EU court adviser signals when AI chatbots may infringe press rights

The core of the dispute

A Hungarian website owner, Like Company Kft., has sued Google over its Gemini chatbot. The company argues that the AI continuously infringed on its rights by copying news articles from its web portals between June 2023 and May 2024. The legal battle began in a regional court before being sent to the European Union's highest court for a ruling on how the rules apply.

The specific right in question is not standard copyright. Instead, it involves a specific press publishers' right created by the EU's Copyright in the Digital Single Market Directive. This right gives publishers established in a member state the power to control the online use of their press publications by information society service providers.

The case centers on a specific example involving a Hungarian singer's plan to bring dolphins to Lake Balaton. When a user asked Gemini to summarize an article from the site, the chatbot provided a summary containing details that were not in the original text. Like Company argues this proves the model 'knew' the article from training and reproduced it for the public.

How the AI retrieves news

The court adviser, Maciej Szpunar, explained that the chatbot likely did not use the article during its initial training phase. Because the article was published in July 2023 and Gemini 1.0 was released in December 2023, the adviser found it implausible that the content was used to build the model's core weights. Instead, the chatbot likely used a process called retrieval-augmented generation, or grounding.

In this process, the chatbot uses a search engine to find relevant pages in real time. It then uses internet bots to retrieve that content and adds it to the prompt fed to the underlying model. The model then generates an answer based on that specific information, which may or may not include a hyperlink to the original source.

The adviser noted that while the model's parameters are too abstract to be considered copies of the training data, the output itself can still infringe. If a chatbot rewrites an article so closely that a reader no longer needs to visit the original site, it crosses a legal line. However, the adviser noted that the specific example provided by the Hungarian site did not appear to meet that test for infringement.

Rules for AI training

The opinion provides a significant signal for the broader AI industry regarding how models are trained. Szpunar advised that training large language models generally falls under the text and data mining exception in EU law. This means that, as a rule, training on protected works is permitted unless the publisher has specifically opted out.

To opt out, publishers can use machine-readable reservations of rights. The adviser noted that the robots.txt protocol, Google-Extended, and metadata embedded in files all qualify as valid ways to signal an opt-out. He also suggested that natural-language reservations in a website's terms and conditions could qualify because modern models are capable of understanding them.

The adviser emphasized that the press publishers' right does not offer protection against model training at all. This is because training happens on dedicated computing infrastructure and is not considered an 'online use' provided to an individual recipient. Therefore, while the press right is limited, standard copyright held by journalists may still be available for training purposes under existing contracts.

Who is most affected

The adviser provided a blunt assessment of which types of content are most at risk from AI. He stated that AI-generated text will not displace the demand for human-made masterpieces of literature or in-depth, analytic pieces of journalism. These types of content require a level of originality and depth that current models do not replicate.

However, the adviser warned that AI could easily replace human creators for texts with low originality. This specifically includes purely informative news stories and sports reports. Because these types of content are more easily summarized or reproduced by a chatbot, they face a higher risk of being replaced by automated systems.

For developers, the opinion suggests that they must demonstrate they are taking measures to prevent model-specific substitution. If a model routinely generates content that resembles specific training material, developers should show they used dataset diversification and deduplication. If they fail to do so, a national court might refuse to apply the training exception to that specific model.

What comes next

While this opinion is a strong signal, it is not the final word from the court. The judges still have to decide on the admissibility of the questions regarding whether training is a reproduction and whether the chatbot's answers communicate publications to the public. The court is expected to guide the lower court, which the adviser noted is 'evidently lost in the AI landscape.'

The ruling also highlights a major tension regarding the territoriality of AI laws. Google argued that only the country hosting the main computing infrastructure can regulate training. However, the adviser noted that if this were the case, EU rights holders would have no protection when their works are used to train models hosted abroad.

The adviser suggested that the legislature should eventually settle the question of where training is regulated. Until then, the court will continue to weigh the balance between the rights of publishers and the technical realities of how generative AI retrieves and summarizes information from the open web.

The legal framework

The case involves the Copyright in the Digital Single Market Directive, specifically Article 15(1). This article was designed to give publishers rights over the online use of their press publications by information society service providers. It was intended to ensure that publishers could manage how their content is shared and used across the internet.

The adviser clarified that the text and data mining exception in Article 4(1) is broad enough to cover machine learning. This position is shared by Google and the governments of Estonia and France. However, other countries like Denmark and Spain have argued that the exception should only apply to 'Big Data' analysis, not the training of generative models.

The court will also have to consider the 'three-step test' from international copyright treaties. Some countries, including Greece and Spain, argue that exempting AI training from these rules breaches international standards. The adviser disagreed with this view, but the court's final decision will determine how these international obligations interact with new AI technologies.

Technical nuances

A key distinction in the opinion is between the 'training' phase and the 'inference' phase. Training involves scraping, conversion, and tokenization of texts to build a model. The adviser found little doubt that these steps involve reproductions of the works within the meaning of the InfoSoc Directive.

In contrast, the trained model itself is not a database. The adviser urged 'utmost caution' regarding the theory that models contain copies of works because they sometimes reproduce them when prompted. He cited a UK High Court ruling finding that a model's parameters are too abstract to be considered copies of the original data.

The distinction is important because if models were considered containers of copies, the mere offering of a model would require consent from every single rights holder. By viewing the model as abstract parameters, the court allows for broader use of the technology, provided that the specific outputs do not infringe on the rights of the original creators.

Summary of findings

The adviser's opinion serves as a roadmap for how the EU might handle the intersection of AI and media rights. It establishes that while the technology is complex, the legal principles of access and substitution remain the primary concerns. If a chatbot makes it unnecessary to visit a news site, it is likely infringing on the publisher's rights.

For publishers, the takeaway is the importance of clear opt-out signals. Using standard protocols like robots.txt or specific metadata is the most reliable way to signal that content should not be scraped for training. This gives publishers a tool to protect their revenue streams while the broader legal landscape continues to evolve.

For AI developers, the opinion highlights the need for responsible data practices. Demonstrating that a model is not being used to substitute for specific training materials is becoming a legal necessity. Developers will need to show they are using diverse datasets and filters to ensure their models provide value without unfairly cannibalizing the sources they learn from.

Key takeaways

The court adviser's opinion clarifies that AI chatbots can be held liable for copyright and press rights in two main ways. First, they cannot bypass technical protections like paywalls to access content. Second, they cannot produce summaries that are so close to the original work that they replace the need for the source.

The opinion also reinforces that training AI models is generally allowed under current EU exceptions. This is a significant win for developers, provided they respect the opt-out signals of publishers. It suggests that the EU is trying to find a middle ground that allows for innovation while protecting the economic interests of the press.

Ultimately, the case highlights the shift from 'training' as a legal issue to 'retrieval' as a legal issue. As chatbots become more integrated into search engines, the way they fetch and present information in real-time will be the primary battleground for digital rights in the coming years.

The role of the court

The Court of Justice of the European Union is now tasked with making the final determination on these questions. The adviser's opinion is a significant guide, but the judges will decide on the admissibility of the specific questions regarding training and reproduction. This is the first generative AI case to reach the Court of Justice, making it a landmark for the entire continent.

The court will have to balance the interests of the press, who rely on traffic for revenue, with the interests of AI developers who need data to build useful tools. The adviser's conclusion that AI will not replace high-quality journalism provides some comfort to the media industry, but the threat to low-originality news remains a point of concern.

The final ruling will likely set a precedent for how other EU member states handle AI litigation. It will clarify whether the press publishers' right is a robust shield against AI or a narrow protection that only applies to specific types of digital use. The outcome will shape the future of how news is consumed and distributed in the age of generative AI.

Future implications

This case signals a move toward a more nuanced regulatory environment for AI. Instead of a blanket ban or a total free-for-all, the court is looking at specific behaviors like paywall bypassing and content substitution. This allows for the continued growth of AI while creating boundaries for how it interacts with existing media ecosystems.

For the reader, this means that the way AI tools provide information may become more regulated. You might see more frequent citations or links to original sources as developers try to stay on the right side of the law. It also ensures that the economic model of journalism—where readers visit sites to see content—remains viable in an AI-driven world.

The next steps will involve watching how the court handles the 'three-step test' and the territoriality of training. These questions will determine how much control EU publishers have over models trained in the United States. The outcome will be a defining moment for the balance of power between big tech, content creators, and the general public.

Summary of the case

The case of Like Company v Google Ireland Limited is a landmark for AI law in Europe. It addresses the fundamental question of whether AI chatbots can be used to undermine the business models of news publishers. The court adviser's opinion provides a detailed analysis of the current legal landscape, covering everything from training exceptions to real-time retrieval.

The adviser's opinion suggests that while training is largely permitted, the way a chatbot delivers an answer to a user is subject to strict rules. If a chatbot acts as a substitute for a news article, it may be infringing on the publisher's rights. This creates a clear distinction between the 'learning' of the AI and the 'doing' of the AI.

As the case moves forward, it will serve as a test case for the Copyright in the Digital Single Market Directive. The final ruling will help define the boundaries of the press publishers' right in the era of generative AI. It will determine how news organizations can protect their content while still allowing for the technological advancements that AI brings to the internet.

Technical details

The technical core of the case involves retrieval-augmented generation, often called grounding. This is the process where a chatbot like Gemini uses a search engine to find live information. It then feeds that information into the model to generate a response that is grounded in current events rather than just its training data.

The adviser noted that this process is distinct from the training phase, which involves scraping massive amounts of data to create the model's weights. Because the training happens on dedicated infrastructure, it is not considered an 'online use' under the press publishers' right. This means that the training of the model is generally exempt from the specific rights of publishers.

However, the output of the chatbot is a different matter. If the chatbot uses the retrieved data to create a summary that is nearly identical to the source, it may violate the publisher's rights. The court will have to decide exactly how 'close' a summary must be before it is considered a reproduction that replaces the need for the original publication.

Impact on publishers

For publishers, the adviser's opinion provides a roadmap for defending their content. It clarifies that the press publishers' right is a specific tool for controlling the online use of publications. This right is separate from copyright, which is usually held by the individual journalists who write the stories.

The opinion suggests that publishers should focus on ensuring their content is 'lawfully accessible.' This means they should not allow their content to be scraped from pirate sites or shadow databases. They can also use technical measures like paywalls and bot blocks to prevent unauthorized access by AI companies.

Furthermore, publishers have the power to opt out of AI training using machine-readable reservations. By using tools like robots.txt or metadata, they can signal to AI developers that their content should not be used to train models. This gives them a way to maintain control over their intellectual property while the legal system works out the details of AI usage.

Impact on AI developers

For AI developers, the opinion provides some clarity on the legality of training models. The adviser's view that training falls under the text and data mining exception is a significant point for the industry. It suggests that, in most cases, developers can use publicly available news content to train their models without seeking individual licenses.

However, developers must be careful about how their models behave in real-time. They need to ensure that their chatbots do not bypass paywalls or create summaries that are so close to the source that they infringe on publisher rights. They may also need to demonstrate that they are taking steps to prevent their models from routinely substituting for specific training materials.

The opinion also highlights the importance of respecting opt-out signals. Developers who ignore machine-readable reservations of rights could face legal challenges. By adhering to protocols like robots.txt and Google-Extended, developers can operate more safely within the European legal framework while still providing powerful AI services to their users.

The role of governments

The case has seen significant involvement from various European governments. Most countries, including France, Germany, and Italy, attended the hearing to provide their observations. Their input reflects the diverse ways in which different member states view the intersection of AI and copyright law.

Germany, for instance, proposed a specific approach where EU copyright would apply to scraping from servers located within the Union. Other governments have argued about whether the text and data mining exception should cover model training at all. These differing views show that the EU is still grappling with how to apply existing laws to new technologies.

The court adviser's role is to help the court navigate these complexities. By providing a detailed opinion that addresses the technical and legal nuances of the case, the adviser is helping to create a unified approach for the entire EU. The final decision will be a major step toward a consistent regulatory framework for AI across the continent.

The future of news

The adviser's opinion touches on the broader implications for the future of journalism. He noted that while AI will not replace high-quality, in-depth analysis, it could easily replace human creators for low-originality content. This means that the economic value of news may shift toward more complex and original reporting.

For the news industry, this means that the business model of providing information may need to change. If AI can provide simple facts and sports scores for free, news organizations may need to focus more on unique