Whichever industry you work in currently, you’re likely hearing or talking about artificial intelligence – its opportunities, challenges and ultimate capabilities. The topic is dominating conversations everywhere, and research and research communication are no different. As eLife’s Editor-in-Chief Timothy Behrens says: “Science is entering a new era in which AI scientists will increasingly work alongside human scientists to enable faster and deeper discovery, as well as allow stronger connections and inferences to be drawn across the research literature.”
Photo by Kevin Ku on Unsplash
One key challenge for AI here is that today’s scientific paper was designed for human communication – not machine reading and reasoning. Modern AI systems are most powerful when they can reason over large bodies of knowledge in context, but the length and structure of a typical research article comes up against the short context windows of large language models (LLMs): articles exceed the limits of how much information these models can ingest, resulting in high computational costs and a hindered ability to help users navigate ever-increasing research outputs.
The solution to this problem is clear: we need a research communication ecosystem that allows findings to be presented in different ways – using narrative and visual modules to create a research-paper-like experience for humans, while still supporting a more granular claim-tree layout with machine-readable evidence for both human and AI scientists.
Collaborative efforts to bring this ecosystem to light would help realise AI’s tremendous potential to improve research discovery and navigation for all.
eLife’s partnerships to explore AI-driven research communication
Since its beginning, eLife has innovated in open-source technologies to support open research publishing, peer review, and assessment in the modern era. In 2023, we introduced our unique publish-review-curate (PRC) model that enables open, independent evaluation of new discoveries – and with which we have firmly established ourselves at the centre of the current ‘AI in Science’ movement.
In our model, every article we review is published as a Reviewed Preprint, a type of scientific publication that includes the article, feedback from the reviewers, a response from the author where available, and an eLife Assessment that summarises the significance of the findings and the strength of the evidence.
The transparent nature of this process has attracted significant attention from the AI in Science community: research groups are using it to create machine-readable versions of research articles to better identify relationships between claims and evidence. In parallel, we’ve been collaborating with funders, preprint servers and research organisations on open-source, AI-ready infrastructure.
Here, we provide an overview of our partnerships to harness AI for the improvement of human and machine reader engagement with research.
OpenEval for making papers machine readable
Sina Booeshaghi, Laura Luebbert, and Lior Pachter, the team behind the OpenEval project, know all too well the challenges that scientists face with the increasing volume of research articles: it is becoming more difficult for them to systematically search, evaluate, and synthesise existing work, while effective tools for navigating the scientific literature remain limited. The narrative form of these articles means that answering basic questions – such as which data and studies support a given claim or how methods differ between studies – remains labour-intensive.
To address this, Booeshaghi and colleagues developed a machine-automated approach for extracting results from papers, and assessed it via a comprehensive review of the entire eLife corpus. As reported in their preprint on bioRxiv, this method, called OpenEval, allows for a direct comparison of machine and peer review, and sheds light on key challenges that must be overcome in order to facilitate AI-assisted science.
Specifically, the results point the way towards a machine-readable framework for disseminating scientific information. As the team says: "The main take-home we hope comes across is that it's cumbersome and difficult to organise science after publication. Science should be published in a machine-readable form from the outset."
We agree. We need to be thinking about how to communicate science differently in an AI age. While narratives will always be important, there should also be smaller, data-backed, and machine-readable claims that live alongside them. This is something we will be working on in the future.
eLife Claim Trees – a prototype for the extraction and verification of claims in research papers
Alongside the work with Booeshaghi and his colleagues, we have also partnered with Zachary Mainen, Principal Investigator at the Champalimaud Foundation, Portugal, on an open-source claim extraction tool designed for AI readers. As Mainen writes, “A paper is a narrative artifact designed to convince a human reader. AI is a different reader – it needs structure it can reason over, verify, and connect to other structure.”
Claim graphs provide a solution. They decompose a scientific paper into typed propositions and the typed logical relations between them. Each proposition is one declarative sentence in an active voice, while relations are statements about logical structure. In other words, A requires B, meaning A's validity depends on B's. If a claim fails to reproduce, then invalidity propagates throughout the graph.
Mainen created a prototype based on 10 neuroscience papers published in eLife. Using his own agentic research platform called HaaK, he extracted a similar claim structure of these papers, linking each claim to data and code, and recording what happened when they ran the analyses.
The 10 papers became dependency graphs from which HaaK extracted 173 claims. This process was 90% automated by the platform, with some human oversight. Mainen says the future success of the project will be defined by a cultural shift toward rapid, modular sharing that enables faster discovery, with the potential to connect papers in a way that could lead to future collaborations.
QED Science’s ‘validity engine’ for helping researchers validate research
Earlier this year, we partnered with QED Science to develop a new approach for helping readers engage with eLife reviews.
Built by Professor Oded Rechavi at Tel Aviv University, Israel, and Dr Niv Samuel Mastboim, QED Science operates a validity engine for researchers to understand which claims hold up across scientific papers, grants, and their day-to-day work. The team released an AI Review tool in October 2025 and has since developed a suite of additional products, including a grant review tool, with the central goal of validating the underlying science within each paper.
QED Science’s software breaks down papers into their core claims. The authors' novel contributions are separated from existing knowledge and ranked against studies in their field to maximise the impact of the research when it comes to publishing it. (For more information about QED Science, see this feature piece.)
Our partnership with QED Science consisted of a small opt-in pilot, where the team provided their AI-extracted claim tree as a structure for displaying eLife's human-generated reviews of a paper, and eLife’s reviewers checked how accurately the tool assigned comments from their reviews onto its claim tree. While human reviews and AI-generated claim structures differ significantly, the goal was to test whether this new communication format could help bring them together.
Preliminary results from the pilot showed that participants marked approximately 80% of the comment-to-claim assignments as accurate, based on a relatively small sample size, while a number of comments were unassigned to any claims.
We hope this project will provide useful insight into AI-backed software for research validation purposes, and potentially pave the way for a model in which AI and human reviews can work together to improve the communication of scientific review.
An AI tool for signalling preprint quality
Finally, while not a partnership, we were also interested to see an update from Mark Hahnel, VP Open Research at Digital Science, on Preprints.ai, a multi-agent peer-review tool for signalling quality in preprints. As Mark explains, this tool moves away from questioning whether AI can help with peer review to asking instead: “if we were designing peer review from scratch for a world where powerful LLMs exist, what would we actually need humans for, and what could we comfortably automate?”
To provide an overview of preprints.ai, 11 specialist AI agents review preprints independently, then deliberate to produce a single Publication Fit tier (1–10) anchored to real journals. One of the layers of the tool’s review pipeline – multi-agent LLM review – is heavily inspired by eLife’s taxonomy, highlighting the benefits of eLife’s transparent and nuanced approach to peer review for machine-readable science.
Next steps and how you can get involved
So, we can see that our PRC model is attracting some attention from the research community and forming the basis of a growing research area in machine-readable science, kicking off with the projects outlined in this update.
While most of this work continues, we also plan to explore other projects as part of our eLife Pathways initiative. By collaborating on three emerging, real-world standards, we are exploring ways to make research more interactive, transparent, and ready for the AI age. This work includes projects such as Open Exchange Architecture (OXA) that makes preprint and paper components "composable"; Modular Interoperable Research Attribution (MIRA), which can represent preprints and papers as specific questions, claims, and the evidence supporting them; and DocMaps, which provides a provenance record and a timeline of a preprint's journey, showing exactly when it was editorially verified, peer-reviewed, who curated it, and how it evolved over time. We already use DocMaps to power eLife’s Reviewed Preprints and share peer review and curation with other organisations such as EuropePMC and OpenRxiv. We’ll have more to say on this and other projects soon.
In the meantime, if you’re exploring how AI could be used for improving research communication, we invite you to get in touch with us, tell us what you’re up to, and discuss potential options to collaborate. As with research itself, research communication is only made better with shared skills, knowledge and perspectives – especially when it comes to the vast possibilities unlocked by AI.
Further reading
- Read more about peer review and publishing at eLife, including our Reviewed Preprints and Assessments
- Find out more about eLife Technology and eLife Pathways
- eLife receives Wellcome boost to build open publishing ecosystem
- Technology at eLife: 10 years of innovation
