Editor’s Note: Today’s post is by Stephanie Decker. Stephanie is a professor of Strategy at Birmingham Business School, UK. Her work lies at the intersection of organization studies and historical research, focusing on using historical methodologies and digital archives for researching organizations. At the British Academy of Management, she leads on Open Access and contributes to the society’s current work on the potential benefits and limitations of artificial intelligence (AI) in research and education.

When the Open Access (OA) movement gained momentum in the early 2000s, its proponents envisioned a world where research findings would be freely available to all readers — breaking down paywalls that limited knowledge dissemination and hindered scientific progress. By the 2010s, an increasing number of countries began to mandate open access publications, arguing that publicly funded research should be freely available to the public. The readers envisaged by proponents of OA were obviously human (academics as well as the wider public). While text mining had been considered as one potential application, they could not foresee the development of large language models (LLMs) which would begin to rapaciously ingest large amounts of text. OA literature has become particularly attractive for AI training precisely because it lacks the legal and technical barriers that might protect traditionally published content.

Barriers to using copyrighted content have not, however, provided effective protection (or remuneration) to published authors in other areas. As media reporting on ongoing lawsuits has made abundantly clear, the major AI models that have transformed the digital landscape since Chat-GPT’s launch in 2023 were trained on copyrighted material on a large scale. To the casual observer, this may appear an obvious breach of copyright. Yet, transformative use cases have previously enabled Google Books’ (Authors Guild v. Google) scanning of millions of books for search functionality. The transformative use doctrine emerged prominently in 1994 with the Supreme Court case Campbell v. Acuff-Rose Music, which established that a commercial parody could qualify as fair use because it transformed the original work. In academic publishing, this has typically covered uses such as quoting portions of articles or books or repurposing content for educational purposes.

However, AI training goes further by not just indexing content, but learning to generate similar content. The ongoing litigation involving companies like OpenAI and publishers/authors may establish new boundaries for what counts as transformative use, especially whether the commercial nature of many AI systems weighs against fair use claims. While this would clearly affect paywalled academic content, the picture becomes far more complicated when we consider OA publishing and its many variations, which is increasingly becoming the norm in academic publishing.

Most open access academic content is published under Creative Commons licenses, with CC-BY being the most common. These licenses were designed with human readers and human reuse in mind:

• CC-BY: This license enables reusers to distribute, remix, adapt, and build upon the material in any medium or format, so long as attribution is given to the creator. The license allows for commercial use.
• CC-BY-NC: This license enables reusers to distribute, remix, adapt, and build upon the material in any medium or format for noncommercial purposes only, and only so long as attribution is given to the creator.
• CC-BY-ND: This license enables reusers to copy and distribute the material in any medium or format in unadapted form only, and only so long as attribution is given to the creator. (can be combined with the previous one into CC-BY-NC-ND)

In a recent consultation on OA for the UK’s Research Excellence Framework, the British Academy of Management, of which I am currently a director, and the British Academy (which represents many professional associations in the social sciences and humanities) jointly raised concerns around both the impact, and the lack of any meaningful discussion, of AI reuses of OA content. While OA licenses clearly allow for machine learning, researchers opt for OA to maximize human readership and reuse, and arguably not to provide free training data for commercial AI companies.

Current licenses, whether traditional copyright or the various CC-BY options, would benefit from explicitly considering what machine reuse means as opposed to human reuse – or indeed machine-mediated human reuse. When AI models ingest text, this is disaggregated into “tokens” (i.e., words) that are transformed into neural networks, which in turn form the basis for query responses. Text is converted into statistical patterns – which does not fit into traditional categories of copying, distribution, or adaptation as it employs aggregation on a massive scale.

This novel use of scholarly publishing highlights that significant economic value can be extracted from free academic content. It is astonishing how the UK research funding bodies, for example, can hold a consultation on OA in 2024 that pays no attention to AI models being trained on academic publications and whether non-commercial restrictions (CC-BY-NC) should be considered more widely. While OA and open data were envisaged to benefit economic development, the focus has been on “small and medium-sized enterprises and unaffiliated researchers [to] have the widest and cheapest possible access to scientific publications of the results of research that receives public funding” (European Commission Recommendation 2012). Not only do the AI companies currently developing LLMs, with their significant venture capital backing, represent a very different class of beneficiaries, their reuse shows distinct signs of significantly intervening in the academic research system in ways that makes it potentially damaging to academics and the society they serve.

It is time for OA proponents to engage in public debate with academic associations, universities and national funding agencies, because the widespread use of academic content in AI models poses significant risks for the research ecosystem. As AI becomes more deeply embedded in academic research practices, it may significantly disrupt knowledge creation and attribution standards, disrupting careers and decontextualising research insights. AI outputs, especially when subsequently cited by human researchers, become a form of “citation laundering”, where original sources are hidden or misattributed through AI generation. Not only will this make it impossible to trace the intellectual lineages of ideas, but the researchers who built foundational knowledge will no longer receive credit for their innovation. Moreover, it is likely to be particularly disruptive to interdisciplinary research, which may miss important cross-disciplinary linkages, and which is of particular relevance when it comes to addressing the global challenges that humanity faces.

Academic citation practices ensure that building on prior work is transparent and makes insights subject to critical review and fact-checking. Unlike a human author who might cite specific sources, AI systems typically don’t maintain clear links between their outputs and specific inputs. New content is generated by synthesizing from many (unattributed) sources. Consequently, the link to underlying sources is interrupted and the origin of ideas, including through recombination of existing insights, becomes impossible to reconstruct.

AI models were not built primarily for academic use, but as their impact on education has become enormous in less than two years, we need to urgently address how we are training future generations of human learners and researchers. Expanding cultural practices of instilling AI literacy and citation attribution is unlikely to be enough. Models could be expanded to track attribution of ideas to distinct academic papers and develop provenance fingerprinting – which would require a policy framework for academic publishing that mandates citation capabilities for AI models. Some of this may well be ongoing work – recent updates (February 2025) to several AI models, for example ChatGPT and Claude.AI, have led to improved citations and fewer hallucinations.

Nevertheless, the central issue remains that commercial AI companies extract significant economic value from OA content without necessarily returning value to the academic ecosystem that produced it, while at the same time disrupting academic incentive structures and attribution mechanisms. We always pay a price for technology, as the cultural critic Neil Postman pointed out in 1998, and all great technology embodies epistemological, political or social prejudices. And as AI technology is about to change everything, it is already clear that it is set to disrupt our information economy in fundamental ways: the much maligned “essay mills” that provide students with their ready-made assessments are collapsing, and business models based on exploiting the attention economy through strategically placed digital advertising or search engine optimization will likely follow suit once Google rolls out its “AI Overviews” to preface or indeed replace ranked search results. Neither are a great loss, but their demise is symptomatic of the wider problem facing any profession or industry that produces knowledge or information.

The very benefits of research that OA is supposed to make available to society at large may be undermined in the medium term. At minimum, OA licenses could mandate that AI tools trained on OA papers need to include citation capabilities in exchange for the free use of high-quality material, to ensure that the creators of academic are appropriately recognized. As OA publishing is rapidly becoming a significant share of academic output, OA policies around AI use could reshape the technologies engagement with research substantially. Otherwise, the underlying incentive structure of academic careers could be quickly eroded through widespread AI use, disproportionately affecting a new generation of researchers. Going forward, we may need to consider alternative impact indicators for AI reuse alongside traditional citation counts, if these become artificially depressed by extensive AI reuse without attribution. These are urgent issues for scholarly publishing, especially when it comes to OA, which will have an impact on future generations of research users worldwide.