By Thorsten Meyer | November 4, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
It started with dolphins. In July 2023, a Hungarian news portal reported that a popular singer wanted to bring dolphins to Lake Balaton, in a pool 25 metres long on the Tihany peninsula. Some time later, someone asked Google’s Gemini chatbot to summarise that article, and Gemini did, neatly, in Hungarian. The publisher took the answer to court. On 3 November 2026, that summary reached the highest court of the European Union.
Advocate General Maciej Szpunar, one of the Court of Justice’s independent legal advisers, delivered his opinion in Case C‑250/25, Like Company v Google Ireland. The court calls it “the first generative AI case to reach the Court of Justice”. His answer: a chatbot that fetches news articles in real time and reuses them in its answers can infringe the rights of press publishers. But only in certain circumstances, and he draws the line in an unusual place. The facts in a news story are free. The way it is written is not.
And then comes the twist. Applied to the dolphin article, his test points against the publisher that brought the case.
Key takeaways
- What happened: On 3 November 2026, Advocate General Szpunar advised the EU Court of Justice in C‑250/25 that AI chatbots which retrieve and reuse online press articles to answer user questions “may, in certain circumstances, infringe the rights of press publishers” (CURIA press release No 155/26).
- The line: An answer infringes if the chatbot reached the article unlawfully, for example by getting around a paywall or anti-scraping block, or if the article’s form of expression is so recognisable in the answer that the user “can dispense with accessing the original publication entirely” (point 240). Reporting the same facts is not enough.
- Training was left out: He told the Court to declare the two questions on AI training inadmissible, because nothing showed that the articles were used for training. Only “for the sake of completeness” did he add that training is, as a rule, covered by the EU’s text and data mining exception.
- The twist: In the actual case, he says, the conditions do not appear to be met: no access barriers, no recognisable wording in the example answer, and Google appears to have safeguards (point 239). The Hungarian court decides.
- Not a ruling yet: The opinion is not binding. No date for the judgment has been set. Preliminary references took 16.9 months on average in 2025, according to the court’s own statistics.
- What it means: Companies running chatbots or AI search over web content need to respect access barriers and keep answers away from near-copies. Publishers get leverage for licences, but not a monopoly on the news. Users may see more synthesis from several sources, and more links.
This article explains a court document. It is general information, not legal advice.
Products Worth Considering
Amazon Kindle Scribe (64GB) with Premium Pen, Metallic Jade
Samsung Galaxy Tab A11 (2026) – 8.7 90Hz Android Tablet, 64GB/4GB Wi‑Fi, Octa-core MediaTek Helio G99, Dual Dolby Speakers, 20W Fast Charger
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
As an affiliate, we earn on qualifying purchases.
The case in one paragraph
Like Company runs several Hungarian web portals. It sued Google before the Budapest regional court (Budapest Környéki Törvényszék), claiming the press publishers’ right that EU law has granted since 2019 (Article 15(1) of the Copyright in the Digital Single Market Directive, 2019/790). According to the court’s press release, Like Company argues that Google used its articles to train Gemini’s language models and “repeatedly reproduced and made them available to the public through the Gemini chatbot” between 13 June 2023 and 31 May 2024. The Hungarian court sent four questions to Luxembourg: two about training, two about chatbot answers. Nine EU governments and the European Commission submitted written observations, and the hearing took place on 10 March 2026 (point 21).
The press publishers’ right itself was written with a different target in mind. The opinion recalls that the EU legislature adopted it so that publishers could control and license the use of extracts by news aggregators “such as Google News” (point 22). Seven years later, the question is whether the same right reaches chatbots.
What the Advocate General said about training
On training, his main answer is procedural. The two questions are “hypothetical” and therefore inadmissible (points 46 to 58), for two reasons.
First, nothing supports the claim that Like Company’s articles were in Gemini’s training data. The dolphin article was published on 21 July 2023, and Gemini 1.0 was released in December 2023. Since training large models takes months and data is collected well in advance, that is “implausible”, he writes (point 49). Gemini could summarise the article because it fetched it live from the website, through a technique called retrieval-augmented generation, or RAG: the chatbot searches the web, pulls in matching pages and writes its answer on top of them (point 50). The user had even typed the domain name into the question.
Second, the press publishers’ right only covers “online use” of press publications by online services. Training a model on dedicated computers is not an online use, he finds, so it falls outside that right altogether (points 51 to 57).
He then discussed training anyway, “for the sake of completeness”, and the press coverage that says he “approved” AI training draws on this part. Training involves copies in the legal sense (point 59). But he tends to think that the EU’s text and data mining (TDM) exception in Article 4 of the 2019 directive covers it “in principle” (point 99), and he reads the EU AI Act the same way: its recital 105 and Article 53(1)(c), which requires providers of general-purpose AI models to have a copyright policy, start from that premise (points 82 and 109). Rights holders can opt out in a “machine-readable” way. He counts robots.txt, Google-Extended and possibly even plain-language terms on a website, as long as current bots can reliably understand them (points 148 and 149). And he calls the AI Act’s duty to publish a summary of training content “a crucial complement” (point 150).
Two caveats sit in the same section. Developers who rely on the exception should show they took steps against outputs that copy training material too closely, such as deduplication and output filters (points 132 and 133). And a model that spits out a protected text almost word for word can infringe through that output (point 89). He is cautious, however, about the idea that the copies are stored inside the model itself, which he calls “far-fetched” at this stage (point 90). That is a different view from the Munich Regional Court in GEMA v OpenAI (more below).
The core: when a chatbot answer infringes
The questions about chatbot answers are the “real crux” of the case, he writes (point 161), and they are admissible. Several findings are worth reading closely.
EU law applies to chatbots offered in the EU, wherever the provider or its servers are located (points 165 and 166). Retrieving articles through RAG and answering a user is an “online use” (points 167 to 169).
Retrieval creates copies. When the chatbot fetches an article and processes it, technical copies of more than “very short extracts” are made. Google “admitted as much at the hearing” (point 177). The TDM exception does not cover this, because RAG is not about finding patterns in large datasets but about using one specific article for one specific answer (point 179). Instead, these copies can fall under the exception for temporary copies in Article 5(1) of the 2001 InfoSoc Directive: transient, part of a technical process, with no separate economic value (points 180 to 184). That exception only works if the end use is lawful. Whether it is lawful depends on two factors (points 187 to 189).
Factor one: how the chatbot got in. If the publisher uses effective technical measures, such as a paywall or a block aimed at scraping bots, the chatbot must not get around them (point 188).
Factor two: what comes out. The press publishers’ right protects the “form of expression”, meaning the way an article is “arranged, structured and phrased”, not the facts it reports (point 199). An answer is a reproduction when that form is recognisable in it and the answer becomes a substitute for the article, because a reader “can dispense with accessing the original article entirely” (point 202). Paraphrasing is no escape: the test is not limited to word-for-word copying, or the right would be “too easy to circumvent” (point 205). But a short summary of facts, presented differently, or an overview put together from several sources, does not infringe (point 208). In between lies a grey area to be judged case by case (point 209).
Making available to the public requires more: the article must be “systematically recognisable” in the chatbot’s answers to related questions (points 213 and 240). He considers that unlikely, given how random chatbot answers are, unless the RAG process is designed that way, intentionally or through negligence (point 214).
Who is liable. Users who ask for something are generally the primary actors, and private, non-commercial use is outside the publishers’ right anyway (points 218 and 219). The provider is liable when such uses are “inherent” to its business model, or when it has not taken the measures a “reasonably diligent operator” would take: bots that respect access restrictions, and output filters that keep answers “sufficiently removed” from any single article (points 221 and 222).
No silent consent. Google argued that a publisher who leaves articles freely accessible implicitly consents to their use in chatbot answers. “I do not think that is the case,” the Advocate General replies (point 225). A paywall is a major commercial decision. And blocking bots is “not as easy as it sounds”: Google appears to use the same bots for search and for AI grounding, so blocking them via robots.txt may hurt a site’s visibility in search (point 228). Google says its Google-Extended control separates the two, which, the opinion notes, “would need to be verified” (footnote 236). If you want the technical background on these opt-out signals, I went through what the new shared opt-out standard covers and what it doesn’t.
Infographic 01 · Where the line runs
When a chatbot answer infringes a publisher's right
Advocate General Szpunar's test in Case C‑250/25 for chatbots that retrieve news articles in real time and reuse them in their answers.
Stays free
- Reporting the same facts. Nobody owns the news itself (pts 199–200).
- A short factual summary, presented differently (pt 208).
- An overview built from several sources, restructured, with links back (pts 208, 234).
- Individual words or very short extracts (pt 201).
Can infringe
- Getting around a paywall or anti-scraping block to fetch the article (pt 188).
- An answer that keeps the article’s structure and phrasing, even paraphrased, so readers can skip the original (pts 202, 205, 207).
- Systematically recognisable articles in answers to related questions: making available to the public (pts 213, 240).
The test, step by step
For a chatbot that fetches a news article live (RAG) and answers a user
- Is it an online use by a service in the EU?Yes for RAG answers to EU users, wherever the servers are (pts 165–169).
- Did the bot get in lawfully?No circumvention of effective technical measures (pt 188).
- Is the article’s form of expression recognisable in the answer, to the point of substitution?Side-by-side comparison; facts do not count (pts 196–209).
- Is the provider responsible?If inherent to its business model, or no reasonable measures such as respecting blocks and output filters (pts 221–222).
- Temporary copies during retrieval are then exempt under Art. 5(1) InfoSoc; the TDM exception does not apply to RAG (pts 179–189).
An Advocate General’s opinion, not a judgment. General information, not legal advice.
Sources: Opinion of AG Szpunar, C‑250/25 (provisional text), CURIA press release No 155/26 (3 Nov 2026). Point numbers refer to the provisional text.
Powered by Thorsten Meyer AI
The twist: the publisher may lose its own case
At the end comes “brief guidance” for the Hungarian court (point 239). Subject to that court’s findings, the conditions “would appear” not to be met. Like Company did not put access barriers on its portals. In the dolphin example, the article’s form of expression is not apparent in Gemini’s answer: besides individual words, only one sentence matched (point 204). Any such uses “do not appear to be inherent” to Google’s business model. And Google seems to have appropriate measures in place.
So the headline “chatbots can infringe” and the outcome “this chatbot probably didn’t” are both true. The opinion sets a general rule, then applies it in a way that favours Google on these facts.
Google welcomed the training part. A spokesperson, quoted by the Irish broadcaster RTÉ, pointed to “the Advocate General’s opinion recognising that EU law permits AI training on publicly available information” and said “European law’s text and datamining exceptions are vital for Europe’s competitiveness and AI development”. I found no public statement from Like Company.
An opinion is not a judgment
This step is easy to skip over. Advocates General propose a solution “in complete independence”; the judges have only now begun deliberating, and “judgment will be given at a later date” (press release). The Court often follows its Advocates General, but not always, and there is no reliable figure for how often. The widely repeated “about 80 percent” figure, the legal news service Out-Law (Pinsent Masons) concluded in 2010, “languishes somewhere between rumour and conjecture”. The best-known econometric study, by Arrebola, Mauricio and Jiménez Portilla (2016), looked only at annulment actions, not at references like this one.
On timing: the Court has set no date. In 2025, preliminary references took 16.9 months on average from filing to decision, according to the Court’s own statistics. This reference was filed in 2025, so a judgment in 2027 is plausible. That is my own estimate, not a court date. After the ruling, the Budapest court still has to decide the case itself.
Infographic 02 · From dolphins to judgment
One summary, four questions, no ruling yet
How a Hungarian article about dolphins became the first generative AI case at the EU Court of Justice, and what still has to happen.
- 2019EU creates the press publishers’ rightArt. 15 of Directive 2019/790, aimed at aggregators such as Google News (pt 22)
- Jun 2023 – May 2024Period of the alleged use, per Like Company13 June 2023 to 31 May 2024, as described in the press release
- 21 Jul 2023The dolphin article is publishedOn a Like Company portal: a singer’s plan to bring dolphins to Lake Balaton (pts 17, 49)
- Dec 2023Gemini 1.0 is releasedToo soon after the article for it to be in the training data, the AG finds (pt 49)
- 2025Budapest court refers four questionsTwo on training, two on chatbot answers (pt 20)
- 10 Mar 2026Hearing in LuxembourgNine governments and the Commission had filed observations (pt 21)
- 3 Nov 2026Opinion of Advocate General Szpunar240 points; “the first generative AI case to reach the Court of Justice”
- No date setJudgment of the Court of JusticeNot bound by the opinion. Then the Budapest court decides the case itself
The 16.9 months is the Court’s 2025 average for preliminary references, not a forecast for this case. There is no reliable figure for how often the Court follows its Advocates General.
Sources: CURIA press release No 155/26 (3 Nov 2026), Opinion of AG Szpunar, C‑250/25 (provisional text), CURIA press release No 44/26 (statistics 2025).
Powered by Thorsten Meyer AI
How it fits with the AI Act and the German cases
EU AI Act. The opinion says the AI Act does not apply to this case in time (point 25) and uses it only as context. Its copyright duties for providers of general-purpose AI models, a copyright policy (Article 53(1)(c)) and a public summary of training content (Article 53(1)(d)), cover training, not the live retrieval behind chatbot answers. The Advocate General suggests similar transparency for what chatbots scrape to ground their answers (footnote 242).
Hamburg, Kneschke v LAION. A German photographer sued LAION, a non-profit that builds AI training datasets. The Hamburg Regional Court dismissed the claim on 27 September 2024 (310 O 227/23), and the Hanseatic Higher Regional Court dismissed his appeal on 10 December 2025 (5 U 104/24), holding that compiling such a dataset fell under the text and data mining and research exceptions. The opinion cites the Hamburg case for its reading of the TDM exception and of machine-readable opt-outs.
Munich, GEMA v OpenAI. On 11 November 2025, the Munich Regional Court (42 O 14139/24) largely ruled for the German music rights society GEMA against two OpenAI companies. It found that the lyrics of nine German songs were “reproducibly” contained in the models 4 and 4o, that outputs reproduced them, and that the TDM exception did not cover this. The judgment was not final when the court announced it. The Advocate General cites it (footnote 80) and is more cautious on the “copies inside the model” point. A Court of Justice ruling would bind national courts on how EU law is to be read, so German courts could have to revisit some of these questions.
What this means for companies
If you deploy a chatbot, an AI search feature or an agent that reads the web and answers customers or staff, the opinion is a design brief more than a ban. In plain terms:
- Respect access barriers. Bots that fetch content must honour paywalls and anti-scraping measures. Getting around them is the clearest risk in the opinion (points 188 and 222).
- Synthesise, don’t substitute. Answers that combine several sources, rephrase facts in a new structure and link back are on the safe side of the line (point 234). Answers that rebuild one article closely enough to replace it are not. Output filters that block long matches with retrieved text are named as a reasonable measure (point 222).
- Keep evidence. The opinion suggests courts should ease the burden of proof for rights holders: a sample of similar answers or links back could shift the burden to the provider (point 163). Logging which sources fed an answer, and which filters ran, could matter later.
- License where you rely on news. He sees publishers in a “strong bargaining position” and expects licensing to grow (points 236 and 237). If your product depends on current news, a licence may be cheaper than a dispute.
- Using a provider’s chatbot is not the same as running one. Liability under the publishers’ right sits with the service provider. A company that builds its own customer-facing RAG system over the open web takes on that role.
What this means for publishers
The opinion gives publishers a usable tool, but not the one many wanted. They cannot claim the facts, and summaries that only extract facts remain free. What they get is protection against substitution, against getting around their barriers, and against the idea that an open website means consent. The practical choices are clearer now: put up effective technical measures and machine-readable reservations, document examples where a chatbot reproduces their pieces recognisably, and negotiate. The opinion even hints at a quality argument: in-depth articles are more distinctive than plain reports of events, and so easier to recognise and protect (point 206; footnote 239). I looked at what AI answers are already doing to publisher traffic in an earlier piece on Google’s AI Mode.
What this means for users
For people who use chatbots, the opinion mostly keeps things as they are, with a push in a good direction. It explicitly wants a “breathing space” for chatbots that combine many sources, link to them and urge users to read more (point 234). If the Court follows it, providers have a legal reason to show sources and to avoid answers that simply rewrite one outlet’s story. That is also good for answer quality: the opinion itself notes that copyright does not solve the bigger issues of source selection, diversity and misinformation (point 238). A summary is a summary. For anything that matters, open the link.
Infographic 03 · What it means
A design brief for chatbots, leverage for publishers
What the Advocate General's test would mean in practice, if the Court follows it.
Companies running chatbots or AI search
- Respect access barriers. Paywalls and bot blocks are the clearest line (pts 188, 222).
- Synthesise, don’t substitute. Several sources, new structure, links back (pt 234).
- Filter outputs that match retrieved text too closely (pt 222).
- Log what fed each answer. Courts may ease the burden of proof for rights holders (pt 163).
- License where you depend on news (pts 236–237).
Publishers
- No monopoly on facts: fact-only summaries stay free (pt 208).
- An open site is not consent to chatbot use (pts 225–228).
- Effective technical measures and machine-readable reservations matter (pts 148–149, 188).
- Document recognisable answers and negotiate from a “strong bargaining position” (pt 236).
Users
- Private use is outside the press publishers’ right (pt 218).
- Expect more synthesis and more links, if the Court follows the opinion (pt 234).
- Copyright does not fix source quality or misinformation (pt 238).
- A summary is a summary. For anything important, open the source.
Not every chatbot is the same. The opinion targets what a chatbot ingests and what it produces.Thorsten Meyer AI, reading of points 233 to 235
General information, not legal advice. The opinion is not binding; the Court’s judgment is pending. Point numbers refer to the provisional text.
Sources: Opinion of AG Szpunar, C‑250/25 (provisional text), CURIA press release No 155/26 (3 Nov 2026).
Powered by Thorsten Meyer AI
Honest limits
- This is an Advocate General’s opinion, not a judgment. The Court can follow it, partly follow it or depart from it.
- The text is a “provisional text” on CURIA. Point numbers are quoted from that version and may shift slightly in the final version.
- The guidance on the dolphin case is “subject to” the Hungarian court’s findings. Other facts, for example a paywalled site or a chatbot that reproduces long passages, could produce the opposite result.
- The training part was not needed to decide the case and is explicitly offered only “for the sake of completeness”. Reports that the adviser “ruled training lawful” overstate it.
- There is no reliable figure for how often the Court follows its Advocates General, and no date for the judgment. The 2027 estimate is mine.
- I have not seen the parties’ written submissions; positions attributed to them come from the opinion and the press release.
Sources
- Court of Justice of the EU, Press release No 155/26, “Opinion of the Advocate General in Case C‑250/25 | Like Company”, 3 November 2026: curia.europa.eu (PDF)
- Opinion of Advocate General Szpunar, Case C‑250/25, Like Company v Google Ireland Limited, delivered 3 November 2026 (provisional text, 240 points): InfoCuria
- Court of Justice of the EU, Press release No 44/26, Judicial statistics 2025: curia.europa.eu (PDF)
- Directive (EU) 2019/790 (CDSM Directive), Articles 4 and 15: EUR-Lex
- Directive 2001/29/EC (InfoSoc Directive), Articles 2, 3 and 5: EUR-Lex
- Regulation (EU) 2024/1689 (AI Act), Article 53: EUR-Lex
- Landgericht München I, press release 11 of 11 November 2025, GEMA v OpenAI, 42 O 14139/24: justiz.bayern.de
- Hanseatic Higher Regional Court Hamburg, 10 December 2025, Kneschke v LAION, 5 U 104/24, summary by DLA Piper: dlapiper.com
- RTÉ, “EU court adviser backs publishers in AI case vs Google”, 3 November 2026 (Google’s statement): rte.ie
- Out-Law (Pinsent Masons), “How often does the ECJ follow Advocates General?”, 18 October 2010: pinsentmasons.com
- Arrebola, Mauricio and Jiménez Portilla, “An Econometric Analysis of the Influence of the Advocate General on the Court of Justice of the European Union”, Cambridge Journal of International and Comparative Law 5(1), 2016, as summarised in a European Parliament briefing: statewatch.org (PDF)
