He Built a Library for AI Before ChatGPT Arrived. Petr Knoth on a Vision Ahead of Its Time – and Much More

Nearly twenty years ago, the idea that scientific papers would one day be read not only by people but also by machines might have sounded like science fiction. Yet Petr Knoth began exploring this possibility during his doctoral studies at the UK's Open University. His work gradually grew into CORE, a research infrastructure now used by approximately 30 million people every month. With the rise of large language models, his early vision is becoming a reality. In this interview, CORE's founder discusses the opportunities AI is opening up for scientific research, why he considers the current funding model for open access publishing problematic, and how the European EOSC AIssistant project aims to help AI agents work with scientific knowledge, including findings still locked behind publishers' paywalls.

13 Nov 2026 Karolína Smetanová

No description

Your academic career took you from Brno to the United Kingdom, where you spent much of your professional life and where the idea for CORE was born. What originally led you to bring together scientific literature and machine processing?

When I started my PhD at the Open University in 2008, I was working on natural language processing and ways to automatically identify connections between texts. I found scientific publications particularly interesting in this respect, and I have always seen them as a kind of reservoir of human knowledge. Unlike an encyclopedia, which generally offers a coherent account of a particular subject, scientific papers present different perspectives, sometimes contradict one another, and engage in scholarly debate. At the same time, it was clear that the volume of scientific literature was growing faster than any person could keep up with.

I therefore began thinking that, alongside a Scholarly Web for Humans, we would also need a Scholarly Web for Machines. In other words, an environment in which computers could work with scientific knowledge on a large scale. At the time, open access to scientific publications was also gaining ground, but those publications remained scattered across many university repositories. That was precisely why I saw an opportunity to build an infrastructure that would connect these resources and enable machines to process them.


What began as a research idea has grown into an infrastructure used by millions of people. How has CORE developed over the years, and what does it offer compared with conventional academic search engines such as Google Scholar?


In 2010, I received my first £40,000 grant, which enabled us to build the initial prototype. The project gradually expanded, and in 2015 it became a UK national service supported by Jisc. Today, CORE does much more than aggregate openly accessible papers. It also works with bibliographic metadata for publications whose full texts are not freely available and provides services to research institutions and repository managers.

People are often surprised when I explain that CORE was never intended to compete with Google Scholar, which I regularly used myself while developing it. Our aim was not to create another search engine, but to enable programmatic access to scientific literature so that software applications could work with it directly. That is why we built an API from the outset.

But when I explained around 2015 why this kind of access mattered, people often asked what it would actually be used for. My answer was that one day it would allow artificial intelligence to read and process scientific texts, something many people could hardly imagine at the time. The emergence of large language models has fundamentally changed that situation. Today, people are less likely to ask why machine access to scientific literature is useful and more likely to ask whether we should be concerned about what it makes possible.


How demanding is it to run CORE, both financially and technically?

“CORE does much more than aggregate openly accessible papers. It also works with bibliographic metadata for publications whose full texts are not freely available, and provides services to research institutions and repository managers themselves.”

From the beginning, I could not imagine asking someone to pay simply to search for or download a scientific paper. That would directly contradict the philosophy of open access to scientific knowledge. Of course, even a service that is free for users has operating costs. CORE is used by approximately 30 million people a month, which involves substantial expenditure on servers, infrastructure and staff.

We therefore combine several sources of funding. The first is community support. More than thirty institutions contribute to CORE's operation, including the University of Cambridge and the University of Oxford. These institutions are our members and also participate in decisions about the organization's direction.

The second source is commercial companies that need access to large collections of scientific publications. These include companies developing plagiarism-detection systems, working on drug discovery, or developing large language models. Their contributions help us keep access free for individual users.


How does this combination of funding sources compare with other open access publishing models, and which approach do you personally consider the most promising?


I strongly support green open access, which means making scientific publications available through repositories. An alternative is gold open access, where openness is often funded through article processing charges, or APCs. And this is where I see a fundamental systemic paradox.

Countries that invest more in research and development generally produce more scientific publications. If they have to pay to make each paper openly accessible, they automatically bear higher costs for making scientific knowledge available to the entire world. In effect, they also finance access to research findings for countries that invest considerably less in research of their own.

Another problem is that a significant share of this money goes to a small number of large commercial publishers. Instead of building shared scholarly infrastructure, public funding often ends up in private publishing structures. And we cannot even be certain that the publications whose open access we have paid for will be preserved in institutional or national repositories over the long term.

If we invested the same resources in high-quality, interoperable repository infrastructure, the cost per article could be orders of magnitude lower. At the same time, we would be building a system that research institutions could manage and develop themselves over the long term.


In the United Kingdom, open access to publications has also been linked to research assessment and university funding. How does this system work, and what lessons could it offer other countries?

“CORE pracuje také s bibliografickými metadaty publikací, jejichž plné texty nejsou volně přístupné, a poskytuje služby samotným výzkumným institucím a správcům repozitářů.”

The UK's national research assessment exercise, the Research Excellence Framework (REF), is extremely important for universities because it affects both their reputation and the allocation of a substantial share of public funding. To be eligible for assessment, scientific papers must meet specified open access requirements, generally by being deposited in a repository. Universities therefore have a strong incentive to make their research outputs available in good time.

We also investigated the effectiveness of this approach in 2019, when Drahomíra Herrmannová, Nancy Pontika and I analyzed more than 800,000 scientific papers from repositories worldwide. Our findings showed that the UK's open access policy significantly accelerated the deposit of publications in repositories.

For other countries, I think the key lesson is that open access need not depend solely on the voluntary efforts of individual researchers. When it becomes part of research assessment requirements, universities have a concrete reason to create the conditions for its systematic implementation.

Such an approach must, however, be accompanied by high-quality repository infrastructure and sustainable long-term funding. I do not know the Czech system well enough to assess its specific arrangements, but I believe the underlying principle is broadly transferable.


For researchers themselves, however, open access publishing can mean additional administrative work. What would need to change to make sharing research outputs easier?

“For other countries, I think the key lesson is that open access need not depend solely on the voluntary efforts of individual researchers. When it becomes part of research assessment requirements, universities have a concrete reason to create the conditions for its systematic implementation.”


I fully understand why some researchers see open access as yet another administrative burden. The current system of scholarly publishing is inefficient in many respects. Researchers often have to enter the same information about a single paper into numerous systems, including the publisher, the funding agency, the university, and the institutional repository. They repeatedly enter virtually identical metadata, even though it should be enough to enter the information once and have the systems exchange it automatically.

That is exactly the kind of problem we are trying to address through CORE. But it is important to emphasize that this administrative burden is not caused by open access itself. It stems from the fragmentation of the scholarly communication system as a whole.

In the age of artificial intelligence, open access is becoming even more important. It is no longer enough for a paper to be available to a person with a subscription to a particular journal. We also need scientific knowledge to be available for machine processing.


Speaking of machine processing, how close are we to a point where AI will not simply analyze scientific literature but will begin contributing to scientific discoveries itself?

“It is no longer enough for a paper to be available to a person with a subscription to a particular journal. We also need scientific knowledge to be available for machine processing.”

I think we are gradually moving towards that point. AI can already significantly speed up certain parts of research, particularly the analysis of scientific literature. Every PhD student knows what it is like to spend perhaps an entire year at the beginning of a research project trying to understand the existing knowledge in their field. This is precisely where AI can save an enormous amount of time.

I supervised a doctoral student who worked on developing software for systematic literature reviews. These are especially important in medicine, for example, where the findings of a large number of studies have to be assessed to establish whether a particular treatment actually works.

But it does not stop at processing existing knowledge. Over the past two years, dozens of systems described as AI scientists have emerged. They attempt to automate further stages of the scientific process, such as generating new hypotheses, designing experiments and evaluating results. I see enormous potential in this. In the future, AI could help researchers both understand what we already know and discover what we do not yet know.

For now, however, we cannot speak of AI as a full replacement for scientific research. It is a fascinating development, but one that also raises many questions and concerns.


The new European EOSC AIssistant project, in which CORE is involved, also focuses on the use of AI in research. What specific tools will the project develop?

EOSC AIssistant is a three-year European research project with a budget of approximately €9 million, bringing together 12 organizations from 7 European countries. It aims to develop a new generation of AI assistants to support scientists at various stages of the research process.

These are not simply chatbots where you ask a question and receive an answer. We want to create AI agents: systems capable of carrying out more complex tasks independently or semi-independently, from searching and analyzing scientific literature to generating hypotheses and proposing experiments.

CORE has a significant role in the project. Our team is the second-largest in the consortium, and I personally lead the first work package, which focuses on the project's data foundations. We will explore how to make scientific publications and other research data accessible, connect them, and prepare them for use by artificial intelligence. Our work will also include developing an AI agent for systematic literature reviews.

Alongside tools with broad applications, the project will also develop specialized agents for biodiversity, energy systems, and materials engineering.


One challenge remains: a  substantial share of scientific literature is not publicly accessible. How can AI agents work with knowledge that is still locked behind publishers' paywalls?

“EOSC AIssistant je tříletý evropský výzkumný projekt s rozpočtem přibližně devět milionů eur, do kterého je zapojeno dvanáct organizací ze sedmi evropských zemí. Jeho cílem je vyvinout novou generaci AI asistentů, kteří budou vědcům pomáhat v různých fázích výzkumného procesu.”

That is one of the most interesting research questions in the entire project. We want to investigate the extent to which European legal exceptions for text and data mining can help. An important distinction here is between scientific knowledge itself and the way that knowledge is expressed in publications. Copyright generally protects a particular creative expression, not the underlying facts, ideas or scientific discoveries.

If, for example, we extract factual information and the relationships between facts from a paper and represent them in a knowledge graph, the question arises as to how far that knowledge can then be used independently of the original copyrighted text.

That certainly does not mean we can automatically process or distribute any content. Legal exceptions come with conditions, including the requirement to have lawful access to the source documents. We therefore want to explore technological and legal solutions that will allow AI systems to work with as broad a body of scientific knowledge as possible without having to make the protected texts themselves available.


As AI develops, concerns about its misuse are also growing. Where do you see the boundary between beneficial and risky uses of these technologies?

It is much the same as with any other major invention. A technology is not inherently good or bad; what matters is how people use it. Artificial intelligence could, for example, help us develop new cancer treatments, but it could also be misused to develop biological weapons. The same capabilities that can accelerate scientific discoveries can pose a danger in the wrong hands.

Cybersecurity research provides a useful analogy. When experts try to break a particular encryption method or security mechanism, they are not necessarily doing so to cause harm, but to identify vulnerabilities that need to be fixed. Yet an attacker can exploit exactly the same knowledge.

I do not think organizations such as CORE can solve this problem on their own. Responsibility must also rest with lawmakers, governments, and other institutions. Regulating artificial intelligence is exceptionally complex, and I think even experts are not yet entirely sure what the optimal approach should look like.


Besides the potential misuse of AI, there is also the question of how reliable its outputs are. How do you plan to ensure that researchers can trust the results of the tools developed within EOSC AIssistant?

This is precisely why the EOSC AIssistant project places considerable emphasis on Responsible AI for Research. We want to develop transparent systems whose outputs can be verified and whose scientific sources can be clearly identified.

Another important principle is human-in-the-loop: maintaining human oversight of decision-making and the interpretation of results. We must be able to distinguish between established scientific knowledge and newly generated hypotheses. Just because AI can produce a convincing-sounding statement does not mean it is a scientifically verified conclusion.

Personally, however, I see an enormous opportunity in these developments. Artificial intelligence could accelerate drug development, research into new materials and the search for solutions to challenges in energy and the environment. To realize this potential, we need not only high-quality AI systems but, above all, open, interoperable and trustworthy infrastructure for scientific knowledge. And I consider that one of the most important goals of EOSC AIssistant.

“important principle is human-in-the-loop: maintaining human oversight of decision-making and the interpretation of results. We must be able to distinguish between established scientific knowledge and newly generated hypotheses.”

No description

Petr Knoth


is a researcher specializing in data science, machine learning, text mining, information retrieval, digital libraries, and Open Science. In 2011, he founded CORE (core.ac.uk), one of the world's largest aggregators of open access scientific publications, and has led the platform ever since. He has experience managing international research and development teams and has contributed to more than 25 national and international projects. Over the course of his career, he has published more than 100 peer-reviewed scholarly papers and secured over $4 million in research and development funding. His work has also been covered by media outlets including Nature, Times Higher Education,  and Physics Today. Previously, he worked as a senior data scientist at Mendeley/Elsevier, focusing on information extraction and recommendations of scientific content.


More articles

All articles

You are running an old browser version. We recommend updating your browser to its latest version.