Research and innovation at risk from copyright-centric AI regulation

• Why it matters: European research and innovation are being held back by copyright policies designed for the Gutenberg press, not laboratories. While AI has the potential to drive breakthroughs in medicine, climate science, and countless other fields, researchers face mounting restrictions on accessing and using the data they need. Publishers and entertainment companies are seeking to reshape copyright policy around their very specific and narrow interests, in ways that will all but kill off large parts of AI research, threatening Europe's scientific competitiveness just when AI leadership matters most. The stakes aren't just academic: this is about whether Europe will lead or lag in the technologies that will define the next decade.
• Key result: European researchers are operating under artificial constraints that fundamentally compromise their work. Copyright-driven restrictions on training data don't just slow down research; they make AI models less safe and more biased. Meanwhile, researchers in more innovation-friendly jurisdictions freely access the data they need to build better AI systems. Europe risks creating a two-tier system where its own scientists become second-class citizens in the global race for AI advancement, undermining both scientific progress and economic competitiveness.

Major challenges for research and education

• There's more to AI and GenAI than the entertainment industry's concerns: The vast majority of AI is domain- and task-specific, from spam filters to medical imaging analysis. Policy debates that focus mainly on the entertainment industry's concerns around GenAI overlook the fact that AI is both a crucial tool for research and a product of it.
• 'Copyright capture' of the AI debate: The entertainment industry has captured the AI debate. This narrow focus creates a distortion that shapes policy in ways that risk, at best, neglecting, and at worst, harming research and innovation.
• The link between TDM and AI is being challenged: Some publishers have denied the connection between Text and Data Mining (TDM) and AI, despite the EU AI Act confirming that EU TDM provisions apply to general-purpose AI models, including GenAI. Such interpretations risk further restricting Europe's ability to keep up in AI technology, sacrificing tomorrow for the sake of yesterday.
• Limitations on AI use: KR21 observes with increasing concern growing pressure from publishers (and collecting societies) for universities and their students and researchers not to make use of AI technologies, including basic research-supporting tools.
• Data access restrictions: The quality of AI models depends on the three 'Vs': velocity, volume, and veracity of data. Restrictions on data access raise the risk of biased, inaccurate, and unsafe AI models. Such restrictions result in "garbage in, garbage out," meaning that limiting training data undermines the very safety and fairness objectives policymakers aim to achieve through AI legislation.
• Collaboration barriers: AI development is inherently collaborative, often involving partnerships between universities, research centres, and industry across borders. The current copyright framework in the EU, particularly the artificial divide between commercial and non-commercial research, creates a stumbling block for these essential collaborations.
• Financial burdens of research: Current research efforts already face a significant financial load, as publicly funded knowledge is being paid for multiple times. Potential data mining costs imposed by rightsholders could add to the initial (public) funding, access costs, and sharing costs. Such quadruple-dipping is not only inefficient but also fundamentally unjust and anti-innovation.

Legal context: Text and Data Mining (TDM)

• Europe: In EU copyright law, the term TDM is used to cover all AI and machine learning technologies. The 2019 Directive on Copyright in the Digital Single Market (CDSM) explicitly defines TDM as:
  "Art 2 (2) 'text and data mining' means any automated analytical technique aimed at analysing text and data in digital form in order to generate information which includes but is not limited to patterns, trends and correlations."
  This broad definition is complemented by Articles 3 and 4 CDSM, which introduce two mandatory exceptions and limitations to copyright for TDM: one specifically for the purposes of scientific research (Art. 3) in a narrow non-commercial context, and a broader one (Art. 4), though the latter allows rightsholders to opt out. Most collaborative research will fall under Article 4 CDSMD and so face the likelihood of opt-outs. The EU AI Act makes it explicit that Articles 3 and 4 CDSMD apply to general-purpose AI models, including generative AI models (see Recital 105 and Art 53.1 (c)).
• UK: The UK has a TDM exception limited to non-commercial research (Section 29A CDPA). UK policymakers are currently engaged in a debate on leveraging AI to boost growth and global competitiveness. However, KR21 has growing concerns about overly restrictive copyright policies that could sideline the significant contribution of R&D and research-intensive industries to the economy.

1. Research-first AI policy

• The needs of science and research must underpin AI policy formulation.
• What's needed:
  • Evidence-based, domain-specific regulation: Avoid one-size-fits-all approaches driven by single industry interests. Approaches targeted at proven problems, sector by sector, offer the best way forward.
  • Research cannot be an afterthought: Consider research needs in all legislation from the outset, not as an afterthought, and conduct proper impact assessments considering effects on research. Learn from jurisdictions where flexible copyright supports both innovation and creative sectors.
  • Focus regulation on outputs, not inputs: Address legitimate concerns about AI outputs (such as substantial similarity in generated content) rather than restricting inputs to training processes.
  • Move beyond a zero-sum game: Stop prejudging that copyright is the optimal solution. Instead, look at the relative merits of other tools. Such an approach, focused on real policy goals, is far less likely to lead to a zero-sum or even negative-sum game than one centred around copyright, which risks benefiting neither the creative sector nor Europe's wider competitiveness.

2. Promote a pro-research and modern information access regime

• Restrictions on data access not only risk driving research and innovation abroad but also harm AI safety and increase bias, undermining policymakers' own safety and fairness objectives.
• What's needed:
  • Volume and veracity matter: Recognise that greater access to diverse, high-quality data leads to better, safer, and less biased AI models. The quality of AI models depends on the three 'Vs': velocity, volume, and veracity of data. Restrictions on legitimately accessed materials not only harm research but society at large.
  • End the opt-out system: The EU's current opt-out mechanism not only risks increasing bias and creating poor AI in certain sectors but also reinforces market concentration and limits competition by favouring only well-resourced players. The UK still has an opportunity to avoid this trap of opt-outs and licensing models.
  • Promote cross-border collaboration: Support research partnerships by ensuring data can be shared across borders for legitimate research purposes without legal uncertainty.
  • Remove the artificial commercial/non-commercial divide: The EU must abolish this unnecessary distinction hindering research-industry collaborations and undermining knowledge valorisation, a key EU policy goal. The UK should extend its TDM exception to commercial uses. Prevent researchers in the EU and UK from being at a severe disadvantage compared to innovation-friendly jurisdictions.

3. Safeguard against excessive interference undermining scientific research

• Restrictions on data access not only risk driving research and innovation abroad but also harm AI safety and increase bias, undermining policymakers' own safety and fairness objectives.
• What's needed:
  • No monitoring by rightsholders: Transparency in research should focus on reproducibility and validation, which are core principles of good scientific practice, not on external monitoring by content owners or their representatives. Research ethics frameworks developed within academia are better suited to address AI challenges than external oversight.
  • Support reproducibility and validation while protecting academic freedom: Enable researchers to retain necessary data and models for verification purposes, contrary to publisher contract terms that require destruction of training materials after project completion.
  • Uphold academic freedoms and human rights: Protect researchers' right to pursue their goals without constant monitoring and questioning. Resist attempts by some rightsholders to control all aspects of AI development and deployment cycles.

4. Address market failures and protect research institutions and students

• Universities and research institutions face unfair bargaining power in negotiations with publishers, who have significant leverage to impose restrictive terms around text and data mining, as well as use of AI tools, that risk harming research and innovation.
• What's needed:
  • Strengthen protections against unfair contracts: Provide legal protections for universities and research organisations similar to consumer protection laws, safeguarding them from exploitative contract terms.
  • Prevent contracts from overriding researchers' rights: Ensure contracts cannot override or narrow TDM exceptions and their application for AI development. TDM is a core research method, not a threat.
  • Address publisher attempts to control AI research and use: Address the growing problem of publishers using licence agreements to restrict AI use in ways going far beyond copyright law, including attempts to prevent the use of research-supporting GenAI tools like Microsoft Copilot, to deny the connection between TDM and AI, or to impose unlimited liabilities on institutions for errors made by researchers and users.

Background | KR21 Knowledge Briefs based on our Reports:

• Copyright and Open Norms in Seven Jurisdictions: Benefits, Challenges & Policy Recommendations