Does training an AI model on online content infringe EU copyright? When does a chatbot answer copy a news article? On 3 November 2026 Advocate General Szpunar gave the first Opinion on AI training and EU copyright in Like Company v Google. He advised the Court not to rule on training in this case. He proposed a test for chatbot answers instead.
The case is C‑250/25, Like Company v Google Ireland Limited, ECLI:EU:C:2026:919. It is a reference from the Budapest Regional Court in Hungary. The Opinion is published as a provisional text. It does not bind the Court, which will give its judgment later.
- The Advocate General proposed that the questions on LLM training be declared inadmissible as hypothetical.
- His provisional view is that training involves reproduction. The text and data mining exception in Article 4 of the DSM Directive can in principle cover it.
- Lawful access and a respected opt out are the key conditions. Robots.txt, metadata and even website terms may work as an opt out if machines can read them.
- A chatbot answer reproduces a press publication if it was accessed unlawfully, or if it keeps enough of the article's expression to replace it.
- He said training carried out outside the EU raises a serious territorial question that the legislature should address.
The facts
Like Company runs Hungarian news portals, including one about Lake Balaton. It earns money from advertising. It claims the press publishers' right under Article 15 of the DSM Directive (EU) 2019/790, as implemented in Hungary. It says Google's Bard and Gemini chatbot infringed that right between 13 June 2023 and 31 May 2024.
Its main example was an article of 21 July 2023 about a singer's plan to bring dolphins to Lake Balaton. A user later asked Gemini for a summary of that article and named the website. Gemini gave a summary. Like Company argued that this proved its articles were used for training, and that the answer itself reproduced the article.
The Hungarian court asked four questions. Two concern chatbot answers. Two concern training and the text and data mining (TDM) exception.
Why the training questions should not be answered
The Advocate General gave two reasons. First, nothing in the file showed that Like Company's articles were in Gemini's training data. Gemini 1.0 was released in December 2023, only months after the article. Large models are usually trained on data collected well before release. The summary was more likely produced by retrieval augmented generation (RAG), where the chatbot fetches a web page in real time and reuses it.
Second, the press publishers' right covers only "online use" by information society service providers. In his view, training a model on dedicated computing infrastructure is not such a use. The link between training and later chatbot answers is too indirect. So training falls outside the press publishers' right altogether.
His provisional view on training
For completeness, he still addressed training. He stressed that these views are provisional.
- Reproduction. He considered it quite clear that using works to train an LLM involves reproductions under Article 2 of the InfoSoc Directive 2001/29/EC.
- The TDM exception. Article 4 of the DSM Directive allows reproductions of lawfully accessible works for TDM unless the rightholder has expressly reserved that use. He considered that this exception can in principle cover training, subject to the three step test in Article 5(5) of the InfoSoc Directive. Article 7(2) of the DSM Directive applies that test. Exceptional cases may fail that test.
- Lawful access. Developers must not scrape obviously unlawful sources, such as pirate sites and shadow libraries. They must not circumvent measures that restrict access.
- The opt out. A reservation is made in an appropriate manner if current crawlers can understand it reliably at the time of access. He mentioned the Robot Exclusion Protocol, Google's own Google Extended token, file metadata and website terms. He did not rule out terms written in plain language.
- Transparency and pay. He pointed to the duty under the AI Act on providers of general purpose AI models to publish a summary of training content. He added that Member States may create a compensation scheme for uses under the TDM exception.
On territory, Google said Gemini was trained in the United States. He noted that copyright is territorial. The German Government suggested that EU law could reach the scraping of content from servers in the EU, wherever training later happens. He was open to that view. He said the legislature should settle the point.
The test for chatbot answers
He proposed that a chatbot answer built by retrieving a press publication involves reproduction in two cases. The first is where the chatbot accessed the publication unlawfully, for example by circumventing technical measures against access or scraping. The second is where the answer keeps so much of the article's form of expression that the user no longer needs the original.
Making available to the public occurs if that expression is systematically recognisable in answers to related queries. The provider is responsible if those acts are inherent in its business model. It is also responsible if it has not taken the measures a reasonably diligent operator would take to counter them.
On the facts before the Court, he observed that the expression of the dolphin article did not appear to be recognisable in Gemini's answer. Google also seemed to have mitigation measures in place. The Hungarian court will decide.
What the Opinion does not decide
- It is not a judgment. The Court of Justice may take a different view.
- It gives no binding answer on training or the TDM exception.
- It does not decide whether Google infringed Like Company's rights.
- It does not resolve whether EU law reaches training carried out abroad.
How the US, India and the UK compare
United States. The US has no TDM exception. Courts apply fair use under 17 U.S.C. § 107, case by case. In 2025 two district courts in California found fair use for training on the narrow records before them. The cases were Bartz v Anthropic and Kadrey v Meta. In Bartz the court treated pirated library copies differently. On 29 September 2026 the Third Circuit rejected fair use in Thomson Reuters v Ross, where the copying built a competing, non generative tool. See our note on AI training fair use after Thomson Reuters v Ross.
India. The Copyright Act, 1957 has no TDM exception. Section 52(1)(a) allows fair dealing only for listed purposes. In ANI Media v OpenAI, the Delhi High Court refused an interim injunction on 24 July 2026. According to reports, it held prima facie that storing ANI's content for training fell within section 52(1)(a). It also found the RAG outputs not substantially similar. ANI has appealed. Separately, a DPIIT committee working paper of December 2025 proposed a mandatory blanket licence for AI training with statutory remuneration. It rejected a TDM opt out model. That paper is a policy proposal.
United Kingdom. Section 29A of the Copyright, Designs and Patents Act 1988 allows text and data analysis only for non commercial research. There is no general commercial TDM exception. In Getty Images v Stability AI, decided in November 2025, the High Court held that the trained model was not an infringing copy.
Practical steps
- Publishers and rightholders should put a machine readable reservation on every site where their content is hosted. Use robots.txt rules for AI crawlers, metadata and clear website terms.
- Restrict access where it matters. The Advocate General treats circumvention of access or scraping controls as decisive for both training and chatbot answers.
- AI developers should log the sources they crawl, honour opt outs at the time of access, and avoid pirate sources. Prepare the AI Act training summary from the same records.
- Chatbot and RAG providers should limit verbatim reuse in answers, link to sources and keep evidence of mitigation measures.
- Indian companies training models for EU users should expect EU rules on scraping and opt outs to apply to EU hosted content. Indian law may still treat the same use differently.
We advise on copyright, licensing and AI data issues in India, the US and Europe. See our guide to the types of intellectual property protection and our IP budget guide for startups.
Frequently asked questions
Is AI training allowed under EU copyright law?
Not yet decided by the Court of Justice. In Like Company v Google the Advocate General's provisional view is that the Article 4 TDM exception can cover training on lawfully accessible content. That is subject to any machine readable opt out by the rightholder.
How do I opt out of AI training in the EU?
Article 4(3) of the DSM Directive requires an express reservation in an appropriate manner, such as machine readable means for online content. The Advocate General mentioned robots.txt, metadata and website terms that crawlers can understand.
Does India have a text and data mining exception?
No. Indian law relies on the fair dealing purposes in section 52 of the Copyright Act, 1957. The Delhi High Court's interim ruling in ANI v OpenAI is under appeal.