Every day, researchers around the world conduct experiments that don't work. Hypotheses fail. Cell cultures contaminate. Statistical analyses reveal no significant results. These moments of scientific disappointment happen countless times, but here's the problem: almost nobody ever hears about them.
When experiments succeed, they get published. When they fail, they often disappear into laboratory notebooks or the memories of researchers who eventually move on to other jobs, other institutions, or different careers entirely. The knowledge vanishes with them, and someone down the line will spend weeks repeating the exact same failed experiment, wasting time and resources on a problem already proven unproductive.
This is just one symptom of a deeper crisis in how modern science preserves and shares knowledge. According to a new perspective published in Nature Communications, the research community faces a perfect storm of challenges: scientific work is being generated at unprecedented scales, experiments are becoming more complex and interdisciplinary, yet the systems we use to document and disseminate findings remain fundamentally broken.
"Science is losing knowledge it cannot afford to lose," the authors write. The consequences extend far beyond individual frustration. Wasted resources multiply. Discovery accelerates toward duplicated effort instead of novel insights. And the scientific record becomes increasingly distorted, shaped more by what journals decide is publishable than by what actually happened in laboratories.
The Scale of Modern Science Has Outpaced Our Ability to Document It
The problem didn't exist in this form even a decade ago. Until recently, science moved at a pace that traditional publication could manage. A researcher would spend months or years on a project, publish the results, and move forward. But the landscape has transformed dramatically.
Laboratory automation now enables experiments at scales previously unimaginable. High-throughput methods can generate thousands of data points where researchers once gathered dozens. Computational power has exploded, allowing mathematicians and biologists to build simulations that produce vast libraries of theoretical predictions. Data has become not just bigger but fundamentally different in character: modern research generates heterogeneous datasets that combine wet laboratory experiments with computational simulations, often spanning multiple scientific disciplines in ways that traditional journals struggle to accommodate.
The result is a knowledge preservation infrastructure that was never designed for this reality. Peer-reviewed journals impose strict page limits, forcing researchers to abbreviate methodologies and omit crucial contextual information. Supplementary materials exist, but they're often incomplete, poorly maintained, and difficult to locate. Version control systems like Git work well for code, but they rarely connect to experimental protocols, metadata, or the practical insights that determine whether an experiment actually works.
Perhaps most troubling is the loss of people. High turnover in research, driven by competitive job markets and short-term contracts, means that expertise walks out the door constantly. When an experienced researcher departs, they take with them subtle technical insights, methodological refinements, and knowledge of what has already been attempted and failed. This tacit knowledge—the stuff that's difficult to write down but absolutely essential for success—exists nowhere but in that person's experience.
The Publication Bias Against Failure
Negative results represent a particularly stark version of this problem. If an experiment shows that a hypothesis is wrong, it rarely gets published. If a computational model turns out to be unstable, that finding typically stays silent. Meta-analyses end up skewed toward positive results, inflating apparent effect sizes and creating a misleading sense of scientific consensus. Machine learning models trained on this biased literature inherit the distortion, potentially amplifying it.
Only a handful of journals like PLOS ONE and the Journal of Negative Results in BioMedicine actively encourage publication of null results. This creates perverse incentives: researchers undertake procedures that have already been shown to fail because there's no accessible record of previous failures. Biomedical researchers, computer scientists, and molecular biologists alike repeat unsuccessful experiments simply because the work was never shared with the broader community.
The authors of the Nature Communications perspective emphasize that this selective preservation extends far beyond individual projects. It undermines the cumulative nature of scientific progress itself. Every wasted repetition of a failed experiment is intellectual effort that could have been directed toward more promising avenues.
Building Better Infrastructure for Knowledge
The good news is that solutions already exist in fragments. The Protein Data Bank preserves protein structures. GenBank stores genetic sequences. Repositories like Addgene distribute plasmids and molecular tools. GitHub has transformed how computational researchers track code and collaborate. These systems demonstrate that when standards are established and curation is strong, researchers will use them.
The challenge is that these tools function in isolation. They use inconsistent interfaces and incompatible standards. Most researchers discover research outputs through published papers, where links to data and code are often buried in footnotes or mentioned inconsistently. Maintenance falls through the cracks. Code that was functional at publication becomes obsolete as underlying computational environments change.
The research community needs an interconnected ecosystem. This would involve several coordinated changes.
Moderated platforms for sharing negative results could capture the value in failed experiments without demanding the formal peer-review process that traditional journals require. These platforms would assign permanent digital identifiers to research outputs, making them citable and formally recognized. Moderation would focus on methodological soundness rather than novelty or significance, allowing valuable negative findings to contribute to the scientific record.
Federated storage systems could replace the dream of a single monolithic repository that faces inevitable challenges of scale and sustainability. A distributed approach, with storage maintained across laboratories, research groups, and existing platforms, reduces overhead and increases resilience. Importantly, it distributes responsibility for curation between human experts and intelligent computational systems.
Interconnected and searchable systems would weave together preprint servers, institutional repositories, domain-specific databases, and community wikis into a comprehensive web. Metadata repositories could serve as connective layers, linking null results and supplementary outputs to published articles. Tag-based systems, similar to those used in online fan fiction archives, could help establish common terminology across complex and dynamic fields.
The Human Element: Documentation, Time, and Incentives
Technology alone won't solve this problem because the underlying issue is cultural and organizational. Researchers have limited time, and most won't devote substantial hours to knowledge preservation efforts unless they're recognized and rewarded for doing so.
Several practical changes could help. Protected time for knowledge preservation activities, similar to the 10 days per year allocated for professional development in some regions, could legitimize these efforts as core academic work. Institutions and funders could recognize contributions to knowledge preservation in hiring, promotion, and grant decisions. Publishers could require authors to release complete datasets, analytical code, and supporting documentation as standard practice.
Educational initiatives matter too. Workshops and summer schools combining hands-on training in documentation and data management with reproducibility at their center could build competence across career stages. Peer-mentoring programs could help researchers adopting new preservation practices. Cross-disciplinary training is essential: bioscientists need foundational skills in software development and AI, while computational scientists need greater fluency in domain-specific research contexts.
Artificial Intelligence as a Tool, Not a Replacement
Artificial intelligence offers potential for reducing the burden on individual researchers. AI-powered systems could automatically generate initial drafts of documentation, flag broken links and deprecated software dependencies, and surface relevant prior work including negative results that might otherwise remain hidden. AI could identify optimal mentor-mentee pairings based on research interests and career stages, and could keep software systems functional as programming languages and libraries evolve.
But the authors emphasize a crucial principle: AI should augment human capabilities rather than replace human judgment. An IBM training manual from 1979 stated that "a computer can never be held accountable, therefore a computer must never make a management decision." This remains true. Important decisions about what knowledge is valuable, how it should be organized, and what it means in context require human expertise.
A Future Built on Community
The path forward depends fundamentally on people. Scientific societies must continue fostering networks through conferences, workshops, and summer schools where researchers learn techniques, share experiences, and build professional relationships. Standards cannot be imposed from above; they must emerge from the collective wisdom of the research community itself.
Workshops should prioritize inclusive participation, actively targeting new attendees to prevent insular expert groups. Sessions should address both technical specifics and broader questions of incentives and governance. Cross-disciplinary collaboration should be actively encouraged, recognizing that biologists, computational scientists, and engineers approach problems in fundamentally different ways.
Community-driven guidelines should take the form of living documentation, version-controlled resources that can be updated systematically as new technologies emerge and community consensus shifts. This approach preserves access to historical versions, allowing researchers to understand how practices have evolved.
The vision, ultimately, is straightforward: knowledge preservation should become an integral part of the research process itself, not an afterthought. An interconnected ecosystem would link experimental outputs, computational simulations, null results, tacit expertise, and published findings into a coherent, navigable web of scientific knowledge.
Success depends on coordinated effort across institutions, funders, scientific societies, and individual researchers. It requires acknowledging that a generation of scientists will need to invest time in these new practices, and that the intrinsic value of establishing their effectiveness must be recognized before reward structures can be expected to catch up.
The stakes are clear. The knowledge generated today must remain accessible, interpretable, and useful for generations to come. Without action, science risks becoming less efficient, less equitable, and less resilient than it needs to be.
Credit & Disclaimer: This article is a popular science summary written to make peer-reviewed research accessible to a broad audience. All scientific facts, findings, and conclusions presented here are drawn directly and accurately from the original research paper. Readers are strongly encouraged to consult the full research article for complete data, methodologies, and scientific detail. The article can be accessed through https://doi.org/10.1038/s41467-026-72667-3






