What happened
On November 3, 2026, Advocate General Maciej Szpunar of the Court of Justice of the European Union (CJEU) released a formal opinion regarding the training of AI models on copyrighted material. The opinion, delivered in the case C-250/25 Like Company v Google Ireland Limited, argues that training AI models on copyrighted books and articles should generally be considered lawful under EU law. However, this permission is contingent upon three primary conditions: developers must respect machine-readable opt-outs (such as robots.txt), avoid bypassing paywalls or bot blocks, and demonstrate that they have implemented technical measures to prevent models from reproducing specific copyrighted works verbatim.
The opinion stems from a dispute between Hungarian news publisher Like Company Kft. and Google, where the publisher alleged that Google's Gemini chatbot infringed on its rights by using its articles for training. The Advocate General advised the Court that the specific questions regarding training were technically inadmissible in this case, as there was no evidence the publisher's specific articles were used to train Gemini. However, he proceeded to provide a comprehensive analysis of the legal framework for AI training to guide future debates.
Szpunar concluded that the text and data mining (TDM) exception under Article 4 of the CDSM Directive applies to AI training. He rejected the argument that training is not 'data mining,' noting that the legislative intent behind the directive was broad enough to cover the development of new technologies. He emphasized that while the training process involves making reproductions of copyrighted works, these are permitted under the TDM exception provided the developer adheres to specific conditions.
The opinion explicitly addresses the 'three-step test' from international copyright treaties, concluding that AI training does not inherently violate these standards. However, he introduced a 'burden of proof' for developers: to benefit from the TDM exception, they must demonstrate they have implemented measures—such as deduplication and output filters—to prevent the model from generating near-copies of training material. Failure to do so could result in the loss of the TDM exception's protection.
Regarding territoriality, the opinion suggests a middle-ground approach: EU copyright law should apply to the scraping of content from servers located within the European Union, regardless of where the subsequent training occurs. This interpretation aligns with the AI Act's requirements for providers to maintain copyright compliance policies, effectively extending the reach of EU law to international developers who scrape European data.
Why it matters
This opinion provides a critical framework for the intersection of and European copyright law, addressing the tension between innovation and intellectual property rights. By validating the text and data mining (TDM) exception for AI training, the adviser offers a path for developers to continue building models while establishing clear boundaries for publishers and creators. The requirement for developers to prove they have taken steps to mitigate ''—where models regurgitate training data—shifts the burden of responsibility onto AI companies. Furthermore, the opinion touches on the complex issue of territoriality, suggesting that EU copyright protections may apply to the scraping of data from European servers even if the actual model training occurs in a different jurisdiction, such as the United States. This has significant implications for global AI developers operating within the EU market.
The opinion is significant because it clarifies that the press publishers' right under Article 15 of the CDSM Directive does not cover the act of training AI models, as training is not an 'online use' of a publication. This provides a degree of legal certainty for AI developers, who have faced uncertainty regarding whether their training datasets constitute copyright infringement.
By framing the opt-out as the primary lever for publishers, the opinion reinforces the importance of machine-readable standards like robots.txt. This creates a clear mechanism for rights holders to protect their content from being used in future training runs, provided they act to signal their preferences to web crawlers.
The focus on '' and regurgitation highlights a shift toward technical accountability. By linking the lawfulness of training to the implementation of safeguards, the opinion encourages the industry to adopt standardized practices for content filtering and model safety, potentially reducing the frequency of verbatim output that currently fuels copyright litigation.
The opinion also touches on the concept of 'digital sovereignty,' suggesting that the EU legislature may prioritize the development of European AI models to reduce dependence on foreign technology. This acknowledges that copyright policy is increasingly intertwined with industrial and geopolitical strategy.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Why can ethical evaluation not be reduced to one model score?
What to watch next
The primary uncertainty remains whether the Court of Justice will adopt the Advocate General's reasoning, as his opinion is non-binding. Additionally, the opinion leaves open the question of how national courts should evaluate the 'reasonableness' of the technical measures taken by developers to prevent regurgitation. Observers should monitor whether the Court addresses the territoriality of training, as the adviser noted that the legislature has not yet settled whether EU law reaches training activities conducted entirely outside the Union. Finally, the practical application of the TDM exception will likely be tested in future litigation, particularly regarding how 'lawful access' is defined when scraping data from the open web.
The Court of Justice's final ruling in the Like Company v Google case will be the next major milestone. While the Court often follows the Advocate General's opinion, it is not required to do so, and any deviation could significantly alter the legal landscape for AI training in Europe.
The practical definition of 'lawful access' remains a point of contention. Future cases will likely explore whether scraping data that is publicly available but protected by terms of service or technical barriers constitutes 'lawful access' under the TDM exception.
The industry's response to the proposed 'burden of proof' regarding technical safeguards will be critical. Developers will need to determine what level of filtering is sufficient to satisfy national courts that they have taken adequate steps to prevent the reproduction of copyrighted works.
The ongoing debate over whether AI models themselves contain 'copies' of training data remains unresolved. While the Advocate General expressed skepticism, the issue continues to be litigated in various national courts, and a definitive EU-wide consensus has yet to emerge.