“Hungarian Dolphins”
In summary: Chatbots can breach EU copyright
The EU Advocate General’s Opinion in Like Company v Google (C-250/25) (“AG Opinion”) finds that Gen AI likely violates news publishers' EU copyright law rights when it reuses an article's distinctive expression so extensively that its answer substitutes for consumers reading the original.
What can website owners do? “Rights Reservations” in the Robots.txt file and written T&Cs.
The case reviews the fact that there is a text and data mining (“TDM”) exemption in EU law. The case advanced by Googe was that its Gemini AIO and AI Mode were entitled to use the content of websites and were protected by the text and data mining exemption. The AG disagrees: he states
“That is another essential aspect of the fair balance inherent to the TDM exception. While, by default, rightholders whose content is ‘lawfully accessible’ online have to tolerate the scraping and use of that content by third parties for text and data mining purposes, including LLM training, they may reserve such use and ‘opt out’ of the TDM exception. This gives rightholders a means of controlling that use and, in particular, licensing it in exchange for appropriate remuneration.” (para 144)
The AG discusses that rightsholders have various “appropriate means” of reserving their rights
“The rather plain Robot Exclusion Protocol is widely used in that regard. (153) Google’s own suggested method of reservation, namely ‘Google Extended’, rests on that protocol. (154) Nevertheless, any appropriate means may be used by rightholders, (155) such as other protocols, (156) including the ‘metadata’ embedded in individual files or even, as envisaged in recital 18 of the CDSM Directive, ‘the terms and conditions of a website’. On the last point, I do not exclude that reservations made in natural language in those terms and conditions could be regarded as appropriate within the meaning of Article 4(3) where they are ‘machine understandable’, as described in the preceding point (which does not seem that far-fetched given the current natural language processing capabilities of LLMs). (157) Obviously, it would be ideal for all the stakeholders to have a limited number of standardised machine-readable formats to express such reservations. Nonetheless, processes for adopting such standards are already underway, and will, in all probability, succeed. (158)”
Preiskel & Co LLP advised Movement for an Open Web on the Search Only Contract, which is embedded in the rightsholder’s robots.txt file and its website T&Cs, which is strongly endorsed by the AG’s guidance.
The AG Opinion also addresses other key points and acts as guidance to the EU’s highest court, the European Court of Justice (“CJEU”), which is due to provide judgment in late 2026 or 2027.
The case
Like Company, a Hungarian news publisher, sued Google Ireland in April 2025 where Gemini had used a Like Company story about a Hungarian singer seeking to release freshwater dolphins into Lake Balaton. Its articles were used, without its consent, to train the Gemini LLMs and during the RAG process when answering user queries, and that Gemini's answers both 'reproduced' and 'made available to the public' those articles.
The Hungarian court deferred four questions to the CJEU regarding whether (1) large language model (“LLM”) training constitutes reproduction (under Article 2 of the InfoSoc Directive), (2) such reproduction falls within the scope of the press publisher’s exclusive rights under Article 15(1) of the CDSM Directive, (3) the text and data mining exception in Article 4 of the DSM Directive can apply, and (s) chatbot output constitutes a communication to the public. The CJEU hearing took place in March 2026.
The Opinion separates two activities:
- Training: processing large amounts of text to develop a language model.
- Retrieving information while answering: looking up online articles in real time to help respond to a user. This is called 'retrieval-augmented generation', or RAG.
The AG also insists that although LLMs, chatbots and even Google Search are interconnected in different ways, they are “fundamentally distinct technologies and services” (para 32). “An LLM is an ‘upstream’, foundational piece of AI technology which has the capability to ‘understand’ and generate text in ‘natural language’” and an “online chatbot is a ‘downstream’ AI system using that technology to function as a conversational parter with its users”.
RAG allows LLM-powered chatbots be more accurate and reliable by finding content on the World Wide Web (essentially, by using Google Search) and this technology is also behind ‘AI Overviews’ and ‘AI Mode’ (para 155 and 156).
Key points that have been made clear:
(a) Evidence and scope. The training claim fails on the facts because the only evidence was that Gemini was able to generate a summary of the article. An LLM takes months to train so this evidence does not show that Like Company's articles were used to train Gemini but rather that the “chatbot retrieved that article in real time from Like Company’s web portal and reused it to provide the chatbot’s answer” (paras 47-50).
Article 15 of the 2019 Copyright (CDSM) Directive covers only “online use of press publications by information society service providers”. A chatbot qualifies as an “information society service” but LLM training does not. The link between LLM training and the chatbot’s output is too indirect (para 51-57).
The AG therefore treats the training questions as 'hypothetical' and inadmissible (para 58) but addresses them anyway.
(b) Training outside the EU: keep it in the EU. The Gemini LLMs were trained essentially in the United States (para 60) and EU law to extend to the US would be counter to the principle of territoriality (like in the UK’s Getty Images v Stability AI).
But the AG accepts the German Government's middle ground: EU law may cover the scraping of works from servers located in the EU, wherever the training later takes place (paras 70–73) but this solution isn’t perfect and needs to be addressed by new laws.
Training involves ‘reproduction’ under Article 2 of the InfoSoc Directive of publisher material (para 75) but near-identical outputs should be determined on a case-by-case basis (para 89) and reproductions are not stored within the model itself (para 92). This interpretation still allows rightsholders to protect their rights.
(c) AI scraping is not like human reading. 'LLMs "learn" their parameters by "consuming" the creative expression contained in the works used… This is not comparable to a human learning to write by reading the works of others, if only because of the industrial scale of the process.'
(d) No temporary copies defence for datasets. Training datasets are not “transient or incidental” and have “considerable "economic significance"' of their own – “many entities now specialise in the creation and marketing of such datasets” (para 96).
(e) The TDM exception applies, but subject to opt-out. The text and data mining (“TDM”) exception in Article 4 CDSM can in principle cover LLM training (para 99) and the opt-out is an essential part of the fair balance (para 144). But it only applies where content is “lawfully accessible” and its use “has not been expressly reserved by their rightsholders” in an “appropriate manner” (para 124). Where rights are reserved, “LLM developers could only do so if they obtained a licence”.
(f) Chatbot answers involve RAG, which is an online use. EU law applies (Article 15(1) CDSM) where the chatbot’s output is shown to users within the EU (para 165). RAG is integral to the chatbot’s service but infringement depends on the amount reproduced and shown to the end user – so the case may be of limited value as a precedent (para 169 and 174).
(g) RAG itself is unlikely to be an infringement because the economic value is the output. Technical copies made during RAG do not have “independent economic significance” but the economic advantage arises when the answer is provided by the chatbot (para 184).
(h) Consent for search is not consent for AI. The AG disagrees that publishers who leave articles freely accessible implicitly consent to their reuse by chatbots. Paywalls are a significant commercial choice and a publisher may accept indexing to be visible in search but not to ground chatbots (paras 224-228).
(i) The balance and the right to property. Publishers have a claim to property (Article 17(2) of the Charter) and their revenue supports media pluralism and the public's right to receive information (para 231). Chatbot must providers have the freedom to conduct a business (Article 16) (para 232). The AG's middle ground is that chatbots that lawfully retrieve content and reuse it in a recognisable way with links for users should operate freely (para 233-234).
Chatbots which “systematically circumvent access restrictions… and/or "recognisably" provide it to their users in a paraphrased fashion should not be freed from the constraints of copyright” (para 235). Without control and compensation, this will make users rely on derivatives rather than original works.
This leaves publishers “in a strong bargaining position vis-à-vis chatbot providers'” who “need access to press publications to be able to generate accurate answers relating to the news” (para 236).
Publishers must reserve rights effectively using “appropriate means” (para 148-149)
- The opt-out 'should not be excessively complicated' for rightsholders but for online content, it should be “machine-readable”, which bots are able to understand.
- For example, the plain Robot Exclusion Protocol is widely used (even by Google in “Google Extended”), which rests on that protocol.
- It can also include "metadata" embedded in individual files or even… "the terms and conditions of a website" so natural-language reservations may qualify where they are 'machine understandable'.
AI players must be transparent (per Article 53(1)(d) of the AI Act) and provide a detailed summary about the content used for training (para 150).
MOW's free-to-use Search Only Contract (SOC) helps publishers legally restrict automated scraping, bulk data extraction and AI reuse, while still allowing standard search engine indexing. The SOC is embedded in the robots.txt file, so it is machine-readable and triggered at the moment a bot accesses the website. It also sets out the reservation in contractual natural language. It therefore appears to meet both of the AG's express statements about what is needed.
Publishers should also, where possible, keep their content on servers in the EU.
Preiskel & Co advises on competition law, abuse of dominance and regulatory matters affecting digital markets. For enquiries regarding section 393 of the Communications Act 2003, Online Safety Act 2023 and conduct or media sector disputes, contact the firm’s antitrust practice.
Please contact Tim Cowen if you have any questions.
The material in this article is only for general review of the topics covered and does not constitute legal advice. No legal or business decision should be based on its content. This article is written in the English language. Preiskel & Co LLP is not responsible for any translation of all or part of its content into any language.