This article examines the principal legal challenges that generative artificial intelligence poses, particularly in the field of copyright law. It explores key issues relating to the authorship of AI-generated content, the training of AI models on copyright-protected works, the use of third-party works by AI systems, and liability for infringements arising from the deployment of artificial intelligence.

Copyright law and artificial intelligence – an overview of the current legal framework

The starting point for analysing the relationship between artificial intelligence and copyright law is the concept of a copyright work as an expression of human creative activity possessing individual character. Copyright protection extends only to the expression of ideas; ideas, procedures, methods of operation and underlying principles remain outside the scope of protection.

Although the Polish Copyright Act does not expressly state that an author must be a human being, the entire copyright system is founded upon that assumption. Only a natural person who has engaged in an act of creative authorship may qualify as the author of a work.

Accordingly, an artificial intelligence system cannot be regarded as the author of a copyright work or as the holder of copyright. As discussed in our previous publication, this also means that content generated autonomously by AI will, in most cases, not attract copyright protection in favour of the user merely because it was produced in response to prompts.

An AI-assisted output may nevertheless qualify for copyright protection to the extent that it reflects the user’s own creative choices and intellectual contribution. Conversely, where the output is generated exclusively through the autonomous operation of an AI system, without sufficient human creative input, it will not constitute a copyright work within the meaning of copyright law.

The relationship between artificial intelligence and copyright law has also been significantly influenced by the AI Act. Among other obligations, the Regulation requires providers of general-purpose AI (GPAI) models to implement a policy ensuring compliance with EU copyright law and to publish a sufficiently detailed summary of the content used for training their models.

The legal framework governing the use of copyright-protected works for text and data mining (TDM) is primarily established by Directive (EU) 2019/790 on Copyright in the Digital Single Market (DSM Directive). Subject to certain conditions, the Directive permits the use of protected works for text and data mining while allowing rightholders to reserve their rights and opt out of such use.

However, these legislative instruments do not provide comprehensive legal certainty in relation to AI-generated content. Several important legal issues therefore remain unresolved and are discussed below.

AI-generated content will generally not qualify for copyright protection.

Under the current legal framework, an artificial intelligence system cannot qualify as the author of a copyright work. What remains uncertain, however, is under what circumstances a user of an AI system may be recognised as the author of AI-generated content for copyright purposes.

Some guidance may be drawn from the practice of the U.S. Copyright Office (USCO). The decision in Théâtre D’Opéra Spatial illustrates that even the use of more than 600 prompts may not be sufficient to establish the level of human authorship required for copyright protection. Similarly, in Zarya of the Dawn, the USCO recognised copyright only in the comic’s text and overall arrangement while refusing protection for the individual illustrations generated by AI.

This does not mean, however, that every work created with the assistance of AI is automatically excluded from copyright protection. A clear distinction must be drawn between situations in which an AI system autonomously generates an entire output on the basis of user prompts and those in which AI is used merely as a technical tool during the creative proces — for example, to adjust image contrast, remove background elements or perform other editing functions on an existing work. The decisive consideration remains whether the work, viewed as a whole, constitutes the author’s own intellectual creation and reflects a sufficient degree of human creative activity.

Nevertheless, the boundary between copyright protection and the absence of protection in the context of AI-assisted creativity remains uncertain. This legal uncertainty may prove particularly problematic for the advertising and creative industries, where the absence of copyright protection may significantly limit the ability to enforce rights against unauthorised use of AI-generated content.

AI models are frequently trained on copyright-protected works without a valid legal basis

The second issue giving rise to significant controversy concerns the use of third-party copyright-protected works for training AI models. Modern AI models require vast quantities of data, including books, newspaper articles, photographs, musical works, films and computer programs. In many cases, these materials are protected by copyright, and their use would ordinarily require the authorisation of the relevant rightholders. From a copyright perspective, it is therefore essential to determine what acts are carried out in relation to those materials during the training process and what legal basis exists for their use.

In the European Union, particular importance is attached to the provisions governing the text and data mining (TDM) exception contained in Directive (EU) 2019/790 (DSM Directive). The Directive provides two separate exceptions.First, research organisations and cultural heritage institutions may carry out text and data mining on protected works for the purposes of scientific research. Secondly, other entities may also rely on the TDM exception, but in that case rightholders may reserve their rights and exclude the use of their works through an opt-out mechanism. As a consequence, entities responsible for training AI models may use only those works whose use has not been excluded by an effective opt-out reservation.

How does it work in practice? Although it is clear that copyright-protected works are used in the training of AI models, it remains uncertain whether Article 4 of the DSM Directive actually extends to the training of generative AI models. This issue is currently being examined in the widely discussed case Like Company v. Google (C-250/25) pending before the Court of Justice of the European Union. The preliminary questions referred to the Court concern, in particular, whether the training of AI models involves acts of reproduction of protected data and, if so, whether such use falls within the scope of the TDM exception. Pending the Court’s ruling, it is nevertheless clear that providers of AI models are likely to rely on Article 4 DSM Directive as the principal legal basis for training activities. This means that any person who does not wish their works to be used for AI training must make an appropriate reservation in accordance with Directive 2019/790.

In practice, however, cases have already emerged suggesting that the opt-out mechanism may not always be effective.An example is the Danish case BoligPortal v. ReData (Sø- og Handelsretten, Case No. BS-42485/2025-SHR). Despite an effective reservation excluding the processing of data, the defendant continued to use data obtained from the claimant’s website. The case demonstrates that the opt-out mechanism does not guarantee that protected works will not be used during AI training.

What’s more, an additional problem is created by so-called shadow libraries, namely unauthorised repositories that unlawfully store copyright-protected works. Data-collection tools may bypass a protected work that has been effectively excluded through an opt-out reservation on the original website, only to retrieve the same work from an unlawful source that does not contain such a reservation. This creates a risk of circumventing the framework established by the DSM Directive and of using protected content without a valid legal basis.

The AI Act addresses certain aspects of the compatibility of AI training with EU law. Providers of general-purpose AI models are required, among other things, to establish a policy ensuring compliance with EU copyright law and to publish a sufficiently detailed summary of the content used for training their models. These obligations enhance the transparency of the training process, but they do not resolve the problems identified above. For authors and other rightholders, the key concern therefore remains not only the possibility that their works may be used for AI training, but also the possibility that attempts to exclude such use may prove ineffective in practice.

AI can generate content that infringes on copyright

One of the most real threats is the possibility that AI could generate content that resembles an existing, copyrighted work. AI can create text, graphics, photographs, music, or audiovisual material containing elements characteristic of a specific work. In extreme cases, the result may also contain excerpts from an existing work or constitute an almost exact reproduction of it.

At this stage, it is important to distinguish between two situations in which plagiarism or other content infringing on copyright or moral rights may arise. The first type of situation involves a user instructing an AI system to generate a work that “resembles” an actual protected work. This could involve, for example, copying the style of a specific artist to generate an image or the voice of a specific narrator for an advertisement. In such a case, it is difficult to attribute liability to the system that generated the content in accordance with the user’s instructions or to the company responsible for the system.

A different situation arises when the system reproduces a protected work without the user’s knowledge or intent. This will apply in particular to situations where, after being given a general prompt such as “generate an image depicting a brick wall,” the AI system generates a specific protected work (e.g., a brick wall on which the AI system has additionally reproduced a work by Banksy). This phenomenon is commonly referred to as “memorization.” It involves the model memorizing data during the training phase in such detail that it allows the model to reproduce a protected work without user input or the model accessing the internet. However, the mere fact that the user obtained such a result by issuing a command to the AI—which did not instruct the model to reproduce a specific protected work—does not exempt the user from liability for any infringements. The issue of liability on the part of the entity responsible for the AI system, however, remains contentious.

It should be noted, however, that generated material containing protected elements of another person’s work may lead to copyright infringement regardless of the user’s intent or knowledge of the infringement. In practice, therefore, one cannot assume that simply because the content was generated by AI, the user is not infringing the rights of third parties.

This risk is particularly significant for businesses using AI in marketing, advertising, customer communications, or the creation of content published online. Verifying the output before its dissemination should therefore be an integral part of a basic legal risk management process.

Liability for incorrect information generated by AI remains unclear

A separate issue is the liability of AI system providers for infringements and errors caused by the operation of these systems, as briefly outlined in the previous section. Errors made by AI are, in fact, so common that special terminology has been created to describe this phenomenon. A situation in which an AI-generated response contains false or misleading information presented as facts is commonly referred to as a “hallucination.”

The consequences of hallucinations can vary—a false perception of reality caused by an incorrect AI response can lead to erroneous business, legal, or social decisions. Focusing solely on the legal sector, one can observe a number of errors that are increasingly being revealed in court cases: AI misinterprets legal provisions, incorrectly applies case law to the facts of a case, erroneously analyzes the legal situation, and even cites nonexistent case law to support a particular argument.

Can AI system providers then be sued for providing incorrect information? The answer to this question is not clear-cut. In the cited case of using AI as a substitute for a lawyer, it is hard to imagine that the AI provider would be held liable for the system’s errors. The providers’ terms of service clearly state that the systems often make mistakes, and the answers they provide should always be verified, especially when it comes to the use of AI for legal, medical, commercial, or other such purposes.

However, the issue of so-called “AI Overviews”—that is, brief answers provided by AI in response to a question asked in a web browser—should be viewed differently. These models are increasingly replacing the functions of a web search engine. A case pending before the German Court in Munich, case no. 26 O 869/26, concerned precisely this issue—the AI system misled users who, when searching for a company name, were incorrectly informed by the AI Overview about suspicious practices and customer service issues. The court’s decision states that Google’s AI Overview cannot enjoy the same protection as a search engine, which merely relays third-party content.

A similar conclusion emerges from an analysis of the Campact v. xAI case (Landgericht Hamburg, case no. 324 O 461/25), which concerned false content published by the “Grok” AI system . The court noted that the burden of proof to demonstrate the accuracy of the information provided by the AI system rested with Grok’s operator (xAI); however, xAI failed to provide any explanations or evidence in this regard.

In summary, it appears that providers of AI systems may be held liable for content generated by their tools; however, this will depend on many factors, such as the nature of the content and whether it is published as part of private correspondence with a user or is publicly available to third parties as well.

Summary

The development of AI poses significant challenges for copyright law. The problem lies not only in the possibility of using others’ works without the rights holders’ knowledge or the lack of copyright protection for AI-generated content, but also in the vaguely defined rules regarding liability for unlawful actions by AI models. In practice, therefore, the use of AI requires an awareness of the associated risks and the responsible use of its capabilities.