Undermining Scientific Research

Blog post by Benjamin White

Why are legacy publishers trying to prevent universities from undertaking AI? (Long Read)

Over the last 12 months, KR21 has been observing with increasing concern the growing pressure coming from publishers (and in one instance a collecting society) for universities and their students and researchers not to make use of artificial intelligence technologies. 

Yes, you read that right. Our research institutions, that are not only at the heart of efforts to build productivity and address the challenges of our time, are being told that they must work with one hand tied behind their backs. They are actively being stymied in their use of AI tools to accelerate the research process – limited in using their own collections for machine learning, developed through spending reaching an estimated $ 10 billion a year on STM content alone.¹

This is not just an issue of competitiveness, but one of basic research and how it feeds into societal advancement, and even encompasses concerns around model safety. We are unfortunately already in a situation where licences, the European AI Act and copyright framework² discourage safe and unbiased AI by overly regulating and restricting access (“inputs”) to training materials.³ This is because where models are limited in their access to high-quality content it will curtail their ability to make accurate predictions. In the interests of science and society more broadly Europe should avoid a “garbage in garbage out” scenario, but on the contrary appears to be facilitating it.

This is one reason why KR21 are very concerned about restrictions being placed on universities by publishers. This phenomenon is not limited just to Europe. 2024 has seen a raft of statements from concerned universities around the globe pushing back publicly on attempts to restrict AI.⁴ Given the public funding of much research and the strength of open access,⁵ the whole situation also raises fundamental questions about the extent to which publishers should be able to control the AI research process, through efforts to influence downstream use of the databases they sell access to.

In particular, examples of contracts in the public domain have helped highlight a number of worrying phenomena:

Denying the link between text and data mining and AI: In Europe, we are aware that STM publishers have asserted that copyright exceptions for machine learning cannot be used for “AI”, based on the (baseless) claim that because copyright law references text and data mining, limitations and exceptions don’t extend to other types of AI techniques and models.

For example, this can be inferred from one contract example where Springer Nature sets out:

This simply doesn’t stand up to scrutiny. In the EU, the Directive on Copyright in the Digital Single Market (CDSM) explicitly defines text and data mining as:

Furthermore, if there was any lingering doubt that this was not intended to cover all forms of AI, the EU AI Act explicitly confirms that these provisions in European copyright law cover the development of general-purpose AI models, including generative-AI models.⁶

Turning to the UK, while the section head refers to “text and data analysis” the exception in UK copyright law itself also clarifies that wider uses are permissible by simply referring to “computational analysis”:

Attempting to prevent use of research-supporting Generative AI Tools: there are also reports that collecting societies (CMOs) in Sweden have written to a university there to say that Microsoft Copilot (a generative AI model) should not be used by students and staff for summarising third party works.

However, this is difficult to understand. The law allows for personal / research copying exceptions that would render lawful many uses of generative AI. It also provides for a “temporary copying exception” that makes temporary computational processing acts lawful for the user and for the software vendor.⁷ Such efforts would appear to show ignorance of the law, as well as a disregard for the harm that they bring to the work of universities and their researchers.

Similarly, (underlining the importance of well-crafted limitations and exceptions) Springer Nature Project DEAL contract provides that: ⁸

Attempting to prevent use in training of general purpose AI research tools: While much of the debate around Generative AI has focused on its ability to replicate cultural outputs, it is also a powerful tool in many other areas, including of course scientific and medical research.⁹ However, despite the potential for such tools to accelerate progress, publishers are holding back their development.

For example, in another Springer Nature licence, we see that amongst prohibited uses, while creation of competitive products is understandable, there is (p.11.4): ¹⁰

Anti-science and anti-safety provisions on retention of content: a particular concern that was regularly raised in the context of the negotiation of the CDSM Directive, and which is at the heart of wider discussions about trust in science, is verifiability and explainability (XAI). And yet publisher contracts can cut across this by obliging the deletion of data after the duration of a project.

The American Chemical Society’s TDM addendum provides a bad example here: ¹²

Governments globally are keen to ensure that AI is safe and doesn’t contain bias or make inaccurate predictions. Part of this requires the long-term storage of data that you have invested in creating,  in order to verify any downstream anomalies that may arise. These contractual terms however don’t support the long-term storage of data beyond the narrow confines of the project at hand, and require data the university has invested in to be destroyed. While Art 3.2 of the CDSM Directive allows those (knowledgeable copyright L&E savvy) research institutions in the EU to ignore this requirement, broadly speaking terms and conditions such as these don’t support explainable and safe AI.¹³

Undermining the knowledge valorisation agenda: despite the priority placed by governments on knowledge valorisation – the flow of knowledge and ideas from public research institutions into new products, markets and wider applications – publisher contracts seek to hold this back. Examples shared above already illustrate this point. Despite the billions spent by universities each year to support their mission, these contracts exclude any potential commercial use or commercial product development by a university or their partners.

Such provisions are also not necessarily easy to apply, given that the line between non-commercial and commercial is not as clear as the contracts indicate. Research-industry partnerships are common, while many ideas that today are purely non-commercial may one day evolve into businesses downstream.

Ironically, of course, this is not a principled stand by publishers against commercial use per se. Large publishers like Wiley, Taylor & Francis¹⁴ and OUP are now licensing “their” content to Microsoft and beyond.¹⁵ Elsevier, of course, is no longer a publisher but refers to itself as an “information analytics business.” Arguably this is a positive development in many ways because it means more high-quality content available, and so a priori safer AI for us all, but of course (most likely) only for the biggest players with the deepest pockets.

As alluded to above, given the proliferation of open access nowadays, from a rights perspective, it does raise the question of what exactly are the publishers licensing? Is it a sui generis database right? A design or type-setting right? Or is it the investment they have made (being in a privileged upstream position) in the aggregation of content which is increasingly itself openly licensed and free for all to use?

Unreasonable barriers to sharing results and developing tools: scientific publishers have built multi-billion dollar companies based on the sharing of the results of scientific research. But licences offered can now seek to hold this back in the case of TDM research. Again, the American Chemical Society’s licence illustrates this:

Yet depending on the service, a sentence from the original may be sufficient or totally insufficient. Such restrictions on access to content, much of which would have been publicly funded, limits the development of effective products and services, and of the wider communication and dissemination of research.

Similarly, the same provisions could imply that final software developed cannot contain any of the data that it has been trained on. Does this extend to a models parameters or weights? Certainly the ACS contract defines a TDM Output as including an algorithm.¹⁶ In this, we see an attempt to control the development of ML software itself by contract law – hardly something a cultural policy like copyright is really intended for.

Few guarantees for the future: clearly, technology is moving quickly, AI in particular, leading to inevitable uncertainty. This indeed provides one of the arguments for taking a proportionate and careful approach to regulation. This uncertainty, however, is also included in contracts, with provisions such as in that signed with Project DEAL: ¹⁷

Similarly, Springer Nature’s contract includes the following: (p.11.)

While adapting to legislation makes sense, the emphasis on technological developments does so less. This introduces an element of uncertainty for universities and research centres who risk being faced, at any point in the future, with requests from publishers to amend terms. Moreover, it will require the university representatives to be uniquely knowledgeable of inter alia AI technology, copyright and database right law, data protection, and the knowledge transfer agenda while robustly defending the needs of the researcher within the context of a potentially cosy working relationship with the publisher. This is no easy ask.

As will be a key theme of the conclusion, this risks exposing these institutions, which are not necessarily equipped to follow the relevant developments, and to obtain optimal results from a negotiation. 

A Sub-optimal Situation Driven by One-sided Contracts: As shown by the various statements issued from universities expressing concerns over publishers’ attempts to curtail AI research and deny the scope of exceptions, at the heart of this lies the issue of unequal and imbalanced bargaining power. This is a point that KR21 has repeatedly highlighted.

In short, a university cannot teach, and a researcher cannot research, without access to journals and books. Supported by intellectual property rights, this puts publishers in position of significant power vis-à-vis the university. While consumers are protected from unequal power relationships with businesses when it comes to a contract, bizarrely public interest educational institutions are not. As outlined in KR21’s EU Action Plan this certainly needs changing.

There are also serious concerns about the impacts on the quality and safety of AI. As highlighted already, provisions limiting access to data and data retention for validation are fundamentally hostile to efforts to improve scientific practice and its results. As underlined in the introduction, there is a direct link between the volume of (the right) content available and the quality of the outputs achieved. More restrictions on inputs (both in terms of access and how they can be used) lead to worse results.

Next steps

Machine learning is an important tool for modern day research and therefore any curtailment of it, limits the ability of researchers and publicly funded institutions to do what they are funded to do. Restrictions on computational research and AI usage curtail fundamental rights such as freedom of expression, data-driven innovation, and undermine the public task of  institutions and ultimately the public interest.  By limiting access to content, it also reduces the quality of AI outputs exacerbating bias and unsafe AI. Counter to these contractual (and in the case of knowledge valorisation in Europe policy¹⁸) trends, it is vitally important for innovation and for safeguarding the public that we seek to maximise the high-quality data a model can train on.

So what can be done?

Universities and publicly funded research organisations:

On the side of senior leaders in the university and research sectors, there is pressing need to develop a robust position around the scope of TDM exceptions and engage in AI related policy developments to defend and / or expand these rights in line with present and future needs. Restrictive clauses should be resisted as far as possible.¹⁹ Institutions should also develop a robust policy based on their exercise of limitations and exceptions, which can support research work into and using AI tools.²⁰ In this, they should draw on fundamental rights such as freedom of expression and freedom of science.

In countries where fundamental rights such as freedom of expression, as represented by the quotation exception to copyright, can be overridden by contract, this requires a delicate balancing act. Defining an institution’s approach to thresholds around the communication to the public right in the context of AI collaborations needs specialist advice. To do this, there needs to be an institutional investment in copyright and licensing policy and its application. Staff dealing with contractual negotiations must have the skills to understand as a minimum copyright, licensing, data protection and contract law. They must also be across the strategic goals of the institution – this may include university spin outs, knowledge valorisation and PPPs.

There must also be strategic engagement in government affairs. This is vital to create the optimal regulatory environment for knowledge circulation – it shouldn’t all be left to the lobbying of the entertainment industry to shape Europe’s legal information and data framework. In addition to copyright law, this should include attention to the details of gold open access knowledge discovery via metadata and the development of publicly owned aggregated services, platforms and infrastructures.

Policy Makers:

Meanwhile, the EU, national governments and the UK need to regulate better and encourage that European AI is based on maximising access to good quality data (volume and veracity). We need better limitations and exceptions to TDM that actively limit bias and maximise safe AI. Throttling inputs to AI models is in no one’s interest – the law should seek to maximise volume and veracity of data. The area of further focus should be on regulating outputs, where the real and legitimate issues arise.²¹ A starting point in the EU for discussions would be Art 4 of the CDSM Directive which restricts inputs for training in the context of knowledge valorisation and public-private partnerships, which only promotes poor AI model predictions and bias.

There also must be concrete steps to protect universities from unfair contracts in similar ways to how consumers are already protected from unequal bargaining power. This would have many benefits for AI and beyond. The ability to carry out research should not depend on the strength of a university or research centre’s contract negotiation team (even if one exists). Where there is fear of publishers, owing to their monopoly position underlined by a lack of protection against unfair contracts, institutions themselves are less able to take a robust position.

Another area to focus on is around expanding open access (and in particular open licensing). It is estimated that in 2023 $2.538 billion was paid globally on author processing charges to six publishers.²² According to Scopus 38% of all articles in 2023 were gold open access.²³ This is an awful lot of content that should be available for machine learning purposes, at least if a properly open licence is used. Surfacing this material on publicly owned infrastructures such as EOSC is vitally important. This also focuses attention on why the EU needs to introduce a secondary publishing right²⁴ as soon as practical.

It is also vital that EU AI Office ensures that freedom of expression and freedom of science in the context of university knowledge valorisation is duly taken account of in the development of the transparency guidelines required pursuant to the EU AI Act. Transparency rules must not be retroactive. They shouldn’t apply to basic research not developed with any commercial intent in mind – AI researchers are intelligent people but they are certainly not able to see into the future.

Much needs to be done to ensure that Europe makes the most of the AI revolution. This requires a strong focus on universities and publicly funded research organisations who are a source of much cutting-edge research and innovation. Governments and university leaders have an important leadership role to play to support data-driven innovation and knowledge valorisation. In an increasingly competitive world where the US and East Asia have far more technology friendly legal regimes  there is not a moment to be lost.²⁵

Further reading

PDF | KR21 Principles on AI science and research

PDF | KR21 AI Primer

Blog Post | Association of Research Libraries – Protecting AI Usage and Training Rights: Three Ideas for Congress

Blog Post | Authors Alliance – Licensing research content via agreements that authorize uses of artificial intelligence

PDF | UCB Library – Comments to the US Copyright Office’s inquiry on AI and copyright

Blog Post | Office of Scholarly Communication – Fair use rights to conduct text and data mining and use artificial intelligence tools are essential for UC research and teaching

We wish to thank Rachael Samberg (Director, Scholarly Communication & Information Policy) at Berkeley University for making us aware of many of the above links.

For universities / researchers / librarians undertaking contract negotiation on AI we recommend

Resource | Berkeley Library – E-Resource licensing explained – Chapter 6. Text and data mining

Resource | COMMUNIA – Articles 3 and 4 DCDSM on Text and data minin

3 September 2024