Today, we continue the conversation started yesterday, where we Asked the Chefs: As participants in scholarly publishing, and with the benefit of 20/20 hindsight, what went right, and what could have/should have been done differently?
This time, the Chefs are joined by special guest Steve Smith, founder of STEM Knowledge Partners and an independent consultant advising scholarly publishers and research organizations.
Ashutosh Ghildiyal
If we had known 20 or 30 years ago what the world of scholarly communication would look like today, would we have done anything differently with open access or open science? Perhaps. The more interesting question is what we might have anticipated about the world that would follow.
The world in which the open access movement took shape was very different from the one we inhabit today. We are connected through our phones, laptops, and the internet. We expect to find information quickly, get answers almost immediately, and move between sources and platforms with little friction. Our attention is also increasingly fragmented. There was a time when information was scarce, and scarcity gave it value. Today, information is abundant.
AI has accelerated this shift. Systems can ingest vast amounts of information and use it to analyze, synthesize, transform, and generate content, often at little or no cost. In that sense, open access was almost inevitable. Once information could be digitized, searched, copied, and distributed at virtually no marginal cost, maintaining access as a significant barrier was always going to become difficult.
I don’t think the lesson is that we should have done less on open access. We might instead have anticipated what would come after access. Open is now ubiquitous and expected. Simply making information available is no longer, by itself, a distinctive source of value. That changes the question for scholarly publishing: If information is abundant and increasingly available to machines as well as humans, where does value reside?
I would argue that it lies increasingly in the creation, validation, interpretation, and preservation of knowledge. The valuable thing is the discovery of something new: an insight, observation, method, connection between ideas, or finding that challenges an established paradigm. Equally important is the ability to recognize its significance. For researchers, this means asking meaningful questions, distinguishing signal from noise, identifying patterns, challenging assumptions, and making discoveries. For publishers, it means supporting the integrity and longevity of the scholarly record through rigorous validation, reliable preservation, accurate machine-readable content, provenance, authentication, and effective discovery as the technologies through which research is accessed continue to evolve.
This becomes particularly important as AI becomes a primary mediator between researchers and scholarly content. A researcher may increasingly encounter the literature through an AI system that finds, summarizes, compares, interprets, and synthesizes research. This raises an important question for publishers: What role do we play when AI becomes part of the interface between researchers and the scholarly record?
I see two important opportunities. The first is to create and maintain trusted, authenticated sources of scholarly knowledge that machines can reliably understand and use. Publishers have decades of experience with peer review, editorial processes, metadata, persistent identifiers, versioning, archiving, and content integrity. These capabilities become increasingly important when machines consume scholarly information at scale. AI systems need to know where information came from, whether it is authentic, which version they are using, and what evidence supports it.
The second opportunity is to help researchers develop the capabilities needed to work effectively with AI. Researchers will need to know how to formulate questions, interrogate AI outputs, recognize limitations, verify sources, protect research integrity, and exercise critical judgment. They will also need to understand how AI can influence discovery, interpretation, and the development of ideas.
This points to an emerging need for cognitive AI literacy: the ability to understand how AI interacts with information and knowledge and to use these systems thoughtfully and responsibly in research. Building this literacy is a shared responsibility. Publishers, libraries, universities, research institutions, and researchers all have a role to play. Libraries and institutions bring established expertise in information literacy, research skills, and scholarly communication. Publishers bring expertise in the scholarly record, research integrity, content provenance, metadata, and the systems through which research is created and disseminated. Working together, these stakeholders can help researchers develop the capabilities needed to navigate an AI-mediated research environment.
This also presents an opportunity for publishers to create new value. They can work with libraries and institutions to develop guidance, educational resources, training, and practical frameworks that help researchers evaluate and use AI effectively and responsibly. In doing so, publishers can contribute to the development of research capabilities alongside their traditional roles in curating, validating, preserving, and disseminating scholarly knowledge.
If we had predicted today’s world 20 or 30 years ago, perhaps the thing we would have done differently is to think earlier about what comes after openness. We might have invested sooner in authenticated and machine-readable content, richer metadata and provenance, interoperability, research integrity, and the capabilities researchers need to navigate an environment of abundant information. The fundamental challenge has shifted. The challenge today is knowing what to trust, what to understand, what to question, and what to do with the knowledge available to us. That may be where the next chapter of scholarly publishing begins.
Author’s Note: I dictated the initial draft using AI and then used AI to edit the transcription. I reviewed and approved the final version.
Alison Mudditt
PLOS was founded on a simple but radical idea: science works best when it’s open. That has never changed in spite of the very different world PLOS now operates in 25 years later, and it’s still our north star as we navigate today’s many challenges. Back in that context, APCs were a reasonable bet to get OA off the ground. PLOS authors were in the biomedical fields — and in the Global North — with healthy research grants that could easily afford a modest fee for universal access and reuse. The success of our first journals, PLOS Biology and PLOS Medicine, proved that open and excellent could coexist, and both PLOS and OA hit a rapid growth curve.
There’s much that we and others got right. OA has grown to a significant proportion of scientific output — as much as 50% or more in some countries. That access has transformed the ability of millions around the world to read the latest science: researchers in low- and middle-income countries (and institutions in the Global North), clinicians, policymakers, journalists, and curious citizens all benefit from the growing body of openly available literature. As a leader who has been part of this movement for a couple of decades, I’m deeply proud of everything that has been achieved. That said, my 20/20 hindsight glasses show that while the goal was right, we got a number of key mechanisms wrong.
I’m going to focus on what I see as the two critical structural issues that have hindered progress. One of the core goals of OA was to democratize access to knowledge, but inequality not only exists, but it has also arguably grown. The paywall has moved from reader to author, with the Global South bearing the brunt of this shift, and where waivers only treat the symptom, not the underlying structural issues:
- APCs have entrenched privilege.
- “Transformative deals” have preserved the power of legacy publishers and are not transforming the system.
- Metrics and incentives still reward prestige over openness.
These are problems that have been building for at least a decade. PLOS, among others, proved the APC model could work in that specific moment: when library budgets were growing, when researchers had APC funding, when the volume and geography of research looked fundamentally different. And PLOS too was slow to course-correct: we treated the early mechanism as the permanent answer for too long. Haseeb is right (in his comment yesterday) that publisher margins went unstressed and that the cause is structural, though I disagree that this means that openness itself has failed.
The second key failing of APCs is that they tied publisher revenue to the same article counts that drive researchers’ careers, so that incentives for both are pulling in the same distorted direction. Researchers behave rationally within irrational systems: they publish for career survival, not community benefit. For the most part, institutions and funders still evaluate based on narrow, article-based metrics. And publishers have built their models around these same distortions — though we have more agency than many like to admit and should stop fueling article growth via our own activities. The misaligned incentives are a key reason why the sector hasn’t course-corrected, and why article growth has fed the integrity crisis.
At PLOS, we’ve spent the past two years deep in research and learning to understand what that better future might look like (details here). Moving “beyond the article” doesn’t mean abolishing it — it means decentering it to better reflect the modular, iterative, and collaborative nature of 21st-century science. That’s at the heart of PLOS’s Knowledge Stack concept: the goal isn’t just technical innovation; it’s cultural realignment. By crediting all forms of contribution, we begin to reshape what “success” looks like in science.
Decentering the article also requires decoupling revenue from the article. We have been experimenting with APC alternatives for over five years and are now ready to shape a wholesale replacement. As we do so, we have four core design principles:
- Recognize a wide range of research contributions
- Transparent and predictable pricing structures
- No per-published-unit (article) fees
- Informed by local geographic and economic situations
Collective models already work at meaningful scale in parts of the system (SCOAP3, Open Library of Humanities, S2O at some publishers). The open question is whether they can work for a large multidisciplinary publisher such as PLOS. As we’ve learned from our partnerships with Jisc and cOAlition S, the models only work if libraries, funders and publishers design them together.
Our R&D work reinforced something that we underestimated 25 years ago: financial models are not simply operational mechanisms. They shape who can participate in research, what gets rewarded, and how widely the benefits of open science are shared. It’s such a cliché, but that matters more now than ever. With the US moving to stop funding publishing from research grants, and budgets under pressure worldwide, there is a real risk that without viable alternatives, research drifts back behind paywalls. And challenges such as climate change, food security, and trust in science can’t be met by limiting access to those with institutional affiliations and financial resources. The goal of open access was and is right, and we now share the awesome responsibility of libraries, funders, institutions, and publishers of fixing the mechanisms.
Author’s note: I worked with Claude to summarize and organize my thoughts from a range of my other talks and written pieces. The comments themselves are written without AI.
Steve Smith
When open access was taking shape, the problem in front of us was clear: too many people could not read the research they needed. Removing that barrier was the right thing to do. But Giorgio Gilestro’s essay, The Weimar of Knowledge, names the opposite problem we did not foresee: a world in which scholarly-looking text could be produced far faster than any human community could read or judge it. The supply of such material outputs may grow almost without limit; human attention will not.
AI can multiply claims far more readily than it can produce trustworthy new evidence, especially where that evidence takes a year in the lab or a season in the field. Access will still matter but, increasingly, the harder question will be, “Which of these millions of outputs are worth my time, and why?” Had I foreseen this shift, I would have designed openness around two scarcities that AI does not automatically remove: trustworthy evidence and human attention.
We should also stop talking about a single “article of the future,” as though the same object can serve machines and people equally well. For machines, the most useful object is a structured research record in which claims stay attached to the evidence behind them. For human readers, the most useful form will be shaped by selection, context, narrative, and judgment. A good conference talk is more than just a paper read aloud; it is a different account, shaped around what a human audience can absorb.
One human-facing synthesis might draw on hundreds of elements in the research record, while one study might generate several structured outputs designed to be queried rather than read linearly. Open science should make as much of the underlying record available as possible without pretending that every component must compete for human attention as another conventional paper.
The separation between the machine record and the human narrative also changes how we think about preprints and journals. Disclosure, review, and certification don’t all have to happen at the same time or in the same place. Preprints can make findings available quickly. Journals and scholarly communities can then show what changed under scrutiny, and why the work is worth your attention. In time, that may mean assessing or certifying selected outputs and commissioning syntheses across many others, rather than converting every preprint into another conventional paper. A journal’s value would lie less in being first than in choosing what matters and explaining why.
Abundance also makes filtering more important. When nobody can inspect everything, the name of a journal, institution, or researcher becomes a convenient shortcut. But that shortcut favors the labs and journals that were already visible, and pushes everyone else down the queue. Open access relocated inequity more than it removed it. Abundance risks hardening what is left.
Provenance and review history offer something better. They can show what supports a claim, how the evidence was produced, and what scrutiny it has received. But neither provenance nor prestige can by itself tell us whether a finding matters; that still requires expert and community judgment. Open science should make both the evidence and the basis of that judgment visible, rather than leaving attention to inherited brands or opaque platform algorithms. In hindsight, opening the record was only half the task. We also needed to make the filters through which we navigate it more transparent.
What would I have done differently? I would have treated open access as the first stage of a larger redesign. Removing barriers to reading was necessary, but we also needed a system built for a future in which machines could generate and process more scholarly material than people ever could. I would also have treated versioning, correction, curation, and synthesis as continuing costs of openness, rather than work that somehow happens after publication. The challenge ahead is to keep the scholarly record open while making abundance navigable, without allowing prestige alone to determine what we trust or what any of us gets to read.
Author’s note: Generative AI was used as an editorial aid in reviewing drafts and copyediting. The ideas, arguments, source selection, and final decisions are entirely the author’s own.
