The EU's top court was advised that AI training is mostly lawful but answers can infringe.

Reported, not verified. The Court's own documents refused every fetch today: press release 155/26 at curia.europa.eu and the opinion on InfoCuria both failed. The account rests on PPC Land, which quotes the opinion at length, with mLex and Courthouse News. Verification event: the Court publishes the opinion text, and later its judgment.

Case C-250/25, Like Company Kft. v Google Ireland Limited, is a reference from the Budapest Regional Court in Hungary. Like Company runs advertising-funded Hungarian news portals. Its disputed article, published July 21, 2023, concerned a plan to bring dolphins to Lake Balaton; Gemini's summary invented a pool 25 metres long holding 1,000 cubic metres. The claim period runs June 13, 2023 to May 31, 2024. The March 10, 2026 hearing drew observations from both parties, the European Commission and 9 governments. Advocate General Maciej Szpunar delivered his opinion on November 3, 2026: 240 numbered points, 243 footnotes, far longer than usual, more than 90 points on training alone. The Hungarian court referred 4 questions, 2 on training and 2 on answers. On training, Szpunar advises declaring the questions inadmissible: the article predates the models' training, and training is not online use under Article 15(1) of Directive 2019/790. Answering anyway: training does involve reproductions under Article 2 of the InfoSoc Directive, and the text and data mining exception in Article 4(1) in principle applies, conditional on lawful access and machine-readable opt-outs under Article 4(3). He names robots.txt and Google-Extended, released September 28, 2023, as qualifying, and writes that a trained model is not a database. Where a model routinely regurgitates, the exception should be refused absent mitigation measures. On answers: retrieval is online use, copying occurs through it, and protection covers the form of expression, not facts. The test is whether the answer substitutes directly, so the reader need not visit the original. Circumventing a paywall or bot block infringes. Making available to the public needs recognisability amounting to substitution, which he calls rather unlikely given non-deterministic outputs, though possible through negligent retrieval design. He rejects Google's implied-consent argument, citing VG Bild-Kunst, and proposes a presumption shifting the burden to the provider once a rightholder shows a reasonable sample of resembling answers; one example rarely suffices. On the facts he is unfavourable to Like Company: a purely informative report, no technical measure apparently circumvented, Google apparently mitigating. Judgment comes at a date not set. Not established: how the Court will rule, and whether Like Company's portals are press publishers.

1. Jun 2023

   Like Company publishes the disputed article on July 21, inside a claim period opening June 13. Google releases Google-Extended, the crawler opt-out the opinion later treats as valid, on September 28, and Gemini 1.0 in December.

2. Mar 10

   The Court hears Case C-250/25. The Commission and 9 governments file observations, an unusually crowded hearing for a reference from one Hungarian regional court.

3. Oct 30

   Anthropic files an internet specification for a single machine-readable AI training opt-out signal. Edition 067. The opinion delivered 4 days later makes machine-readable opt-outs the condition on which the training exception turns.

4. Oct 30

   Britain's competition regulator opens an abuse case over Meta's terms for publishers' content. Edition 067. Publishers press the same grievance through competition law and copyright at once.

5. Nov 3

   Advocate General Szpunar delivers a 240-point opinion: training in principle inside the text and data mining exception, chatbot answers infringing where access was unlawful or the answer substitutes for the article. Advisory only.