16.06.2025
The launch of OpenAI’s ChatGPT in November 2022 brought generative AI to the forefront, transforming how people work and create. Since then, models that can process and generate not only text but also images, audio, and video have gained momentum. However, these advancements raise significant copyright concerns: generative AI uses copyrighted materials for training and produces realistic, original content, thus challenging traditional concepts of authorship and originality.
Thank you
Generative AI refers to a type of artificial intelligence capable of creating original content—such as text, images, video, audio or software code—in response to a user’s prompt or request.
Generative AI uses advanced machine learning models known as deep learning models, that analyse large amounts of data to identify patterns and relationships. With this information, they can process a user’s natural language request and generate relevant new content in response.
The original GPT-3 model from Open AI was trained on 570 gigabytes of text, with some of this data being scraped from the internet and including copyright-protected text and images. As expected, this raised concerns about potential copyright infringement.
One argument is that AI models do not retain permanent copies of any copyrighted work as they only use the data to generate weightings (the numerical values adjusted during AI training) in their neural networks, akin to how humans learn by reading books or viewing artworks. Therefore, training models could fall under existing EU copyright exceptions (EU Digital Single Market Directive (2019/790), Article 3) and the US Fair Use Index (US Copyright Act, Section 3) as the data is not directly copied as such. As such, it has been reasoned that generative AI is not replacing human creativity as it serves as a tool to foster new forms of creativity.
Unsurprisingly, copyright holders argue that using their works in training data sets without permission constitutes copyright infringement. Several lawsuits (discussed more below) have been filed against AI developers, but clear legal answers are unlikely to emerge soon. Determining how much a single work contributes to training an AI model—and how to fairly compensate the copyright owner—is challenging. Adding to this, since copyright registration is not mandatory in many jurisdictions, identifying copyrighted works can be difficult.
AI generated content also sparks debate over whether it can receive copyright protection, consequently challenging legal concepts of originality, authorship, and ownership. If a human significantly contributes to the creative process using AI, the resulting work might be eligible for copyright protection. However, defining the extent of human contribution and establishing a clear threshold for copyright eligibility remains a complex issue. For instance, the approach and requirements for copyright rights to subsist in computer-generated works vary from jurisdiction to jurisdiction.
In the UK and Ireland, the author of a computer-generated work is the person who made the arrangements necessary for its creation (UK Copyright Act 1988, Section 9(3) and Irish Copyright and Related Rights Act, 2000, Section 21). That is to say, if a person makes the key decisions and exercises control over how the AI generates content, they may be considered the author.
For instance, if a user inputs specific prompts into an AI system, selects elements, or refines the output in a meaningful way, the user may be considered as the author of the work in its own right. However, this is dependent on the level of originality of the expression rather than the effort or labour put into creating the work.
The work created should reflect the author’s personality and choices and so involves some degree of creativity, even if it is minimal, as indicated by the CJEU through its decisions, notably under the InfoSoc Directive including Infopaq International A/S v Danske Dagblades Forening (C-5/08) and Bezpečnostní softwarová asociace – Svaz softwarové ochrany tegen Ministerstvo kultury C-393/09. Thus, it could be argued that if the AI is merely a tool used by the person, and the person exercises creative control over the output, the person can be considered the author.
However, if AI operates autonomously with minimal human input, the resulting work may not qualify for copyright protection, as the absence of a clear human creator diminishes the possibility of assigning authorship or granting copyright.
A significant ongoing case in the UK is Getty Images v. Stability AI (2023), in which Getty Images accuses Stability AI, the developer of Stable Diffusion (a generative AI model), of infringing copyright by using millions of Getty’s images without permission to train the AI system.
In December 2023, the UK High Court allowed Getty’s claims to proceed to trial, rejecting Stability AI’s arguments for a summary judgment. Stability AI’s key argument was that the training and development of Stable Diffusion did not take place in the UK. It asserted that all of the computing resources it used for training were, at all times, located outside the UK and that, in particular, all of the training computing infrastructure was located in two U.S. data centres operated by Amazon Web Services Inc.
Controversially, the judgment left open the possibility that an AI model could qualify as an infringing copy of its training data, which implies that the Stability AI model could potentially embed copies of its training data. The Court also held that secondary infringement of importing, possessing or dealing with an infringing copy, might extend to intangible information such as making software available via a website. This demonstrates the Court may be willing to reconsider established copyright principles.
This signals a potential shift in how copyright law could be applied in the future and could set a precedent, forcing AI companies to reassess how they source training data and prompting significant changes in how copyright is enforced in digital spaces. The outcome of this trial may redefine the legal responsibilities of AI developers in the UK and shape the boundaries of copyright protection in the modern era.
EU copyright law remains grounded in traditional concepts of authorship and creativity, similar to the UK and Ireland. The EU Copyright Directive (Directive 2001/29/EC) and national implementations across member states emphasise human creativity as a central requirement for copyright protection.
As in the UK and Ireland, the degree of human control and creativity is critical in determining whether a work involving AI attracts copyright protection. Minimal human involvement may disqualify AI-generated works from protection under EU law, but significant input and originality could allow the user to claim authorship.
Currently, the EU's legal framework for generative AI and copyright is based on existing copyright law that requires the use of copyrighted material, including for AI training, to have the permission of rights holders unless it falls under specific exceptions, such as text and data mining (TDM).
Under the EU Digital Single Market Directive (2019/790), Article 3 permits certain uses of copyrighted material for TDM (the process used to train generative AI systems) for research purposes with lawful access. Lawful access includes content accessed through open access policies, subscriptions or contractual arrangements. The UK has a similar TDM exception (Copyright, Designs and Patents Act, section 29A), allowing use for non-commercial research purposes, though recent policy discussions suggest potential changes to how broadly this exception may apply.
Copyright holders in both regions have raised concerns about the scope of data mining and whether sensitive or proprietary information is being extracted during the process. While lawful access is required, the widespread use of TDM could impact how copyrighted works are commercially exploited in the future.
In the milestone case of Robert Kneschke v. LAION (2024), Hamburg Regional Court dismissed a copyright lawsuit by German photographer Robert Kneschke against LAION, a non-profit organisation, aiming to make large-scale machine learning models, datasets and related code available to the general public. This decision was not based on the actual use of a copyrighted work for AI training, but instead was limited to acts preceding a possible AI training.
Kneschke argued that the download of his photos for the purpose of building an AI training database constituted copyright infringement, noting the stock-image site's terms of service explicitly prohibit automated downloads or scraping. However, the court ruled that LAION's actions were protected under Section 60d of the German Copyright Act, which permits TDM for scientific research. Since LAION provides its dataset for free, the court deemed it non-commercial and research-focused. This ruling suggests that non-commercial AI developers acting in the public interest and reinvesting profits into research may qualify under this exception, while commercial entities likely do not.
In the U.S., copyright law generally requires human authorship for a work to qualify for protection. For instance, as established in cases like Thaler v. Perlmutter (2023), U.S. courts have ruled that works created entirely by AI without human involvement do not qualify for copyright protection.
Although human authorship plays a central role in copyright protection, the human must exert some creative control over the output. As an example, an AI-assisted image that a human significantly edits or transforms could receive copyright protection, but only for the human-made modifications. In the Zarya of the Dawn (2023) case, artist Kristina Kashtanova created a graphic novel using the AI platform MidJourney to generate images. While Kashtanova was initially granted copyright for the entire work, the U.S. Copyright Office later re-evaluated this after learning that AI was used to create the images.
The Copyright Office ultimately allowed copyright protection for Kashtanova’s text and the overall arrangement of the AI-generated images but refused to extend the copyright protection to the individual images produced by the AI. The key issue was whether the AI-generated images met the requirement of human authorship under U.S. copyright law. While Kashtanova influenced the creative process, the Office concluded that her level of direction was insufficient to establish authorship. As a result, the images were not eligible for copyright protection, though the arrangement of text and images was protected.
In another case against Stability AI (Andersen v. Stability AI (2023)), artist Sarah Andersen and other plaintiffs filed a lawsuit against Stability AI, DeviantArt, and MidJourney, alleging copyright infringement related to the training of AI image-generation models. The core legal claim is that these companies scraped billions of online images, including copyrighted works, without authorisation, and used them to train AI systems capable of producing new images that may infringe on the original works.
A central argument is that the AI models retain “compressed copies” of the training images, effectively embedding and reproducing protected works. While some claims were dismissed due to lack of copyright registration, Andersen’s registered works allowed key claims to proceed. The case raises unresolved legal questions about how copyright law applies to data used in AI training and the nature of outputs generated by such systems.
As AI technology advances, the UK, Ireland, EU and U.S. face growing pressure to adapt their legal frameworks to balance the rights of creators, developers, and the public. Ongoing lawsuits highlight tensions between human creativity and AI-generated content, with outcomes likely to shape future copyright rules.
AI users should review agreements with tool providers to clarify ownership of content created through their platforms. To help prevent unauthorised use in AI training, some creators include statements restricting such use or adopt technical measures like Digital Rights Management (DRM). While these approaches do not guarantee full protection, they may help clarify your intentions and support your ability to assert your intellectual property rights.
The use of copyrighted material in AI training remains a legal grey area, with several ongoing lawsuits seeking to clarify the boundaries. Copyright owners may wish to stay informed about evolving case law and legislative developments—such as the UK’s proposed Data Bill on AI and creative rights. Likewise, AI users and developers should monitor key cases like Getty Images v. Stability AI, which may set important precedents for how copyright applies in the AI context.
Thank you
