Europe's Top Court Adviser Drew a Line: Training Is Copying, Live Scraping Is Not

On 3 November 2026, Advocate General Maciej Szpunar delivered an opinion in case C-250/25, Like Company v Google Ireland. It is the first generative AI dispute to reach the Court of Justice of the European Union, and the reasoning matters more than the outcome, because the outcome is not binding and may not even be answered.
The case came from the Budapest Regional Court. A Hungarian publisher objected to a chatbot answering questions about its articles, and the referring court sent four questions up: two about using the articles as training data, two about reproducing them in answers. The disputed period ran from 13 June 2023 to 31 May 2024.
Szpunar's split is clean. Training a model on protected works is a reproduction, which means it needs a legal basis. That basis can be the text and data mining exception in Article 4 of the Digital Single Market Directive, provided the works were lawfully accessed and the rightsholder did not reserve rights in a machine-readable way. Copying at answer time is a different act, judged under a different rule.
Why the second half is the dangerous one
A chatbot that reproduces an article in full, close enough that the reader no longer needs to visit the publisher, can infringe the press publishers' right in Article 15. Szpunar went further than the bare act of copying: where that effect is built into the business model, or where there are no effective safeguards against it, the service provider carries the liability.
That sentence is the one to read twice. It makes the design of the product part of the legal question. A retrieval layer that quotes three sentences and links out sits in a very different place from a generator trained to produce a complete substitute reading experience. Same model, same corpus, different exposure.
For anyone shipping a news-adjacent assistant in Europe, the practical work therefore sits in the output layer rather than the training pipeline: how much is reproduced, whether the source gets anything back, and whether you can show a mechanism that stops the pathological case.
What a reservation of rights actually looks like
The Article 4 exception is conditional, and the condition is procedural. A rightsholder can reserve the use of their works, and if they do it in a machine-readable way, the exception does not apply to them.
In practice that means the crawler has to read a signal and honour it, which is a different obligation from the one most teams built for. robots.txt governs access. A reservation of rights governs use, and the two are not the same instruction. A site can allow crawling and still reserve the right to prohibit mining, and a pipeline that only checks the first will believe it has permission it does not have.
The awkward part is the archive. Rights reservations change over time, and a training corpus assembled over two years contains documents that were freely minable in 2024 and reserved in 2026. Nothing in the opinion resolves what happens to the earlier copies. A conservative reading says the dataset has to be re-filtered, which is expensive. A loose reading says the state at crawl time governs, which is convenient and probably not what a rightsholder would agree to.
The court may decline the training questions
Szpunar recommended that the Court declare the two training questions inadmissible, on the grounds that they are hypothetical for this dispute. If the Court follows that advice, the EU will keep its Article 4 exception without any ruling on what "lawfully accessed" means at model scale, or on what counts as a machine-readable reservation of rights.
That is a strange result for an industry that has spent two years asking exactly those questions. It is also a realistic one. Courts answer the dispute in front of them, and this dispute is about answers, not about crawl logs from 2023.
The judgment is expected in 2027. Until then, national courts will keep improvising, and they are already disagreeing.
A quiet disagreement about memory
Szpunar is sceptical of the idea that a model "memorises" training works in a legally meaningful sense. The Munich Regional Court took the opposite view in an earlier case. That conflict is not resolved by this opinion, and it will not be resolved by the eventual judgment either, because memorisation is a factual question wearing legal clothes.
The distinction matters commercially. If memorisation is a legal fiction, then the harm from training is diffuse and hard to prove, and the exception does most of the work. If memorisation is a finding of fact that can be shown with prompting, then every model that can reproduce a paragraph verbatim is a reproduction machine, and the exception starts to look narrower than it did.
Both readings are live in Europe right now. A compliance memo that picks one and assumes the other is wrong is not a compliance memo.
What actually changes this week
Nothing, in the sense that an opinion binds nobody. The Court will rule later, and might rule narrowly.
Something, in the sense that the shape of the argument is now public and it is not the shape most people expected. The fight moved off the training corpus and onto the answer. A publisher's lawyer will spend very little time on whether you scraped. The question that decides the case is whether, after your product finished answering, anyone still needed to click.
Three things follow for teams building in this space.
First, log your reservations. If you rely on Article 4, the machine-readable opt-out is the hinge, and it has to be checked at crawl time and re-checked later. A crawl that respected opt-outs in 2024 and ignored them in 2026 is a single continuous act with a bad second half.
Second, treat quoting as a product decision with a legal shadow. How many sentences, whether the source is named, whether the link is the point of the answer. Those choices are now evidence.
Third, do not wait for 2027 to write the internal policy. National courts are already taking positions, and the ones taking the hardest positions are the ones hearing the smallest cases.
The opinion's most useful line is also its dullest: two acts, two tests. Training copies. Answering republishes. Teams that built one pipeline and assumed one legal theory now have to defend two behaviours with one architecture, and the second behaviour is the one with a plaintiff attached.
Related articles
Google replaced the model behind Search, and the number that moved was a shell benchmark
Gemini 4 Hydrogen became the default on 12 November, taking over from Gemini 3.8 Flash across Search, the Gemini app and the developer API. It runs about 2.5 times faster, scores 57% on Terminal-Bench 4.1 where its predecessor scored 18.2%, and costs $1.50 in and $7.50 out per million tokens.
Three ways a coding agent wastes money, measured on 30 repositories
TokenEconomizer watches what an agent did in a repository you work in often, then stops it from rediscovering the same build command, loading tools it never uses and looping on one error. Across five agent and model combinations it cut mean cost by 18.4% to 30.9%, cheaper in all 30 repository cells, with Pass@1 up in four of five.
earshot gives coding agents 922 sounds, and every one is rendered from code
An agent can write a game, an interface or a film and never hear any of it. earshot answers requests like "three variations of a footstep on gravel" from a catalogue of 922 designs built from public-domain recordings, with no audio files in the repository at all. It runs on CPU, as a CLI, a Python package, an MCP server and a Claude Code plugin.
The agents stealing API keys were not sophisticated. The gateways were open.
Amazon's decoy servers spent October posing as inference APIs and logged agents built on ARTEX pulling provider keys out of LiteLLM, Dify, Langflow and Open WebUI. The way in was a default master key and a half-finished install. ARTEX itself ships with budgets, scope gates and evidence for every finding.