Executive Summary

On March 10, 2026, in the Grand Courtroom of the CJEU in Luxembourg, a fifteen-judge Grand Chamber took its seat. The Grand Chamber is the court's highest formation, convened only for cases the CJEU considers especially significant or of fundamental principle. The case that came before it, in six hours of oral argument, was Like Company v Google Ireland, C-250/25 — the first time Europe's top court squarely examined generative AI and copyright.

The claimant, Like Company, is a Hungarian media publisher running several news portals. The defendant is Google Ireland, which operates Gemini (formerly Bard). The case reached Luxembourg through a preliminary reference from a Budapest district court — the procedure by which a national court asks the CJEU what a piece of EU law means, then applies the answer to the dispute in front of it. That mechanism is the reason this hearing does not stop at Hungary: its outcome applies with equal weight to courts across the entire EU.

Three core questions — and where the data question sits

The Budapest court referred three core questions (some sources split the TDM sub-issue into two and count four). The first concerns the right of communication to the public: when a chatbot returns output that partly matches a protected article, does that amount to communicating the work to the public? The second concerns the reproduction right: is LLM training, which observes patterns and adjusts to them, an act of reproduction under EU copyright law? And the third: if training does involve reproduction, can that reproduction fall within the TDM exception in Article 4 of the DSM Directive (2019/790)?

This report focuses on the third. Where the first two ask whether training collides with copyright at all, the third asks how wide the door is that could make the collision lawful. That door is the TDM exception. Fixing the width of the door goes beyond deciding one infringement case: it draws the boundary of what qualifies, across all of Europe, as lawful training data.

One more thing. March 10 was not only the day of the CJEU's first hearing; it was also the day the European Parliament adopted a resolution on generative AI and copyright. The judiciary and the legislature seizing the same subject on the same day reads less like coincidence and more like a signal — that the institutional center of gravity on this issue is shifting toward Europe.

To understand the TDM exception, you first have to see how Europe designed its copyright exceptions. The DSM Directive splits text-and-data mining into two provisions. Article 3 covers mining done by research and cultural-heritage institutions for scientific research, and rightholders cannot exclude it. Article 4 broadly permits all mining, including commercial purposes — but any work whose rightholder reserves it in a machine-readable form drops out of the exception. Whether commercial LLM training falls inside Article 4 is the crux of C-250/25.

How a 'machine-readable opt-out' actually works

The weight of Article 4 rests on that one condition: a "machine-readable reservation." To keep their content out of training, rightholders cannot simply say no in prose the way they would to a person; they have to mark it in a form a machine can read. In practice, that reservation is expressed through robots.txt, metadata, terms of use, and standards such as TDMRep, ai.txt, and C2PA. The EUIPO catalogs roughly eight such techniques. The signaling tools themselves are settling in fast: 57.2% of France's top 250 sites have already implemented TDMRep.

Yet the design has several practical fractures. robots.txt can be bypassed, and typos or misconfigurations are common. The standards are fragmented across several tracks, so it is unclear which one to trust. Above all, the boundary of what counts as "machine-readable" is still being hammered out in case law. The Hamburg appellate ruling we'll see below held that a reservation written in natural language alone is not machine-readable. Both those who signal and those who scrape are still feeling their way toward where a valid opt-out ends.

How firmly opt-out is nailed into law varies by member state. That means the very same work can be a valid or an invalid reservation depending on whose national law you apply — and for anyone handling data across 27 countries, that is that many more status values to verify. The breakdown below groups EU member states by how far they have written machine-readable opt-out into law.

One more norm sits on top of all this. In Recital 105 and Article 53, the EU AI Act requires that any general-purpose AI (GPAI) provider placing a model on the EU market respect Article 4 TDM opt-outs and publish a summary of its training content. This obligation does not ask where the training happened. Train outside Europe, but commercialize the model inside Europe, and Europe's rules apply. The boundary of the TDM exception does not stop at Europe's borders — it reaches outward.

For most readers, the AI-copyright story is a story of U.S. lawsuits. The New York Times sued OpenAI; authors won a $1.5 billion settlement from Anthropic. The American approach is fair use. Train first, and if it becomes a problem, a court weighs the legitimacy after the fact by balancing four factors: the purpose and character of the use, the nature of the work, the amount used, and the effect on the market. The outcome stays uncertain until the litigation ends.

Europe flips the timing. Not after training, but at the moment of collection, legality is decided in advance by each work's machine-readable opt-out status. It is an ex-ante rule, not an ex-post balancing. Facing the same "AI training and copyright" problem, the two systems place the burden at different moments and on different parties.

How large the American "settle-it-later" world has grown is a matter of numbers. By the Copyright Alliance's count, U.S. AI-copyright suits have passed 70, and Bartz v Anthropic closed with a settlement covering roughly 500,000 works at $3,000 each — $1.5 billion in total, the largest copyright settlement in U.S. history. That $3,000 figure is now hardening into a benchmark for future damages negotiations. The world that prices things through litigation has, in fact, started to put a price on it.

Europe draws that boundary outside the courtroom, through a rule, and it draws it first. A defense after the fact is only needed once a suit is filed; proof beforehand is needed from the first day you scrape. From a data-practice standpoint the difference is decisive, because the center of gravity moves from "defend if a problem arises" to "prove it from the moment you collect."

There is no judgment in C-250/25 yet, but Europe's lower courts have already handed down several. Lined up in chronological order, they reveal the terrain the Grand Chamber now faces. Below are the major European rulings on the TDM exception, and this case's timeline.

Member states line up; the Commission stays cautious

Several member states filed observations at the hearing. The confirmed five (Hungary, Denmark, Greece, Spain, and France) broadly backed a broad, extraterritorial reading: that training and deployment are one integrated process, and that even training done outside the EU triggers EU copyright law once the model is commercialized inside the EU. The European Commission, by contrast, suggested the reference might be partly or wholly inadmissible, on the ground that the questions focused only on Gemini's functionality without identifying a specific infringing act.

So the lower-court trend and the member-state positions converge on a single direction: the exception's boundary runs up to the data-collection stage, and memorization and commercialization sit outside it. There is, of course, no guarantee the Grand Chamber will confirm that trend as is. But whichever way the ruling lands, the fallout splits. If Like Company wins, obtaining licenses becomes the de facto default for training inside Europe; if Google wins, the duty to respect opt-outs still remains — because AI Act Article 53 demands a published training-content summary and a copyright policy independent of the ruling, with enforcement beginning August 2, 2026.

Translate the hearing and the case law into the language of data practice, and one sentence remains. For data to be lawfully trainable in Europe, you have to be able to prove its provenance, its licenses, and its opt-out status. A dataset you cannot prove becomes a liability for European deployment regardless of which way the ruling goes. In a world of ex-ante rules, failing to answer "when, and in what status, did we obtain this data" is itself a defect.

The Munich ruling's "memorization equals leaving the exception" principle adds a second implication here. If a model that memorizes training data wholesale and regurgitates it verbatim leaves the exception, then lowering the probability of memorization is the same as lowering legal risk. Data that is heavily duplicated and dense with copyright raises the probability of verbatim reproduction. So deduplication, source filtering, and license tagging (data curation) become compliance tools at the same time as quality work. The characteristics of the training data run through the model's interior and out into the courtroom as evidence.

This demand is not Europe's alone. Because Article 4 and AI Act Article 53 apply extraterritorially to models placed on the EU market, a Korean or American company faces the same question the moment it ships an AI product into the EU. A GPAI provider must publish a summary of its training content, and a violation can draw a fine of up to 3% of global turnover or €15 million, whichever is higher. Enforcement begins August 2, 2026. "Prove the opt-out, license, and provenance status of your training data" enters as a precondition of European deployment.

Official rulings & policy

• 1.Court of Justice of the European Union. Like Company v Google Ireland Ltd., C-250/25. curia.europa.eu / InfoCuria.
• 2.EU IP Helpdesk. (2026). "First CJEU hearing on generative AI and copyright." European Commission.
• 3.Landgericht München I. (2025, November 11). GEMA v OpenAI.
• 4.Landgericht / Oberlandesgericht Hamburg. (2024–2025). Kneschke v LAION (310 O 227/23 · 5 U 104/24).
• 5.Rechtbank Amsterdam. (2024, October 30). DPG Media v HowardsHome.
• 6.European Union. EU AI Act, Article 53 & Recital 105; DSM Directive (EU) 2019/790, Articles 3–4.

Law-firm & academic briefings

• 7.Bird & Bird. (2026). "Like Company v Google: CJEU holds first-ever hearing on generative AI and copyright." Bird & Bird.
• 8.Taylor Wessing. (2025). "GEMA, Getty and beyond: AI copyright litigation in the EU and the UK." Taylor Wessing.
• 9.European Copyright Society (CREATe). (2026). Comment on C-250/25 (Grand Chamber). Springer IIC, 10.1007/s40319-026-01717-6.
• 10.ODIPI. (2025). "Like Company v. Google Ireland Ltd (C-250/25)." ODIPI.

Policy & data

• 11.EPRS. (2025). "AI and Copyright: The Training of General-Purpose AI." European Parliamentary Research Service (EPRS_ATA(2025)769585).
• 12.Press Gazette. (2026). "How many news websites block AI crawlers." Press Gazette.
• 13.Copyright Alliance. (2025). "Copyright and AI Lawsuit Tracker." Copyright Alliance.
• 14.NPR. (2025, September 5). "Anthropic reaches $1.5 billion settlement with authors." NPR.

※ C-250/25 is at the oral-hearing stage as of March 10, 2026; no judgment has been issued (Advocate General opinion expected September 3, 2026). Some figures — market sizing, blocking rates — are estimates based on commercial research and industry tallies and vary with sample and definition. Originally surfaced via: aimadetools.com.