For generations, the scientific manuscript has been optimized primarily as a human-readable narrative: structured text containing an abstract, introduction, methods, results, figures, tables, and references, accompanied by supplementary files where necessary. Although journals may ultimately convert accepted manuscripts into PDFs and other publication formats, the material submitted by authors is increasingly more complex than a document alone. Modern research can generate vast datasets, sophisticated computational workflows, versioned code, and machine-learning models, while researchers increasingly rely on AI systems to search and synthesize the literature. The manuscript is therefore becoming less a self-contained narrative and more an interface to a network of interconnected research objects. Standards such as NISO's Journal Article Tag Suite (JATS) illustrate how journal content can be represented in structured, machine-readable form, while the FAIR Principles emphasize that research data should be findable, accessible, interoperable, and reusable.
This shift also raises the possibility that future manuscripts will be written simultaneously for humans and machines. Structured abstracts, standardized terminology, persistent identifiers, machine-readable metadata, and explicit links between claims and underlying evidence could allow computational systems to interrogate the scientific literature at a much deeper level than conventional text search. Rather than simply retrieving papers containing particular keywords, AI systems could potentially trace a finding to its experimental dataset, identify the statistical method used, compare it with related studies, and assess whether subsequent research has supported or challenged the original conclusion. For authors, this may mean that clarity and reproducibility increasingly depend not only on how well a study is explained, but also on how effectively its underlying research objects are structured and connected.
Peer review is likely to change alongside the manuscript itself. Automated systems can already assist with checks for statistical inconsistencies, image manipulation, plagiarism, missing reporting information, inappropriate references, and other potential problems. More sophisticated systems could eventually perform standardized methodological and computational checks before a manuscript reaches a human reviewer. This would not necessarily make peer review obsolete; instead, it could shift the role of reviewers toward questions that remain difficult to automate, such as whether the experimental question is important, whether the interpretation is biologically convincing, and whether the evidence genuinely supports the broader scientific claims. The reviewer of the future may therefore spend less time detecting procedural errors and more time evaluating scientific judgment.
For researchers, the implication is that preparing a manuscript may increasingly involve building a reproducible research package rather than simply writing a compelling narrative. Authors may need to document software versions, analytical environments, data provenance, model versions, reagent information, and computational dependencies alongside conventional experimental details. Journals, in turn, may need to rethink what constitutes the “version of record” when datasets, code, and analyses can continue to evolve after publication. The PDF is unlikely to disappear soon, but it may become only one view of a much richer scientific object. The research paper of the future could be simultaneously a narrative, a dataset, a computational workflow, and a machine-readable evidence map—changing not only how scientists publish, but also how scientific knowledge is discovered, evaluated, and reused.
Click here for the Japanese version.