Editors’ note: Today’s guest post is by Ben Kaube and Steve Smith. Ben is the co-founder of Cassyni, a platform for helping publishers harness events, video, and AI to engage their journal communities. Steve is the founder of STEM Knowledge Partners and an independent consultant advising scholarly publishers and research organizations on AI-era strategy, content licensing, and the trusted use of scholarly content. Reviewer credit to Chef Alice Meadows.
In a post last year, we imagined a researcher receiving a useful scientific answer from an AI tool without visiting a journal website, crossing a paywall, or downloading a branded PDF. The “front door” of an article has now moved far enough to raise a harder question: what happens to the bundle of functions that used to sit behind it when the researcher asks not just for an answer, but whether that answer can be trusted?
Suppose that the AI assistant draws on the published paper, its dataset, the methods, and the citation and correction record. It picks up a recent conference session where the finding’s central assumption was questioned. It notices that the lead author is giving a seminar next week, and offers to register the researcher.
This would be a good answer: evidence, context, expert reaction, and a route into the conversation. In some ways, it may be a better starting point than the article alone.
However, it also raises an awkward question for the organizations whose work made the answer possible. Which part(s) of that answer were they actually trying to own?
Who supplied the corpus the assistant drew on? Who made the context legible, so the assistant could tell what the material meant and whether to trust it? And who convened the community whose questions, debates, and peer judgment shaped what the finding now appears to mean?
For most of its modern history, scholarly publishing answered all three at once. Journals organized review, distributed content, certified its place in the record, and sat within communities that societies, conferences, and editors helped convene. Corpus, context, and community formed a bundle, with each reinforcing the others, even when they ran on separate systems.
AI is beginning to pull that bundle apart. No organization can assume equal strength across corpus, context, and community. It must decide where it still has an advantage, what it needs to keep doing itself, and where it can rely on partners.
The Bundle Comes Apart
We’ve seen unbundling before. Music went from whole albums to individual tracks; newspapers saw classifieds, crosswords, and reporting split off into separate digital tools. Scholarly publishing has had its own remarkably durable bundle in the journal: discovery, dissemination, peer review, citation, and community. AI now allows each of those parts to be encountered separately.
The bundle always looked more unified than it was. The earliest journals grew out of scholarly communities: they recorded what was read and argued at the meetings of the first societies. Convening remained part of the enterprise, but the record became the business. The article was the product, the journal the container, and the publisher or society made the whole thing possible and visible.
AI is more than just one more discovery channel. It changes the relationship between the parts. An assistant can retrieve content from one source, interpret it using signals from several others, and deliver the result inside an interface that belongs to none of them. The reader may get a useful answer without ever learning which journal published the work, which society developed the field, or which researchers supplied the criticism that made the answer reliable.
The scale of this shift is now being measured rather than debated. The cross-industry Taming the Crocodile study is quantifying what zero-click discovery does to usage, while a recent Scholarly Kitchen post by Angela Cochran and Todd Toler drew the distinction that matters here: between content sold once for model training, where it becomes part of a system’s underlying statistical machinery, and content accessed repeatedly as live context in an ongoing, licensed relationship in which use can be measured.
But one question has largely dropped from view: who owns the interface where the researcher enters their question? When we wrote about the moving front door a year ago, that still felt contested. It now looks increasingly settled, and not in publishers’ favor. That leaves publishers and societies with a less comfortable question: what remains recognizably theirs when research is discovered, interpreted, and used elsewhere?
The components still matter, but they can now travel separately. A society may host the session where a result is first challenged, yet the AI assistant learns of the challenge only through a paper published elsewhere. A publisher may invest heavily in review and integrity, yet downstream systems treat the article as undifferentiated text because those judgments were never encoded in a portable form. What breaks is the automatic reinforcement between corpus, context, and community.
The bundle can therefore be taken apart economically as well as technically. Different actors can capture value from different layers. So the useful question is not which function matters most, because all three do, but what each one actually is and how and why they pull against one another.
Corpus, Context, and Community
Corpus makes research available; context makes it interpretable; community is where new judgments form. In practice, the same scholarly initiative may do all three. A guest-edited special collection begins with convening, as editors bring a field together around a question. Its framing and selection provide context, while the articles themselves become part of the corpus.
A corpus is the body of material that people and machines can reach and use: articles, books, datasets, figures, proceedings, transcripts, and other research outputs. Its value grows with breadth, consistency, and machine accessibility. Much of the visible AI activity sits here, including via publisher platforms such as Wiley’s AI Gateway, research workspaces such as Elsevier’s LeapSpace and Claude Science, and services such as Cashmere that provide controlled access to licensed content.
The corpus is most exposed to commoditization when it is separated from the context and convening that make it authoritative. When content is consumed through someone else’s interface, the supplier can disappear inside the product. A high-value journal article may be expensive to license, while an open access review, preprint, or secondary summary reproduces much of its substance soon afterwards. Breadth and access still matter, but may not be enough to preserve visibility or pricing power.
Context matters because the gist of a paper can be restated in seconds; the assurance behind it cannot. An AI summary of a paper can capture what a paper says, but it rarely contains enough information to judge whether it should be trusted. Some of that missing context is factual: provenance, review status, version history, corrections, data availability, and rights. Other context is evaluative: whether the claim was novel, whether the evidence held up, and whether later work supported it. Review articles, editorials, commentaries, and citation patterns provide some of this, but it is captured and maintained unevenly. Without it, an AI tool can summarize what a paper says without knowing how it should be interpreted.
Community brings together the people, evidence, and questions through which judgment forms. It is sustained through conferences, seminars, editorial networks, working groups, and informal exchanges.
In a research community, who is speaking matters. A challenge from a respected methodologist or a cautious endorsement from an editor carries weight because peers know the speaker and can judge the remark accordingly. That kind of authority comes from the community around the research, and it cannot be manufactured simply by producing more content.
The Trilemma: Where Corpus, Context, and Community Pull Apart
These three layers are entering the AI era at very different speeds. Corpus is already being licensed and ingested at scale. Tools to enrich it with context are starting to emerge. Community is barely represented at all.
That may be less a lag than a clue: it is the layer whose value resists tokenization. Some traces of community activity belong in the scholarly record, but much of its value comes from living human exchange: who is speaking, how others react, and where the exchange takes place.
Corpus and context: scale against scrutiny. A corpus grows simply by adding more research. Context has to be built around that research in different ways: tracking provenance and rights information, linking data, or adding expert commentary. Some of this can be captured as structured metadata; some depends on editorial or expert judgment. Applying this consistently across a large back catalogue is expensive, and maintaining it turns publication into an ongoing stewardship obligation.
Automation can recover much of what is already on the page: models can extract entities, link citations, and identify relationships across a document. But they cannot recover context that was never recorded. Researchers in a subfield may know the cell line behind a contested result, which findings have quietly failed to replicate, or that an apparently novel claim appeared decades ago. That knowledge belongs to the field, but not always to the formal record. It circulates through conference questions, review discussions, seminar exchanges, and the accumulated judgment of people who know the area well.
Because bridging that gap requires real human effort, not every paper will warrant the same treatment. Publishers will need to decide where additional enrichment can be justified.
Corpus and community: consumption against participation. Reaching readers through third-party systems can expand the reach of the corpus while making the originating organization less visible. Search often still delivers the reader to a journal page, where a relationship can begin; an assistant may remove that encounter altogether. The corpus is used, but no relationship forms between the reader and the source organization.
That matters because vibrant communities depend on participation. Communities hold together when people present, review, organize, mentor, and return to the same spaces over time. Content can travel easily; communities cannot. The more content is consumed in third-party systems, the less reason readers have to return and participate.
The significance of new work is often recognized socially before it becomes visible bibliometrically. If a result rumored to be announced at the annual meeting would fill the room, that tells us something no citation count can yet show. In those cases, convening reveals the value of the corpus before the formal metrics catch up.
The corpus may travel without carrying much connection to the community that produced and evaluated it. A reader may get the answer without any route into the discussion or expert network around the research. In our opening vignette, the assistant surfaces an upcoming event, and the corpus becomes a route into a relevant community rather than a substitute for it. That does not happen by default. Events, people, and topics need to be described in discoverable ways, and connected to the research they surround. Societies feel this most sharply because their advantage was never simply owning content. Large publishers face the mirror image: buying an event or platform does not automatically also transfer the community’s willingness to participate.
Context and community: what can be recorded, and what cannot. Much of what we call context can be attached to a research object: who produced it, what evidence it rests on, whether it has been reviewed, and whether it has since been corrected. Community judgment is harder to separate from the exchange in which it was made.
A post-publication discussion may reveal that a result rests on a disputed assumption. An audience question can expose an error that later becomes a correction. These exchanges are valuable precisely because they create context that did not exist before, so the temptation is to capture more of them: transcribe the recordings, link the questions to the papers, give presentations identifiers, fold expert commentary into the record. PubPeer offers one longstanding example: informal judgment, made permanent and searchable.
Capturing these exchanges creates a real opportunity, but it can also change the conditions under which communities work. In seminars, conferences, and working groups, people often think aloud. They speculate, test ideas, and change their minds; a sharp comment may be informative without representing anyone’s settled view. Make every remark permanent and machine-readable, and the price of speculation rises to match the price of publication.
A prepared presentation is not an offhand answer, and a recorded keynote is not a closed working group discussion. Some exchanges can reasonably become part of the record; others should remain ephemeral or be summarized only with consent.
Taken together, these tensions show that corpus, context, and community do not naturally move in the same direction. A larger corpus can make context thinner; context can travel without the community that produced it; and some of what happens within a community loses value if it is made permanent and machine-readable.
What Remains Recognizably Yours?
A journal portfolio, platform, or membership base is no longer enough on its own. As AI transforms how research is discovered, every organization will need to decide where its real strength lies, what it must build, what it can let go, and where it can rely on partners.
Corpus, context, and community will all still matter, but they do not need to belong to the same organization or appear in the same place. When the reader never reaches your interface, what remains recognizably yours?
Authors’ note: Generative AI was used as an editorial aid in drafting, revising, and copy editing this article. The ideas, argument, source selection, and final decisions are entirely the authors’ own, and all factual claims were reviewed by the authors.
Discussion
13 Thoughts on "Guest Post — The Unbundling of Scholarly Publishing: A Trilemma for the AI Era"
Nice article. If I’m reading you right, you’re mapping “context” mostly onto content and the metadata and signals around it. That’s useful, but I think context is also a **process** category. Editorial thinking and decisions, provenance, review status, version history, corrections etc are all generated through editorial and production workflows.
That distinction matters because it may be where the larger AI risk sits. To replicate the value publishers provide, AI platforms need to learn not just the corpus, but the processes that produce and shape it.
And prompts, instructions, decisions, and workflow context fed into AI systems are precisely the signals from which those processes can be learned. If that’s right, the risk isn’t only disintermediation of the corpus. It’s platforms absorbing the **how**, not just the **what**.
Thanks, Adam. I think that is exactly right, and a very useful extension of the argument.
We probably treat context in the piece mostly as the layer that makes content interpretable: provenance, correction status, review signals, commentary, and so on. But as you say, those signals are themselves generated by editorial, production, and community processes.
That may be where the deeper risk sits. If AI platforms learn not only the corpus but the workflows, prompts, decisions, and judgment patterns that produce context, then the issue is not just content capture but process capture.
For me, that points back to the importance of the live human layer: the continuing capability to convene expertise, make accountable judgments, revise them, and maintain trust over time. Static context can be absorbed. The harder thing to replace is the living process that keeps producing it.
You raise a good point. The three constituents were framed mainly around what the researcher encounters, which abstracts away the processes that produce them. That said, I’d expect many of those processes to undergo a similar unbundling too.
Thanks, Steve, Ben.
Just switching back to your point mapping content as context… it’s also probably worth asking where that content-and-signal generation actually starts. Researchers are already using large LLM providers well before a paper exists – during ideation, lit review, drafting, analysis. So a good deal of that context (=value) is being formed and absorbed at pre-authoring and authoring stages, long before a publisher’s platform or AI Gateway ever sees the work.
That suggests publishers may need to reach further back than the interface layer this piece is mostly addressing — into the researcher’s actual lab and desktop workflow, with tooling that’s genuinely more useful to them there than a generic LLM. Otherwise that early layer just defaults to whichever provider the researcher already has open. It’s entirely possible to consider this these days as what seemed like an immense technological vision 5 months ago is now possible when it comes to tool building.
Thanks for the interesting post and discussion.
Thanks, Adam. I think that is a sharp extension of the argument. Context does not begin only after an article exists. A lot of it is formed earlier, during ideation, literature review, drafting, analysis, and revision. If those stages increasingly happen inside general-purpose AI systems, then the risk is not only that publishers lose visibility at the point of discovery. They may also lose sight of the workflow in which much of the context is being created.
That pushes the question further upstream. It is not only “how does the article travel into an AI assistant?” but also “where does the researcher do the work?” If the answer is always a generic LLM, then publishers and societies may arrive too late.
The challenge, as you suggest, is that any move upstream has to create real value for researchers, not just better visibility for publishers. But I think the underlying point is right: the risk is not just losing the interface to the article. It may be losing the workflow in which context is made.
I agree this is a valuable opportunity, though I’m less certain that scholarly publishers are best placed to operate here.
Research workflows can vary enormously between fields, so the tooling would need a lot of field-specific nuance. It also requires product skills that many publishers have not historically needed in-house.
I expect it would have different competitive dynamics compared to those many publishers are used to. The usual advantages around journal brands, content ownership and distribution do not automatically carry over into researcher workflow tools, where switching costs are low and incumbency offers limited advantage.
That does not mean publishers should stay out of it, but I suspect the route in will often be through partnerships or very focused vertical tools rather than trying to own the whole research workflow.
Seems like the context layer is where we will find the premium value and utility.
I think that’s right, Roy, though I’d add one qualification. A lot of premium utility will show up in the context layer, but context is not self-generating. It depends on editorial, review, production, and community processes that keep it current and accountable. So I’d say the value is not metadata alone, but context connected to the human processes that produce it.
This is a thought-provoking article, thanks Ben and Steve, and it calls to mind two other articles I think are relevant to your argument.
The first was published in Nature yesterday, “Reimagining research papers as interactive and reliable AI agents”, https://www.nature.com/articles/s41586-026-11044-y. As the abstract puts it: “By turning static papers into interactive AI agents, Paper2Agent introduces a paradigm for knowledge dissemination and a collaborative ecosystem of AI co-scientists.” In other words, it enables AI to substitute, at least in part, for the community you describe.
The second was written by Daniel Hook in Research Agenda, and explores what actually happens when AI becomes our collaborator in place of people: https://www.researchagenda.news/articles/the-waymo-effect.html. His conclusion is that “funders and institutions should recognise collaboration [or, in your framing, community] for what it is – not a nice-to-have, but a form of infrastructure – and price it accordingly.”
Many thanks, Rob. These are very helpful references, and they pull in exactly the right tension.
Paper2Agent is a fascinating example because it pushes the paper itself further up the stack: from static object to something queryable and partly conversational. In that sense it may well substitute for some of what we used to get by going back to the author, the lab, or the immediate research group: how do I run this, adapt it, reproduce it, or apply it?
But I’m less sure it substitutes for community in the stronger sense we had in mind. It may help make the paper active, but the field still has to decide how far the work should be trusted, whether its assumptions hold, whether it matters, and how it changes in light of later evidence or challenge.
That is why Daniel Hook’s piece feels so relevant. If AI makes it easier to avoid the friction of human collaboration, the answer can’t simply be better paper-agents. We also need to preserve and fund the human settings in which disagreement and accountable judgment happen.
So perhaps the distinction is this: the paper may become an agent, but the field cannot become just a collection of paper-agents. Community, or convening in our terms, remains infrastructure.
This is a fascinating article and it raises some very important questions about how human knowledge is created through trial and error and often chaotic processes, where once held truths are swept away by new insights. Whilst today’s research and publication process has served us well for many years, it can often present matters as finalised and publishers have relied on controlling access to knowledge to monetise the research. As with so many discussions about AI, we come back to provenance. But in the case of Scientific research that provenance extends beyond the data, to Who is the authoritative source? Who is critical of the research? Have the findings been replicated? What alternative theories are there? The provenance to all these questions becomes equally important. What potentially emerges from this is a trust framework that can score these individual components (I’m sure there are many more). With such an approach AI could play a role in the constant assessment of research adding value for researchers beyond retrieving and summarising the research itself. It in many ways has parallels with my time in the credit risk industry, which over the last 20 years has moved from initial risk assessment, to periodic reviews to today’s continual risk assessment. In such a model lies the next monetisation opportunity for whoever can build an authoritative research assessment framework.
To be clear I don’t see this as a FICO for scientific research but more a framework to support answering the question How robust is the evidence supporting this claim and what lies behind that assessment? Creating an auditable assurance system.
Thank you, Phil. This is a helpful way to frame it, especially the analogy with the move from point-in-time assessment to something closer to continuous assurance.
I agree that provenance in research has to extend well beyond the data or the article itself. It includes who is making the claim, what evidence supports it, who has challenged it, whether it has been replicated, what alternatives remain plausible, and how that picture changes over time.
I’d be cautious about reducing this to a single score, but your clarification about an auditable assurance framework feels right to me. The opportunity is not simply to retrieve or summarize research more efficiently, but to help users understand how robust a claim is, what sits behind that assessment, and how it changes as the field continues to test it.
