Scientific Papers, AI Training and Copyright: Who Controls Access to Scientific Research?

Elsevier, one of the largest and most influential corporations in scientific journal publishing, is one of five major publishers that have filed a putative class-action lawsuit against Meta, the parent company of Facebook and Instagram, which in recent years has also become a major player in artificial intelligence. The plaintiffs claim that Meta violated copyright law by using books, textbooks and scientific articles without permission to train its large language models, known as Llama models. Although several similar lawsuits are already under way, including the lawsuit filed by The New York Times against OpenAI, the company behind ChatGPT, this is the first time such a significant body in scientific publishing has taken part in a lawsuit of this kind.

The ambition of science and education publishers to control the copyrights to the content they publish is not new, and long predates the rise of language models. Academic and educational content is usually kept behind high paywalls, which prevent the general public, and sometimes also researchers at smaller and underfunded institutions, from accessing material that is important for research and science. On the internet, several repositories bypass these paywalls, either through pirate websites dedicated to that purpose or through databases voluntarily shared by researchers and others. In these disputes, courts generally tend to rule in favor of the publishers, but it is far from clear that they give sufficient weight to the public interest.

In an era in which the distinction between high-quality, reliable information and careless or even false content is becoming increasingly blurred, access to reliable sources of information and research is more important than ever. In this situation, the balance between publishers’ economic interests and the public’s need for reliable information may need to change. The vision of open science is a movement that seeks to make scientific knowledge available to everyone, by easing access to articles and even abolishing paywalls altogether. As part of this movement, major research bodies such as CERN, the particle physics research center in Geneva that operates the world’s largest particle accelerator, conduct many important studies, giving them leverage over publishers. Some use that power to promote funding models that make the articles resulting from their research freely available. More broadly, in the United States, federal policy now requires peer-reviewed publications and supporting data from federally funded research to be made freely accessible to the public.

Now, with the rapid development of the artificial intelligence market, including Meta’s activity in the field, we need to understand how this new player is changing the balance of power and the status of copyright over scientific content. In the specific context of scientific knowledge and content, the interests of AI developers and the public may even converge.

Access to Databases

Large language models rely, during training, on enormous databases in order to develop the ability to generate rich text in the context requested by the user. The information is collected in various ways. Some content organizations are already choosing to sign direct and profitable agreements with AI companies, allowing them to use their content. The academic publisher Taylor & Francis, for example, signed such an agreement with the software giant Microsoft. Under the agreement, the publisher was reportedly set to receive more than ten million dollars in exchange for granting access to its content for the purpose of training artificial intelligence systems.

Alongside agreed-upon access to content from rights holders, AI companies also routinely collect up-to-date information from the internet. Here, difficult clashes arise between content creators and the companies that collect publicly available information and use it.

The copyright question becomes even sharper when the content was collected from pirated sources available online. In the lawsuit involving Elsevier, the publishers claim that Meta copied content they own from shadow libraries such as Sci-Hub and LibGen. These libraries are huge repositories that provide access to academic content protected by paywalls and to scanned books that are not legally open to the public. Legal battles over the legitimacy of these libraries are taking place independently of the question of artificial intelligence.

The cost of academic articles and books can often reach several dozen dollars per article. Since regularly reading up-to-date articles and keeping abreast of new developments is an essential part of advancing research and human knowledge, universities and research institutes usually pay publishers subscription fees in order to give their researchers access. As a result, the main victims of the existing policy are private individuals and institutions with limited financial resources. It was in this gap that pirate libraries emerged.

Elsevier and other publishers have filed numerous lawsuits against pirate websites, and courts in the United States have ruled that Sci-Hub must pay damages. In India, where many universities operate on limited budgets, the case against Sci-Hub and LibGen sparked a particularly intense debate over the tension between copyright and access to scientific knowledge. Indian copyright law does recognize certain exceptions for research and education, but the court did not accept the argument that this justified the operation of the shadow libraries. For now, access to Sci-Hub and related sites from India has been blocked.

Fair Use For Whom?

The struggle over copyright has another side as well: the scientists themselves. Many researchers and academics support opening access to their own articles and those of others. Some even take an active part in pirate activity, even though it runs counter to the economic interests of the publishers that publish their papers. Sci-Hub, for example, builds its repositories with the help of usernames and passwords that subscribed researchers secretly provide. Under the banner “The Cost of Knowledge,” researchers have long been conducting public protest campaigns in which they declare that they will not publish articles with Elsevier, cite Elsevier journals or assist the publisher with peer review, in protest against the high prices it charges.

Scientific publishers charge high fees from researchers and scientists who seek to publish articles with them. In many cases, researchers are required to pay publication fees out of their research budgets, in sums that often reach thousands of dollars. And as if that were not enough, the payment flows in only one direction: publishers do not compensate researchers for the services they provide. In order for an article to be published in a reputable scientific journal, it must clear the hurdle of peer review — critical evaluation by other scientists who are experts in the relevant field. The reviewers, who are themselves researchers, receive no financial payment for their work. Thus, while the major publishers benefit from their status, which makes them highly desirable venues for publication, researchers and the general public are harmed by the process.

War of Giants

We live in an era in which research and theoretical papers, as well as other academic learning materials, are valuable content for human society as a whole — and also for the training of artificial intelligence models. They contain a great deal of high-quality information, and their reliability is generally higher than that of many other texts written by humans. The higher the quality and reliability of the information on which models are trained, the more reliable and useful their outputs are expected to be.

This places academic content high on the priority list of developers of generative artificial intelligence. In addition, researchers increasingly use bots based on language models while writing articles and searching for relevant research information. In this situation, researchers themselves also have an interest in ensuring that the models are trained on the highest-quality information. In practice, however, it seems that no one is really asking researchers what they want done with the copyrights to the papers they wrote. Even in the agreement in which Taylor & Francis sold Microsoft legal rights to use its content, it was reported that the authors of the content — the scientists themselves — were not involved in the deal and would not be compensated for it.

From a legal standpoint, the issue brings us back to the question of whether using scientific content to train large language models can be considered fair use, and therefore legal. Meta’s representatives argue that the company’s use of the materials is fair because it is transformative — that is, it uses the content in a way that changes it significantly compared with its original form. In 2025, U.S. courts handed AI companies some important but narrow victories on this question, ruling in certain cases that training generative AI models on copyrighted works could be fair use. At the same time, those rulings were narrow and did not settle all the major questions — especially in cases where the training material was allegedly taken from pirate databases, or where publishers argue that AI companies should have paid to license the content.

Alongside this lawsuit, many other lawsuits are under way concerning copyright infringement during the training of artificial intelligence models or in the outputs they produce. Scientific and educational content may deserve different treatment from other types of content, such as works of art or commercial material. What is clear is that the original creators — the researchers — do not appear to be taken into account in ownership decisions.

In this arena of conflict, the profits of giant commercial publishers are being weighed against those of wealthy technology companies. In a particularly optimistic scenario, one might hope that the confrontation will eventually lead to more open and accessible science. For that to happen, the decision will have to take the common good into account — and as we know, that can never be taken for granted.