The Week AI Rules Stopped Arguing About Training
On Tuesday, an adviser to Europe's highest court said the part out loud. Training a model on your articles is, as a rule, allowed. The answer it builds out of them may not be.
0:00 / 5:09
The Week AI Rules Stopped Arguing About Training
Europe's highest court put a test on paper this week, and the test is about the answer rather than the training set. Meta took its disputed commercial terms off the web altogether, two days after a judge in São Paulo barred them. And The New York Times sued not a model maker but the company that sells the models their pages. None of the three settles the law for everyone. Here is what each one does.
The first generative AI case at the EU court
On Tuesday 3 November, Advocate General Maciej Szpunar delivered his Opinion in Like Company v Google, Case C-250/25, a reference from the Budapest Regional Court. It is the first generative AI case to reach the Court of Justice of the European Union. 1
Like Company runs Hungarian news portals. It claims the right European law gives press publishers over the online use of their publications — Article 15(1) of the 2019 directive on copyright in the digital single market. The dispute covers Gemini's use of its articles between 13 June 2023 and 31 May 2024.
The Hungarian court asked four questions. The advocate general advised the Court to declare the two about training inadmissible. Nothing in the file shows that Like Company's articles were used to train the Gemini models, he wrote, and training is not a form of online use of a press publication. For completeness he gave a provisional view anyway. Training reproduces works, and the text and data mining exception in Article 4 of the directive covers those copies as a rule. That holds where the material was lawfully accessible and the right holder had not expressly reserved its use. He added that a developer relying on the exception should take steps to limit the risk of a model reproducing protected material closely in its output, and that national courts should weigh whether it did. 12
The live questions were the other two, about the answers Gemini gives. The advocate general proposed a test with two limbs.
The first limb is how the machine got in. Where the publisher had effective technical measures restricting access to or scraping of the publication — a paywall, or a measure aimed at bots — the chatbot should not have circumvented them. Circumvention is what makes the access unlawful. The second limb is what came back out. The press publisher's right protects the specific form of expression, not the facts. Whether an answer is a reproduction turns on whether the protected elements are recognisable in it on a side-by-side comparison, to the point that the reader is entirely dispensed from opening the original. The right of making available is, as a rule, infringed where the form of expression is systematically recognisable across answers to related queries. 1
On liability, the advocate general said the provider answers for those acts of use where they are inherent to its business model. The provider also answers where it has failed to put measures in place to counter them credibly and effectively. He rejected Google's argument that leaving an article freely available amounts to consent. The Opinion is not binding, and the Court has not set a date for judgment. 13
What it changes for a newsroom: the technical measure stops being a product decision and becomes the fact the legal test turns on. A paywall a crawler respects, or a bot block that is actually enforced, separates an unlawful taking from a permitted one.
Meta's terms are off the web
Meta took down the preview page for its proposed Commercial Terms, Product Catalog Terms, Self-Serve Ad Terms and Ad Creative Generative AI Terms. It posted a notice in its Business Help Center. The updates "are now paused and will not take effect on November 20, 2026", the notice says; Meta is continuing to consult stakeholders, and "our existing terms remain in effect". The notice adds that crawling practices are unchanged and that content owners can continue to signal preferences in
robots.txt. 4"We're pausing the changes as we continue to engage with stakeholders and take our time to get this right." — a Meta spokesperson 5
The Brazilian order came first. On Tuesday 3 November, the 44th Civil Court in São Paulo granted a preliminary injunction sought by the National Newspaper Association, the Brazilian Association of Radio and Television Broadcasters and the National Association of Magazine Publishers. Judge Guilherme Madeira Dezem suspended the provisions of the new terms. The judge also barred Meta from collecting and using the outlets' journalism without authorisation, including to develop and train AI systems. The fine is BRL 1 million a day, capped at BRL 100 million. Meta must keep the outlets' accounts and tools operational if they refuse the terms. The order holds until a final judgment, and Meta declined to comment to Valor. 6
Norway's court blocked the same terms on 28 October, outside this edition's window. Scope matters here. The São Paulo injunction protects the outlets those three associations represent, and it binds nobody else. For every other publisher, the terms that govern the account are the ones already signed.
The Times sues the company that supplies the pages
On Monday 2 November, The New York Times Company filed The New York Times Co. v. Brave Software, Inc. in the Southern District of New York, No. 1:26-cv-09549. The complaint runs to 72 pages and demands a jury. The Daily News filed separately the same day. 78
Brave operates a browser and a web index, and sells access to that index to AI companies. The complaint names Anthropic, Perplexity, Mistral, xAI and Cohere as recipients. None of the five is a defendant.
"Brave acts both like a burglar that misappropriates the content of publishers, and an AI 'fence' that sells stolen content to third parties." — the complaint 7
The five counts are direct copyright infringement, contributory infringement, vicarious infringement, circumvention of technological measures under the Digital Millennium Copyright Act, and unjust enrichment under New York law. Exhibit A lists more than 10 million registered works.
The complaint alleges that Brave has never disclosed the name of its crawler, so site owners had no way to block it. It also alleges that the browser ships a "Paywall blocker" setting Brave describes as allowing users to "Bypass Paywalls", with Times articles as the worked example. A further count concerns the advertising Brave removes from Times pages and replaces with its own. 7
"Brave has flouted established copyright law and ignored repeated requests by The Times to stop its infringing conduct." — Graham James, a New York Times spokesperson 8
Spokespeople for Brave, Anthropic and xAI did not immediately respond to Reuters. What is new is the target. The cases of the past three years were aimed at the companies that build models. This one is aimed at the company that supplies them the pages.
Dates to hold
- 20 November 2026 — replies due on the cross-motions for summary judgment in the consolidated OpenAI copyright case in Manhattan. It was also the date Meta's paused terms would have taken effect; Meta now says they will not.
- No date — judgment in Like Company (C-250/25). The Opinion binds nobody until the Court rules.
- Until final judgment — the São Paulo injunction against Meta's terms holds, and so does the Norwegian one of 28 October.
- 15 December 2026 — Bing's temporary grounding-transition preferences open in Bing Webmaster Tools for sites on
NoArchiveorNoCache. - 1 February 2027 — Microsoft begins honouring the new AI Training and AI Grounding preferences as it recrawls.
- 1 March 2027 — deadline for sites on the old tags; after it,
NoArchiveandNoCachestop being read for training or grounding.
References
- 1
- 2
- 3Wolters Kluwer Copyright Blog, first reflections on AG Szpunar's Opinion in Like Company, 3 November 2026
legalblogs.wolterskluwer.com
- 4
- 5
- 6Valor International, Brazilian court blocks Meta from using news content to train AI, 4 November 2026
valorinternational.globo.com
- 7
- 8
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.