts may be treated as carried out by the chatbot provider where they are inherent to the provider’s business model, or where the provider failed to put in place measures a reasonably diligent operator would use to counter those acts credibly and effectively. Occasional mistakes are treated differently from a service designed to substitute for press reading without safeguards.
CEA Digital Law’s 4 November commentary frames the balance as avoiding two extremes: no monopoly over the news, but no free pass to substitute the original outlet. Open questions remain — how courts will measure “recognisability,” which countermeasures count as credible, and how unlawful access will be proven in practice.
What publishers, SEO, and licensing teams should do now
Even before the Court’s judgment, the Opinion already reshapes how European publishers should talk about generative AI:
- Separate training from grounding. Negotiate and document TDM opt-outs / licensing for model training separately from RAG answer licensing. Collapsing both into one “AI fee” muddies remedies the AG treats as legally distinct.
- Treat effective access controls as evidence, not theatre. Paywalls and anti-scrape measures are not merely commercial pricing tools; under this framework they feed the unlawful-access prong.
- Watch substitution, not only citation. A chatbot that cites a URL while still delivering a recognisable rewrite that dispenses users from the article sits closer to the infringement zone than a short factual pointer that sends readers through. That sits beside, but is not identical to, product SEO work on getting cited in AI search retrieval and Google’s generative AI SEO guidance.
- Document provider safeguards. Publishers evaluating partnerships or complaints should ask what countermeasures exist against paywall bypass and long-form regurgitation — the AG’s liability test centres on business-model design and diligence.
- Do not treat the Opinion as final law. It is advisory. Related Hungarian generative-AI referrals — Manchester United Fanatics Klub (C-806/26) and Mobilguru (C-829/26) — and national cases elsewhere will continue to develop the map.
For traffic and monetisation teams already tracking answer engines versus classic Search — including Alphabet’s Q3 2026 AI Mode monetisation pressure and OpenAI’s claim that ChatGPT search-like queries are ~20% of Google query volume — the Like Company Opinion adds a European rights layer: when AI answers become substitutes for journalism’s form of expression, Article 15 may become a litigation and licensing lever, not only a robots.txt debate.
FAQ
Is the Advocate General’s Opinion already EU law?
No. An Advocate General’s Opinion proposes a legal solution to the Court but is not binding. The judges deliberate afterward; judgment is delivered later. National courts apply the Court’s eventual ruling.
Does the Opinion ban AI training on news?
No. On these facts the training questions are treated as inadmissible. For completeness, the AG suggests that qualifying training reproductions are generally covered by Article 4 CDSMD (TDM) if lawful access, no opt-out, and the three-step test are satisfied.
When can a Gemini-style chatbot answer infringe press publishers’ rights?
Under the AG’s framework: when the chatbot accessed the publication unlawfully (for example by circumventing effective paywall / anti-scrape measures), or when the answer reuses a recognisable form of expression so the user is entirely dispensed from reading the original — with making-available liability where that substitution is systematic across related queries.
Does Article 15 give publishers ownership of the news?
No. The Opinion stresses that Article 15 protects the form of expression and the investment in the publication, not a monopoly over mere facts reported in the news.
Who may be liable — the user or Google?
The AG proposes provider liability where the acts are inherent to the chatbot business model or the provider failed to implement appropriate, credible countermeasures. A user initiating a query does not automatically shield the provider.
Are other generative-AI copyright cases pending at the CJEU?
Yes. Commentary notes two further Hungarian referrals: Manchester United Fanatics Klub (C-806/26) and Mobilguru (C-829/26). The Like Company judgment will matter for those dockets as well.
Bottom line for Web Observer readers: the first generative-AI Opinion at the CJEU draws a bright line between training (largely TDM-framed, and inadmissible on these facts) and RAG chatbot reuse that can trigger Article 15 when access is unlawful or answers become recognisable substitutes for press expression. Publishers should harden access controls, split licensing asks, and watch the Court’s forthcoming judgment — without assuming today’s product AI surfaces already cleared EU press-publisher risk.