Should infrastructures do more than protect knowledge? My perspective is that the future of infrastructures designed to safeguard the scientific record should focus not only on protecting knowledge, but on ensuring it remains trustworthy, transparent, correctable, and governed in the public interest.

Foundational principles should be redundancy, governance, and decentralization. No single publisher, repository, institution, commercial platform, or technology provider should become a point of control or potential failure for the scientific record. Academic libraries and librarians fill a critical role and must continue to serve as trusted stewards of the scholarly record, with missions oriented toward enduring access.

Relatedly, I would like to see less commercial capture and control of knowledge and greater investment and emphasis on community-owned and governed open infrastructure. Knowledge is a global public good, and the systems that govern it should reflect and facilitate this reality. Research communities and institutions around the world need meaningful agency rather than relying on commercial platforms whose overall incentives may not align with the interests of science. Funders can advance this goal by providing sustained, rather than project-based, support for open infrastructure and aligning their policies to ensure adoption. Also, funders should use open infrastructure in their own grantmaking and research management practices. Doing so not only strengthens the open infrastructure ecosystem but also embeds transparency and trust throughout the research lifecycle —from funding and research management through

publication and dissemination.

We need to reconsider the concept of the “version of record.” Scientific knowledge is not static, yet retractions, corrections, and errata largely reproduce practices developed for print publishing. A digital record could instead be genuinely versioned: preserving the original while clearly linking subsequent corrections, updated analyses, critiques, and new supporting evidence, with transparent provenance showing what changed, when, and why rather than treating correction as a singular event applied to a static article.

Finally, we should be more ambitious about using technology to strengthen trust. Emerging technologies, including AI, introduce risks, but they can also help identify manipulation and inconsistencies, establish provenance, connect evidence, and detect problems earlier.

Ultimately, safeguarding the scientific record should mean designing better systems: open, distributed, community-governed infrastructures that treat knowledge as a global good and use technology proactively to make science more transparent, accountable, and resilient.

Ashley Farley, Senior Officer of Knowledge & Research Services, Gates Foundation

Should data repositories be funded like libraries? About ten years ago, we saw that our field needed a domain-specific data repository, so we built one, LDbase. Building LDbase was the easy part. Finding money to keep the lights on is the hard part.

Open Science Framework announced it is phasing out the project feature many researchers use to share data, citing the cost of maintaining it. The common funding model for repositories, grant funds, is not working. Grants come with term limits, forcing repositories to compete for renewal, which is never assured. Renewals also ask for innovation through new aims, when what a working repository needs is continuation.

We were recently awarded a new grant to maintain and extend LDbase. It took three years and five submissions. We were fortunate. From the start we assumed grant money would not last, and made development choices that keep LDbase cheap to run, though not free. Fewer features is the price, and the work those features would have done falls to the researchers using it. If the money ran out, our university library partner would take LDbase in as a special collection.

That is survival, not a funding model. We should fund data repositories the way we fund libraries. A library serves a unique role, giving everyone in the community access to knowledge and preserving it permanently. It is not asked to win a competitive grant every five years to keep existing. It is understood to be permanent, paid out of operating funds.

Many libraries are governed by a board of trustees, a structure designed to keep collections stable across changes in leadership and priority. Data repositories should adopt the same model, where paying the running costs is what earns a seat on the board, and no single funder leaving can close the repository. OSF had paying institutional members, but they bought a service rather than a seat.

Borealis, the Canadian Dataverse Repository, is run by Canada's academic library consortia. Member institutions pay annual fees, and those consortia hold its steering committee seats. The International Nucleotide Sequence Database Collaboration goes further. Its partners in the United States, Europe and Japan are separately funded, separately governed, and each holds a complete copy of the sequence data, so no one government can shut it down.

Operational funding should come from the organizations with a stake in the repository's survival. For LDbase, that means the international scientific societies whose members deposit and reuse these data, and the agencies and foundations that already name LDbase a recommended repository for their grantees. Those who require data sharing, and those who benefit from it, are the natural parties to share its cost.

Preserving scientific data takes continuous, skilled work that does not stop when a grant does. LDbase will need to exist in 2040. Repositories like it need what libraries have, a line in an operating budget.

Sara Hart, Canada Excellence Research Chair in Developmental Science; Professor at the University of Waterloo

What wins: AI curation or AI creation of academic research? The Shangri-La of academic publishing is to be fast, trusted, open, and cost-effective. We have rapid, open and affordable dissemination through preprints, data and software. But they lack the assurance associated with peer review, while trusted journal publishing (e.g. gold OA) remains slow and expensive. Applying transparent, auditable AI checks to open research outputs could provide a scalable new layer of verification earlier in the process. This offers a potential route to publishing that is fast, open, trusted and cost-effective at the same time. However, the other side of the coin is the use of AI tools to mass produce research outputs. We already see content published as preprints that is a rehash of what we already know in order to pump publication numbers, or pseudoscience wrapped in academic language.

Computer checks are likely to proliferate alongside computer-generated content. Both will often rely on the same models and carry similar operating costs, so fabrication and detection get cheaper together as volume rises on both sides. Screening may move beyond the submission gate and run continuously across material published years ago as well as work arriving today.

Generated content may also fill gaps in the record. Synthetic data can stand in when underlying records cannot be released for consent or privacy reasons. Negative results and failed replications, which few researchers have time to write up properly, can now be created by our robot friends. Both are useful when their provenance is declared and problematic when it is not. That puts more weight on clear labeling. 

Human checking may become a status marker. When verification is free and automatic, the scarce signal is a named person who has spent time assessing the work and is willing to attach an identifier to the outcome. Journals and repositories may begin recording who checked what and against which version, then display that information. Humans and machines can be corruptible; we see this with citation rings or nefarious actors publishing cherry-picked data. Transparency throughout can help prevent this. “Who checked what, where, and how” is a layer of provenance that will be needed for human and AI checkers. This is work. Badges posted on one platform will not travel well into other systems looking to reuse the content.

Standardized provenance of curation and checks needs to flow with the objects throughout academic platforms. In order for research infrastructures to maintain their integrity, they need to embrace new tooling as aggressively as the AI enabled researchers, looking to enhance their career prospects (or worse), using new tools.

View the interactive publishing quadrilemma

Mark Hahnel, VP of Open Research, Digital Science

Over the past several decades, it has felt that we have been playing continuous catch-up with the massive disruptions across the scholarly communication ecosystem. What these changes have shown is that this is a distributed, fragile system, one that has no truly centralized infrastructure and no universally agreed (let alone enforceable) rules or governance, for either individual parts of the system, or the system as a whole. We can see the tensions more clearly than ever as a result of an acquired function of the system — one it was never designed for — of how the academics who generate knowledge are rewarded and promoted.

Looking back, it is clear that the enormous possibilities from these disruptions, beginning at the end of the last century with online publications and followed by the possibility of universal access to academic knowledge in all its diversity, have not been accompanied by the requisite open, resilient infrastructure. As AI tools become more prevalent, we risk repeating recent mistakes if we fail to build the right infrastructure and guardrails.

What Ashley Farley, Mark Hahnel, and Sarah Hart have laid out clearly is that the organization of the current system can barely support the processes we have now. 

Organizational and financial infrastructure are fragmented and operate in silos. As Farley notes, we need infrastructure that is “trustworthy, transparent, correctable, and governed in the public interest.”, for which we require “redundancy, governance, and decentralization”. But as Hart notes, we are very far from that. Using just one subject-specific data repository as an example, she highlights LDbase’s fragile existence, such that “repositories [must] compete for renewal funding, which is never assured [whereas] what a repository needs is continuation.” She calls for ongoing funding for data repositories akin to “the way we fund libraries”.

There is no shortage of ideas for innovation in academic communication — and the pace of change is unlikely to slow. As Hanel notes, innovation is already coming with the advent of AI tools, but we don’t yet know how they will truly affect the system. What does seem clear is that trust will become even more important, especially if the escalation of AI-generated work continues. Hanel notes a potential model of “transparent, auditable AI checks to open research outputs [that] could provide a scalable new layer of verification earlier in the process.” Ultimately, it is possible to imagine an assessment system that rewards the process of research — not just the final output — underpinned by trusted AI checks that can be interrogated. He argues, and I agree, that we are, however, always likely to value “the scarce signal [of] a named person who has spent time assessing the work”. Such systems will require careful design to incorporate tools and humans optimally.

But more than ever, any system requires publicly funded and communally governed open and transparent infrastructure. If we were concerned about the private lock-in of journals, and the current patchwork of public infrastructure for data, we should be truly alarmed by a future where private companies could be allowed to capture the trust infrastructure that may come to underpin academic knowledge. 

It’s essential we start planning now for the public infrastructure we need.

Ginny Barbour, Editor-in-Chief, Medical Journal of Australia; Co-Chair DORA; Adjunct Professor, QUT