LexBeat in Focus: Who Owns the Answer?

1LexBeat is the column in the German Legal Tech Summit magazine that summarises and puts into context the key developments in the legal market. LexBeat in Focus takes an in-depth look at a single topic.

2So far, the case has been discussed mainly as a media and copyright story. It is just as much a legal tech story, because almost every legal AI tool retrieves third-party content in real time and turns it into answers.

The case

3Like Company, a Hungarian media group, operates several news portals. It accuses Google of using its articles without consent in two ways: to train the language models behind the Gemini chatbot, and to generate answers for users. Gemini relies on retrieval-augmented generation (RAG): the chatbot retrieves content from the web in real time, including press articles, and uses it to build its answers. Like Company does not rely on copyright but on the EU press publishers' right under Article 15(1) of Directive (EU) 2019/790 on copyright in the Digital Single Market (DSM Directive).¹

4The Hungarian court stayed the proceedings and referred four questions to the Court of Justice. Two of them concern training: is training a reproduction reserved to press publishers, or is it covered by the text and data mining (TDM) exception? The other two concern the chatbot's answers: do they engage the rights of reproduction and of making available to the public?²

5The reference reached Luxembourg in April 2025. On 10 March 2026, the Grand Chamber held the Court's first-ever hearing on generative AI and copyright. The opinion, originally scheduled for September, was delivered on 3 November.³ The list of participants shows how fundamental the case is: besides the parties and the European Commission, nine Member States submitted observations, including Germany. Szpunar himself notes that his opinion is considerably longer than usual.⁴

What the Advocate General says

6Training: permitted, as a rule. Szpunar advises the Court to declare the two training questions inadmissible as hypothetical. There is no evidence that Like Company's articles were used to train Gemini, and training a model is not an "online use" of press publications, which is all the press publishers' right protects.⁵ For the sake of completeness, he adds that using protected works and subject matter to train a language model is, as a rule, permitted under the TDM exception.⁶ That freedom comes with conditions: the commercial exception in Article 4 of the DSM Directive (implemented in Germany as section 44b of the Copyright Act) requires lawful access and does not apply where rightholders have reserved their rights in a machine-readable way.⁷

7Answers: this is where it gets tight. Unlike training, Szpunar treats RAG as an online use of press publications, and one that can involve legally relevant reproductions.⁸ At the same time, he draws clear limits. The press publishers' right protects the specific journalistic and linguistic expression of an article, not the facts or news it reports. Whether an answer infringes depends on two things: whether the chatbot accessed the article lawfully, without circumventing paywalls or anti-scraping measures, and what the answer contains. Merely reproducing facts is generally unproblematic. It is different when the answer recognisably takes over an article's characteristic form of expression so extensively that the user no longer has a reason to read the original.⁹

8Liability: on the provider. Where the expression of press articles systematically and substantially resurfaces in a chatbot's answers, Szpunar considers the right of making available to the public to be infringed as a rule. The provider is liable if such uses are an integral part of its business model, or if it failed to take appropriate measures to prevent infringements.¹⁰

Why this is a legal tech story

9At the heart of the Advocate General's analysis of the answers lies one technique: retrieval-augmented generation. That is precisely the architecture on which almost every serious legal AI tool is built. A legal AI assistant is only as good as the sources it retrieves at the moment of the query: statutes, case law, commentaries, journal articles. The model writes; the content layer makes the answer reliable.

10The market has organised itself around that content layer:

• OpenAI built a dedicated search index of case law and statutes for Astra for Law.¹¹
• Wolters Kluwer has integrated its German content into Libra since January 2026: legislation, case law, commentaries, literature and journals from Wolters Kluwer Online, plus selected titles from the publisher Dr. Otto Schmidt, with citations and source traceability.¹²
• Beck-Noxtua uses exclusively legal content from C.H.Beck's beck-online database and returns results with beck-online citations.¹³ The underlying Noxtua platform also works with content from MANZ in Austria, Swiss Legal Publishing and Ciela in Bulgaria.¹⁴
• Generic assistants and many smaller tools, by contrast, retrieve from the open web, much like Gemini in the case before the Court.

11This is why Like Company matters for legal tech. The question the Court has to answer — when does an AI answer built from retrieved content cross the line from using information to taking over someone else's expression? — is the same question every legal AI product faces each time it summarises a commentary or a journal article. And the answer will determine what the content layer is worth.

Where existing protections fall short

12If the case is so relevant, why has it attracted so little attention in the legal tech community? Partly because the right at its centre does not cover much of the content that legal AI relies on. On closer inspection, most of the existing legal instruments work less well for legal publishers than for news publishers:

• The press publishers' right does not cover legal journals. The DSM Directive expressly excludes periodicals published for scientific or academic purposes from the definition of press publications.¹⁵ Many legal journals are likely to fall on that side of the line.
• Primary law is free. In Germany, statutes, court decisions and official headnotes are official works without copyright protection.¹⁶ As far as copyright is concerned, legal AI tools may retrieve and reproduce them freely.
• Copyright protects expression, not legal reasoning. Commentaries, articles and handbooks are protected works. But an AI answer that paraphrases a legal argument in its own words takes over the reasoning, not necessarily the expression. That is exactly the line Szpunar draws for press articles, and it is even harder to police for legal analysis.
• The opt-out targets the wrong stage. Rightholders can reserve their works against text and data mining in machine-readable form. The Higher Regional Court of Hamburg has held that a reservation written in natural language does not automatically meet that standard.¹⁷ But the opt-out governs training and analysis. Szpunar treats RAG answers as a separate online use, assessed under the rights of reproduction and making available.
• The chain of rights is contested. To use authors' works in AI products, publishers need the corresponding rights. C.H.Beck has asked its academic authors to sign AI supplements to their publishing contracts. The copyright scholar Michael Grünberger has publicly refused, arguing that it is doubtful whether existing contracts already cover AI uses and criticising the extension of the publisher's market power into AI.¹⁸

13The most effective protection for legal content is therefore not a special right but something more mundane: the paywall and the contract. Commentaries in beck-online or Wolters Kluwer Online sit behind logins. Under Szpunar's test, a tool that circumvents such access controls fails the first condition.

14The United States shows how differently this can play out. On 29 September 2026, the Third Circuit ruled in Thomson Reuters v. ROSS Intelligence that Westlaw's editorial headnotes are protected by copyright and that using them to train a competing AI legal research tool was not fair use. It was the first federal appellate decision on fair use in AI training.¹⁹ The court also recognised a developing licensing market for editorial content as AI training data.²⁰ In the EU, by contrast, training on such content may well be covered by the TDM exception, unless rightholders have effectively opted out.

The publishers' play: licensed AI

15Legal publishers have understood that their content is the scarce resource in legal AI. Three licensing models are emerging.

161. The walled garden: Beck-Noxtua. C.H.Beck bundles an AI workspace with its own database. Beck-Noxtua PLUS costs €299 per licence and month and includes content from eight areas of law with around 40 commentaries. PREMIUM costs €499 and covers eleven areas with more than 360 commentaries, over 200 handbooks and around 130 journals.²¹ Beck stresses that only Beck-Noxtua has direct access to beck-online, that hosting is sovereign in the EU, and that user data is not used for model training.²² Beck's message to authors shows how closely content and AI are now linked: according to Grünberger, the publisher argues that AI-optimised works are cited more often as sources and clicked more often, which leads to a higher share of royalties.²³

172. Acquire and integrate: Libra by Wolters Kluwer. Wolters Kluwer bought the Berlin start-up Libra in November 2025 for up to €90 million, €30 million of it up front.²⁴ Since January 2026, its German content has been available inside Libra.²⁵ The pricing is modular: according to a comparison by the competitor Lulius, primary sources are included in every Libra licence, while commentaries come as a paid content module.²⁶

183. Licensing to AI platforms. The third model brings publisher content to independent AI vendors. Harvey and LexisNexis have built joint agentic workflows that combine Harvey's platform with LexisNexis content.²⁷ Thomson Reuters is one of the launch partners of OpenAI's Astra for Law.²⁸

19The Advocate General's opinion strengthens all three models. If retrieving content without a lawful basis and reproducing its expression in AI answers creates liability, licensed access becomes a legal asset as well as a quality feature. The flip side is a concentration risk: whoever controls the commentaries controls the most reliable legal AI. For a market in which a handful of publishers hold the leading commentaries, that is a competition question in its own right.

Responsibility along the chain: still open

20A legal AI answer typically passes through several hands: the developer of the underlying model, the legal tech vendor that builds the RAG pipeline and selects the sources, the publisher that licenses the content, and the law firm or legal department that uses the output.

21As noted above, the Advocate General's opinion addresses only one of these roles: the provider of the chatbot. He considers the provider responsible where infringing uses are an integral part of its business model or where it has failed to take appropriate measures against infringements.²⁹ In GEMA v OpenAI, the Munich Regional Court rejected OpenAI's argument that the user who writes the prompt is responsible for the output³⁰ and held OpenAI responsible instead.³¹

22How these approaches apply to the multi-layered structure of legal AI products has not yet been decided by any court. The answer will be shaped by future case law and by the contracts between model developers, vendors, publishers and users.

What comes next: cases, dates, legislation

23The opinion opens a dense phase of court decisions. These are the dates to watch:

• 17 December 2026: Federal Court of Justice, Kneschke v LAION (I ZR 281/25). The court heard Germany's first case on creating datasets for AI training on 3 September and has scheduled its decision for 17 December.³² According to reports on the hearing, a referral to the Court of Justice appears likely.³³
• Pending: Higher Regional Court of Munich, GEMA v OpenAI (6 U 3662/25 e). OpenAI's appeal against the Munich memorisation ruling has been pending since December 2025; no hearing date has been published yet.³⁴
• Not yet final: Munich Regional Court, GEMA v Suno (42 O 763/25). On 31 July 2026, the same chamber extended its line to the AI music generator Suno.³⁵
• 2027: Court of Justice, Like Company v Google (C-250/25). A judgment is not expected before 2027.³⁶

24On the legislative side, Brussels is moving:

• European Parliament. On 10 March 2026, Parliament adopted a non-binding resolution on copyright and generative AI, drafted by Axel Voss, by 460 votes to 71. It calls on the Commission to examine fair remuneration for past uses of protected works by providers of general-purpose AI models.³⁷ MEPs also want EU copyright to apply to all generative AI systems on the EU market, regardless of where training took place, and want rightholders to be able to opt out of training and automated crawling.³⁸
• European Commission. In its July 2026 follow-up to the resolution, the Commission states that it is preparing the review of the DSM Directive.³⁹ That review is where the future of the TDM exception, the opt-out and possible remuneration models will be decided.
• AI Act. For providers of general-purpose AI models, the copyright obligations of Article 53 already apply, and since August 2026 the Commission has been able to enforce them with fines.
• Germany. We have found no specific German bill on AI and copyright. For now, Berlin is waiting for Luxembourg and Brussels, while the courts in Munich, Hamburg and Karlsruhe set the pace.

What is likely to happen

25The opinion does not settle the law, but it sets the direction. Several developments seem likely.

26The Court will probably follow the Advocate General on the answers. Szpunar's test for RAG, lawful access plus recognisable takeover of expression, is cautious and leaves room for facts and summaries, which makes it easy for the Court to adopt. Whether the Court will also address training is less certain. If it follows Szpunar on inadmissibility, the training question will move to the next case, possibly a referral from the Federal Court of Justice in the LAION proceedings.

27Licensing will become the norm. The Third Circuit has already recognised a licensing market for editorial content as AI training data.⁴⁰ The European Parliament wants fair remuneration and voluntary collective licensing.⁴¹ Szpunar's opinion adds a liability argument for licensing at the point of retrieval. More deals between publishers and AI vendors are likely, as are rising prices for licensed legal content, while vendors without such deals will concentrate on freely available primary law.

28Product design will change. Legal AI answers will probably become more citation-heavy: shorter summaries, more precise references, direct links to the source and fewer long verbatim passages. This will be less a matter of style than of legal risk management.

29Contracts will be renegotiated. Copyright indemnities and content warranties are likely to become standard clauses in legal AI procurement, along the chain from model developer to vendor to law firm. At the same time, the AI Office will look more closely at how providers of general-purpose models implement their copyright policies under Article 53(1)(c) of the AI Act.⁴²

30The conflict will shift to authors and competition. The more valuable licensed content becomes, the more intense the debate about authors' share will be, as the controversy over Beck's AI contract supplements already shows. And the stronger publishers' bundles become, the more likely it is that competition authorities will take an interest in who controls access to the leading commentaries.

31The rules for the content layer of legal AI are being written now, in courtrooms and in the DSM review. Many of these questions will also be on the agenda of the German Legal Tech Summit in Hannover on 3 December.