Abstract
This dissertation examines the evolving landscape of United Kingdom artificial intelligence policy, specifically analysing whose interests are prevailing in the contentious debate surrounding copyright, text-and-data mining, and generative artificial intelligence model training. Through a comprehensive literature synthesis of contemporary academic scholarship, policy documents, and legal commentary, this study investigates how stakeholder dynamics have shaped the current regulatory framework. The analysis reveals that copyright holders and the creative sector have achieved significant policy victories over technology companies and public interest advocates. The UK’s reversal of proposed commercial text-and-data mining exceptions, following sustained lobbying from creative industries, has reinforced a licensing-centric approach that grants rights holders substantial control over how their works are utilised in artificial intelligence development. While ongoing governmental consultations explore potential reforms—including voluntary codes of practice and transparency requirements—no fundamental shift away from licensing-first principles has occurred. This dissertation concludes that the current policy equilibrium favours established creative industries, potentially at the cost of innovation capacity and equitable access for smaller technology firms and academic researchers. Future research should evaluate alternative licensing models and assess the real-world impacts of current policies on the UK’s artificial intelligence ecosystem.
Introduction
The rapid advancement of generative artificial intelligence technologies has precipitated one of the most significant intellectual property debates of the twenty-first century. At the heart of this controversy lies a fundamental question: should artificial intelligence developers be permitted to utilise copyrighted works for model training without explicit authorisation or compensation? This question has become particularly acute in the United Kingdom, where policymakers have grappled with competing interests from the creative industries, technology sector, academic community, and the broader public.
The intersection of copyright law, text-and-data mining, and artificial intelligence model training represents a critical juncture in the development of the UK’s digital economy. Text-and-data mining refers to the computational analysis of large datasets to identify patterns, trends, and relationships within information (Margoni and Kretschmer, 2021). For generative artificial intelligence systems, this process is fundamental to model training, requiring the ingestion and analysis of vast quantities of text, images, audio, and other copyrighted materials to develop the statistical patterns that underpin their outputs (Kretschmer, Margoni and Oruc, 2024).
The policy significance of this debate cannot be overstated. The United Kingdom has positioned itself as a global leader in artificial intelligence development, with the government’s national artificial intelligence strategy emphasising the technology’s potential to drive economic growth and societal benefit (UK Government, 2021). Simultaneously, the creative industries contribute substantially to the UK economy, generating over £100 billion annually and employing more than two million people (Department for Digital, Culture, Media and Sport, 2022). The tension between these two economic pillars has created a policy environment characterised by competing claims to legitimacy and influence.
Recent years have witnessed dramatic shifts in UK policy positions. The government initially considered expanding its text-and-data mining exception to encompass commercial uses, including artificial intelligence training. However, following sustained opposition from creative industry stakeholders, this proposal was abandoned, leaving only a narrow non-commercial research exception in place (De La Durantaye, 2025; Thongmeensuk, 2024). This reversal has reinforced a licensing-centric approach wherein rights holders retain significant control over how their works are used in artificial intelligence development (Rosati, 2025).
This dissertation contributes to the academic discourse by systematically analysing whose interests are prevailing in this policy contest. By synthesising contemporary scholarship and policy developments, this study provides a comprehensive assessment of stakeholder influence, policy outcomes, and potential future trajectories. The analysis is particularly timely given ongoing governmental consultations that may reshape the regulatory landscape in coming years.
The academic significance of this research extends beyond immediate policy considerations. The UK’s approach to balancing copyright protection with technological innovation serves as a potential model—or cautionary tale—for other jurisdictions grappling with similar challenges. Understanding the dynamics of stakeholder influence in this context illuminates broader questions about the relationship between established industries and emerging technologies, the role of lobbying in shaping intellectual property policy, and the challenges of protecting public interest in an era of rapid technological change.
Aim and objectives
The primary aim of this dissertation is to critically analyse whose interests are prevailing in the United Kingdom’s artificial intelligence policy debate concerning copyright, text-and-data mining, and generative artificial intelligence model training.
To achieve this aim, the following objectives have been established:
1. To examine the current legal framework governing text-and-data mining and artificial intelligence training in the United Kingdom, identifying the scope and limitations of existing exceptions to copyright protection.
2. To analyse the stakeholder dynamics shaping UK artificial intelligence policy, with particular attention to the relative influence of copyright holders, technology companies, the research community, and public interest advocates.
3. To evaluate the policy outcomes of recent governmental consultations and legislative developments, assessing whether these outcomes favour particular stakeholder interests.
4. To compare the UK approach with international frameworks, particularly the European Union’s text-and-data mining exceptions and the EU Artificial Intelligence Act’s transparency requirements.
5. To identify gaps in the current policy framework and propose directions for future research that might inform more balanced regulatory approaches.
Methodology
This dissertation adopts a literature synthesis methodology, systematically reviewing and analysing contemporary academic scholarship, policy documents, and legal commentary to address the stated research objectives. Literature synthesis represents an established methodological approach within legal and policy research, enabling researchers to identify patterns, assess evidence quality, and develop comprehensive understandings of complex regulatory landscapes (Snyder, 2019).
The research drew upon a comprehensive literature search encompassing peer-reviewed academic journals, government publications, and reports from reputable international organisations. Primary databases consulted included Semantic Scholar, Westlaw, and HeinOnline, with supplementary searches conducted through Google Scholar and institutional repositories. The search strategy employed Boolean operators combining key terms including “text-and-data mining,” “copyright,” “artificial intelligence,” “generative AI,” “model training,” “United Kingdom,” and “policy.”
Inclusion criteria required that sources be published in English, address UK copyright law or policy directly or through comparative analysis, and focus on the period from 2021 onwards to capture the most recent policy developments. Sources were excluded if they failed to engage substantively with legal or policy analysis, if they originated from non-authoritative websites, or if they predated significant regulatory changes.
The analytical framework employed a stakeholder analysis approach, categorising interests according to four primary groups: copyright holders and creative industries; technology companies and artificial intelligence developers; the academic and research community; and public interest advocates. This categorisation enabled systematic assessment of how different stakeholder interests have been represented and addressed within policy outcomes.
Evidence from the literature was assessed according to strength and consensus, with particular attention to empirical claims about policy effects. Where scholarly opinion diverged, the dissertation presents multiple perspectives while identifying the weight of evidence supporting different positions.
Limitations of this methodological approach include the reliance upon published academic commentary, which may lag behind rapidly evolving policy developments, and the potential for selection bias in literature identification. These limitations are mitigated through triangulation with primary government documents and attention to the most recent available scholarship.
Literature review
The legal framework for text-and-data mining in the United Kingdom
The United Kingdom’s legal framework for text-and-data mining is primarily governed by Section 29A of the Copyright, Designs and Patents Act 1988, which establishes a narrow exception permitting text-and-data mining for non-commercial research purposes. This exception, introduced in 2014, requires that users have lawful access to the works being mined and prohibits the transfer of copies to others for non-research purposes (De La Durantaye, 2025; Rättzén, 2025).
The scope of Section 29A has been subject to considerable academic debate. Scholars have noted that the exception’s limitation to non-commercial research creates significant uncertainty for academic institutions undertaking mixed-purpose research or for collaborations between universities and commercial partners (Johnson, 2024; Ørstavik, 2025). The requirement for lawful access has been interpreted to mean that researchers must have legitimate rights to access the underlying works, typically through subscription or purchase, before text-and-data mining activities can proceed lawfully (Thongmeensuk, 2024).
Critically for the current debate, Section 29A provides no exception for commercial text-and-data mining activities, including the training of artificial intelligence models by technology companies. This means that commercial artificial intelligence developers must either obtain licences from rights holders or risk copyright infringement when using protected works for model training (Rosati, 2025; Kretschmer, Margoni and Oruc, 2024).
The failed expansion of text-and-data mining exceptions
In 2021, the UK Intellectual Property Office consulted on proposals to expand the text-and-data mining exception to include commercial uses, recognising the potential benefits for artificial intelligence innovation and competitiveness. This proposal would have aligned the UK more closely with the approach adopted in some other jurisdictions and would have permitted artificial intelligence developers to mine copyrighted works without individual licensing agreements (De La Durantaye, 2025).
However, the proposal encountered substantial opposition from creative industry stakeholders, including publishers, visual artists, musicians, and their representative organisations. These stakeholders argued that a broad commercial exception would undermine their ability to license content and receive fair remuneration for the use of their works in artificial intelligence training (Thongmeensuk, 2024; Balch, 2024). The House of Lords Communications and Digital Committee explicitly cited potential harm to creative industries as justification for recommending against the proposed reforms (De La Durantaye, 2025).
By 2023, the government had abandoned its plans to expand the exception, representing a significant policy victory for the creative sector. The decision was widely interpreted as reflecting the superior lobbying capacity of established creative industries compared with the more fragmented technology sector (Chesterman, 2024; Kiyani, Idrees and Fatimah, 2025). As a result, the licensing-first approach remains firmly embedded within UK law.
Stakeholder interests and influence in policy formation
Academic literature has extensively analysed the stakeholder dynamics shaping UK artificial intelligence policy. Copyright holders, represented by well-organised industry associations and collective management organisations, have demonstrated considerable capacity to influence policy outcomes. Their arguments have centred on three primary claims: that unrestricted text-and-data mining would deprive creators of revenue from licensing; that generative artificial intelligence poses existential threats to creative markets by enabling the production of derivative content at scale; and that the moral rights of authors to control how their works are used deserve legal protection (Balch, 2024; Chesterman, 2024).
Technology companies have advocated for broader exceptions, arguing that licensing requirements create transaction costs that impede innovation, particularly for smaller firms unable to negotiate comprehensive agreements with multiple rights holders. They have further contended that artificial intelligence development generates broader public benefits that justify some limitation of copyright exclusivity (Tyagi, 2024; Vesala, 2023).
The academic and research community occupies an intermediate position. While researchers benefit from the existing non-commercial exception, many have expressed concerns about the uncertainty surrounding collaborative and mixed-purpose research, and about the use of academic outputs by commercial artificial intelligence systems without clear compensation or attribution mechanisms (Kochetkov, 2025). The journal Nature has editorialised that artificial intelligence firms must “play fair” when using academic data in training, highlighting unease within the scholarly community (Nature, 2024).
Public interest perspectives have been less prominently represented in policy debates. Some scholars have argued that broader exceptions could facilitate research and education, enhance access to information, and promote the development of beneficial artificial intelligence applications (Ørstavik, 2025; Johnson, 2024). However, these arguments have not significantly influenced recent policy outcomes.
Comparative international frameworks
Comparative analysis illuminates the distinctive features of the UK approach. The European Union’s Digital Single Market Directive, adopted in 2019, introduced two text-and-data mining exceptions: Article 3 provides a mandatory exception for research organisations and cultural heritage institutions, while Article 4 establishes a broader exception for any entity, provided that rights holders have not expressly reserved their rights through appropriate means (Quintais, 2025; Margoni and Kretschmer, 2021).
The EU framework thus represents a different balancing of interests, permitting broader text-and-data mining while allowing rights holders to opt out of commercial uses. The effectiveness of this opt-out mechanism remains contested, with scholars questioning whether it provides meaningful protection for rights holders or merely creates barriers for artificial intelligence developers (Hacker, 2021).
More recently, the EU Artificial Intelligence Act has introduced transparency requirements for general-purpose artificial intelligence systems, mandating disclosure of training data summaries to enable rights holders to identify potential infringements (Buick, 2024; Quintais, 2025). These transparency measures represent an alternative regulatory approach that focuses on information provision rather than ex-ante licensing requirements.
In the United States, the question of whether artificial intelligence training constitutes fair use under copyright law remains unsettled, with ongoing litigation likely to provide significant precedents. The flexibility of the fair use doctrine potentially permits broader uses than the UK’s specific exceptions, but also creates legal uncertainty for artificial intelligence developers (Gans, 2024; Dornis and Stober, 2025).
Proposals for reform and alternative models
Academic literature has proposed various mechanisms to address the current tensions between copyright protection and artificial intelligence innovation. Extended collective licensing has attracted particular attention as a potential middle ground that could provide fair compensation to rights holders while reducing transaction costs for artificial intelligence developers (Axhamn, 2025; Liu, 2025). Under such models, collective management organisations would negotiate licensing terms on behalf of rights holders, with revenues distributed according to established principles.
Statutory licensing represents a more interventionist approach, whereby legislation would permit artificial intelligence training in exchange for remuneration determined by regulatory bodies rather than individual negotiations (Liu, 2025; Senftleben, 2024). Proponents argue that statutory licensing could ensure fair compensation while enabling beneficial uses of copyrighted works.
Technical measures, including machine-readable opt-out mechanisms mirroring EU developments, have been proposed as a means of respecting rights holder preferences while permitting broad access to willing licensors (Quintais, 2025). The UK government has indicated interest in developing voluntary codes of practice for licensing negotiations between rights holders and artificial intelligence firms (De La Durantaye, 2025).
Critics of the current UK approach have warned that overly restrictive policies risk stifling innovation, particularly among smaller UK-based technology firms unable to afford large-scale licences (Kretschmer, Margoni and Oruc, 2024). Some scholars have suggested that jurisdictional competition may lead artificial intelligence developers to locate their training activities in more permissive legal environments (Rättzén, 2025).
Discussion
The evidence synthesised in this dissertation supports the conclusion that copyright holders and the creative sector are currently prevailing in the UK’s artificial intelligence policy debate concerning text-and-data mining and model training. This conclusion requires careful examination against each of the stated research objectives.
Assessment of the current legal framework
The first objective sought to examine the current legal framework governing text-and-data mining and artificial intelligence training in the United Kingdom. The analysis confirms that Section 29A of the Copyright, Designs and Patents Act 1988 provides only a narrow exception for non-commercial research purposes, with no equivalent provision for commercial artificial intelligence development. This framework places the United Kingdom among the more restrictive jurisdictions in its treatment of text-and-data mining for artificial intelligence training.
The practical implications of this framework are significant. Commercial artificial intelligence developers must obtain licences from rights holders before using copyrighted works for model training, creating substantial transaction costs and potentially limiting the range of materials available for training purposes (Rosati, 2025). The requirement for individual licensing negotiations favours well-resourced technology companies capable of concluding comprehensive agreements while potentially disadvantaging smaller firms and start-ups (Kretschmer, Margoni and Oruc, 2024).
Stakeholder dynamics and influence
The second objective focused on analysing stakeholder dynamics and relative influence in shaping UK policy. The evidence strongly indicates that creative industry stakeholders have exercised greater influence than technology companies or public interest advocates in recent policy developments.
The most compelling evidence of this influence is the government’s reversal of its proposed commercial text-and-data mining exception. The abandonment of this reform, following lobbying from creative industry stakeholders and the explicit intervention of the House of Lords Communications and Digital Committee, represents a decisive policy victory for copyright holders (De La Durantaye, 2025; Thongmeensuk, 2024). The strength of this evidence is substantial, with multiple independent sources confirming both the sequence of events and the causal relationship between industry lobbying and policy outcomes.
Technology companies have been less successful in advancing their preferred policy positions. While they have articulated arguments for broader exceptions, these arguments have not translated into legislative or regulatory change. This may reflect several factors, including the relative fragmentation of the technology sector compared with established creative industries, the reputational challenges facing major technology platforms, and the political salience of protecting creative workers and traditional cultural industries.
The academic and research community, while benefiting from the existing non-commercial exception, has had limited influence over the broader policy debate. Concerns about the use of scholarly outputs in commercial artificial intelligence training have been articulated but have not generated policy responses (Kochetkov, 2025). Similarly, public interest arguments for broader access have been acknowledged but not prioritised in policy outcomes.
Evaluation of policy outcomes
The third objective required evaluation of whether recent policy outcomes favour particular stakeholder interests. The analysis confirms that outcomes have decisively favoured copyright holders over other stakeholders.
The preservation of the licensing-first approach ensures that rights holders retain control over commercial uses of their works for artificial intelligence training. This control extends to both the decision whether to license and the terms upon which licences are granted. In economic terms, this allocation of rights enables rights holders to capture value from artificial intelligence development, potentially including both upfront licensing fees and ongoing royalties.
However, the extent to which this theoretical control translates into actual compensation for creators remains uncertain. Individual creators, particularly those without significant bargaining power, may struggle to negotiate favourable terms with major technology companies. The benefits of the licensing-first approach may therefore accrue disproportionately to large rights holders and collective management organisations rather than to individual artists and writers (Balch, 2024).
The evidence regarding potential costs of the current approach is more moderate in strength. Scholars have warned that restrictive policies may impede innovation and disadvantage smaller technology firms, but empirical evidence of these effects remains limited (Kretschmer, Margoni and Oruc, 2024; Tyagi, 2024). Similarly, claims about the necessity of copyright protection for sustaining creative markets are plausible but lack robust empirical foundation in the specific context of artificial intelligence training.
Comparative analysis and international perspectives
The fourth objective involved comparing the UK approach with international frameworks. This comparison reveals that the United Kingdom has adopted a more restrictive stance than the European Union, which provides broader text-and-data mining exceptions subject to rights holder opt-out, and potentially than the United States, where fair use doctrine may permit some artificial intelligence training without licensing.
The transparency requirements introduced by the EU Artificial Intelligence Act represent an alternative regulatory approach that the UK has not yet adopted. These requirements may help rights holders identify unauthorised uses of their works but do not fundamentally alter the balance between copyright protection and artificial intelligence innovation (Buick, 2024; Quintais, 2025).
The comparative analysis also highlights the possibility of jurisdictional competition, whereby artificial intelligence developers may choose to locate their training activities in more permissive jurisdictions. This phenomenon could undermine the effectiveness of restrictive UK policies while failing to generate compensatory benefits for UK rights holders (Rättzén, 2025).
Gaps and limitations in current policy
The fifth objective focused on identifying gaps in the current policy framework. Several significant gaps have been identified through this analysis.
First, the current framework may inadequately address the interests of smaller rights holders who lack the bargaining power to negotiate effective licensing terms with major technology companies. The benefits of licensing may therefore be concentrated among larger rights holders and collective management organisations.
Second, the framework provides limited attention to public interest considerations, including the potential benefits of artificial intelligence for education, research, and the development of beneficial applications. While the non-commercial research exception partially addresses these concerns, its scope and clarity remain contested.
Third, the effectiveness of proposed reforms, including voluntary codes of practice and transparency requirements, remains uncertain. These mechanisms have been discussed in consultations but have not been implemented, and their likely impact is difficult to assess without empirical evaluation.
Fourth, the framework has not resolved questions about the training of artificial intelligence systems by academic institutions for purposes that straddle the commercial/non-commercial divide. This uncertainty may impede beneficial research collaborations between universities and industry partners.
Implications for theory and policy
The findings of this dissertation have significant implications for both academic theory and policy practice. From a theoretical perspective, the analysis contributes to understanding of stakeholder influence in intellectual property policymaking. The success of creative industry stakeholders in shaping UK artificial intelligence policy illustrates the continued importance of organised industry lobbying in regulatory outcomes, even in contexts where alternative considerations might seem compelling.
The analysis also contributes to debates about the appropriate balance between intellectual property protection and technological innovation. The UK’s choice to prioritise copyright holder interests over potential innovation benefits reflects particular normative commitments about the value of creative work and the importance of sustaining creative markets. Whether these commitments are well-founded, and whether they are optimally implemented through current policies, remains open to debate.
From a policy perspective, the findings suggest that ongoing consultations should give greater attention to the interests of stakeholders who have been less effectively represented in policy debates to date. This includes individual creators without significant bargaining power, smaller technology firms, the academic community, and the broader public. Mechanisms such as collective licensing and transparency requirements may offer opportunities to address some of these gaps, but their design and implementation will require careful attention to ensure effectiveness.
Conclusions
This dissertation has systematically analysed whose interests are prevailing in the United Kingdom’s artificial intelligence policy debate concerning copyright, text-and-data mining, and generative artificial intelligence model training. The evidence strongly supports the conclusion that copyright holders and the creative sector are currently winning this policy contest.
The five stated objectives have been addressed through comprehensive literature synthesis. First, the current legal framework has been examined, revealing that Section 29A of the Copyright, Designs and Patents Act 1988 provides only a narrow non-commercial research exception, with no provision for commercial artificial intelligence training. Second, stakeholder dynamics have been analysed, demonstrating the superior influence of creative industry stakeholders over technology companies and public interest advocates. Third, policy outcomes have been evaluated, confirming that recent developments—particularly the abandonment of proposed commercial exceptions—decisively favour rights holder interests. Fourth, comparative analysis has situated the UK approach within international frameworks, revealing its relatively restrictive character compared with the European Union. Fifth, gaps in the current framework have been identified, including inadequate attention to smaller rights holders, public interest considerations, and the academic community.
The significance of these findings extends beyond the immediate UK context. As other jurisdictions grapple with similar policy challenges, the UK’s experience offers lessons about the dynamics of stakeholder influence in intellectual property policymaking and the challenges of balancing established industries with emerging technologies.
Future research should address several questions identified through this analysis. Empirical research is needed to assess the actual impacts of current policies on innovation capacity, particularly among smaller technology firms. Evaluation of alternative licensing models, including collective and statutory licensing approaches, could inform more balanced policy reforms. Investigation of the effectiveness of transparency requirements, as implemented in the EU context, would provide evidence relevant to ongoing UK consultations. Finally, research attention should be directed to ensuring equitable representation for all stakeholders, including individual creators, academic researchers, and the broader public, whose interests may not be adequately served by current policy arrangements.
The UK’s approach to artificial intelligence and copyright represents a significant policy choice with lasting implications for the creative industries, the technology sector, and society more broadly. While the current framework gives clear priority to copyright owners’ interests, ongoing debate means that this balance may continue evolving as new evidence emerges and stakeholder arguments develop. Continued academic attention to these questions will be essential to informing evidence-based policymaking in this rapidly evolving field.
References
Axhamn, J. (2025) ‘Extended collective licensing for use of copyrighted works for machine learning’, *The Columbia Journal of Law & the Arts*, 48(4). https://doi.org/10.52214/jla.v48i4.13933
Balch, T. (2024) ‘Copyright and generative AI’, *Journal of IP in Practice*, 1(1). https://doi.org/10.5750/jipp.v1i1.2198
Buick, A. (2024) ‘Copyright and AI training data—transparency to the rescue?’, *Journal of Intellectual Property Law and Practice*. https://doi.org/10.1093/jiplp/jpae102
Chesterman, S. (2024) ‘Good models borrow, great models steal: intellectual property rights and generative AI’, *SSRN Electronic Journal*. https://doi.org/10.2139/ssrn.4590006
De La Durantaye, K. (2025) ‘Control and compensation: a comparative analysis of copyright exceptions for training generative AI’, *IIC – International Review of Intellectual Property and Competition Law*, 56, pp. 737–770. https://doi.org/10.1007/s40319-025-01569-6
Department for Digital, Culture, Media and Sport (2022) *DCMS sectors economic estimates 2021*. London: DCMS.
Dornis, T. and Stober, S. (2025) ‘Generative AI training and copyright law’, *ArXiv*, abs/2502.15858. https://doi.org/10.48550/arxiv.2502.15858
Gans, J. (2024) ‘Copyright policy options for generative artificial intelligence’, *SSRN Electronic Journal*. https://doi.org/10.2139/ssrn.4716669
Hacker, P. (2021) ‘A legal framework for AI training data—from first principles to the Artificial Intelligence Act’, *Law, Innovation and Technology*, 13(2), pp. 257–301. https://doi.org/10.1080/17579961.2021.1977219
Johnson, A. (2024) ‘Generative AI, UK copyright and open licences: considerations for UK HEI copyright advice services’, *F1000Research*, 13. https://doi.org/10.12688/f1000research.143131.1
Kiyani, S., Idrees, R. and Fatimah, R. (2025) ‘A critical analysis of UK legal frameworks of intellectual property laws and artificial intelligence (AI)’, *The Regional Tribune*. https://doi.org/10.63062/trt/wr25.088
Kochetkov, D. (2025) ‘Science and artificial intelligence: a copyright perspective’, *International Conference on Scientometrics & Informetrics*. https://doi.org/10.51408/issi2025_047
Kretschmer, M., Margoni, T. and Oruc, P. (2024) ‘Copyright law and the lifecycle of machine learning models’, *IIC – International Review of Intellectual Property and Competition Law*, 55, pp. 110–138. https://doi.org/10.1007/s40319-023-01419-3
Liu, X. (2025) ‘Research on the adjudication of infringement in the use of works in AI data training and the statutory license system’, *Transactions on Social Science, Education and Humanities Research*. https://doi.org/10.62051/pwgck217
Margoni, T. and Kretschmer, M. (2021) ‘A deeper look into the EU text and data mining exceptions: harmonisation, data ownership, and the future of technology’, *SSRN Electronic Journal*. https://doi.org/10.2139/ssrn.3886695
Nature (2024) ‘AI firms must play fair when they use academic data in training’, *Nature*, 632, p. 953. https://doi.org/10.1038/d41586-024-02757-z
Ørstavik, I. (2025) ‘Development of large language models: copyright law perspectives for research institutions and research libraries’, *International Journal of Legal Information*, 52, pp. 200–211. https://doi.org/10.1017/jli.2024.46
Quintais, J. (2025) ‘Generative AI, copyright and the AI Act’, *Computer Law & Security Review*, 56, p. 106107. https://doi.org/10.1016/j.clsr.2025.106107
Rättzén, M. (2025) ‘Location is all you need: copyright extraterritoriality and where to train your AI’, *Science and Technology Law Review*, 26(1). https://doi.org/10.52214/stlr.v26i1.13338
Rosati, E. (2025) ‘Copyright exceptions and fair use defences for AI training done for “research” and “learning,” or the inescapable licensing horizon’, *European Journal of Risk Regulation*. https://doi.org/10.1017/err.2025.10035
Senftleben, M. (2024) ‘AI Act and author remuneration—a model for other regions?’, *SSRN Electronic Journal*. https://doi.org/10.2139/ssrn.4740268
Snyder, H. (2019) ‘Literature review as a research methodology: an overview and guidelines’, *Journal of Business Research*, 104, pp. 333–339.
Thongmeensuk, S. (2024) ‘Rethinking copyright exceptions in the era of generative AI: balancing innovation and intellectual property protection’, *The Journal of World Intellectual Property*. https://doi.org/10.1111/jwip.12301
Tyagi, K. (2024) ‘Copyright, text & data mining and the innovation dimension of generative AI’, *Journal of Intellectual Property Law and Practice*. https://doi.org/10.1093/jiplp/jpae028
UK Government (2021) *National AI Strategy*. London: Department for Digital, Culture, Media and Sport.
Vesala, J. (2023) ‘Developing artificial intelligence-based content creation: are EU copyright and antitrust law fit for purpose?’, *IIC – International Review of Intellectual Property and Competition Law*, 54, pp. 351–380. https://doi.org/10.1007/s40319-023-01301-2
