Future experts and scholars will need to devote a stable share of their time and resources to building a public knowledge presence and communicating their work. For those who want their research to enter the public knowledge space and exert lasting influence, this is moving from an optional extra toward something that can no longer be entirely separated from professional work.
This does not mean that every scholar should become an influencer, nor that academic value should be determined by clicks. The deeper issue is that the structures of knowledge production, knowledge filtering and knowledge distribution are changing at the same time. As they change, the old division of labour around expertise also has to be reconsidered.
Having something genuinely original to communicate is, of course, the basic premise. The point here is not to keep restating that premise. The more important question is why people who already possess original ideas, foundational theories or systematic research should still take communication seriously. Why can the work once largely entrusted to universities, journals, publishers and specialist media no longer be left entirely to those institutions?
My view is that the expansion of knowledge-production capacity is turning discoverability into an independent and increasingly important problem. At the same time, the recipients of knowledge are changing. Researchers now need not only to make their work discoverable by the right people, but also to make it possible for AI systems to retrieve, identify, cite and use that work accurately.
Together, these changes are redefining the relationship between experts and society.
For a long time, researchers could rely to a considerable degree on an established academic division of labour. They posed questions, conducted research and wrote papers; journals and peer review handled filtering and evaluation; universities, publishers and specialist media performed part of the work of certification and dissemination. This system never guaranteed that every excellent contribution would be recognised, but it did provide a relatively clear route into the knowledge system.
AI is changing more than the speed of one step in that chain. It is entering several parts of the chain at once.
In August 2025, Wiley surveyed 2,430 researchers. Eighty-four per cent reported using AI in some aspect of their work, up from 57 per cent in the previous year, while the share using AI specifically for research or publishing tasks rose from 45 to 62 per cent. These are self-reported figures from a survey sample and should not be treated as a precise measure of global adoption. Even so, they clearly indicate that AI is moving from an experimental tool used by a minority into an ordinary part of research work. Source: https://johnwiley2020news.q4web.com/press-releases/press-release-details/2025/AI-Adoption-Jumps-to-84-Among-Researchers-as-Expectations-Undergo-Significant-Reality-Check/default.aspx
A Nature study published in January 2026 used a dataset covering 1980 to 2025, six fields in the natural sciences and 41,298,433 papers to examine the relationship between the adoption of AI tools, scientific output, career development and the overall direction of scientific research. The study found that scientists participating in AI-augmented research produced 3.02 times as many papers and received 4.84 times as many citations as those who did not. But this was an observational study. It should not be read as meaning that an individual researcher who begins using a chatbot will suddenly triple their output. The study also found an association between AI adoption and a contraction in the range of scientific topics being explored. Higher individual productivity, in other words, does not automatically mean that the knowledge system as a whole expands into more directions. Source: https://www.nature.com/articles/s41586-025-09922-y
The change is not confined to writing. The Coscientist work published in Nature in 2023 connected large language models with literature and web search, code execution and laboratory automation, enabling semi-autonomous planning and execution in bounded chemistry tasks. It still depended on human-defined goals, laboratory infrastructure and necessary intervention, but it was already doing more than polishing prose. It was participating in the research process itself. Source: https://doi.org/10.1038/s41586-023-06792-0
The more accurate description, then, is not simply that AI lets people write more. Some operations involved in knowledge production are being separated from direct human cognition and relocated into working systems composed of models, databases, tools and institutions.
I discussed this shift in After the Knowing Subject. As knowing becomes infrastructure, knowledge is no longer only something possessed after an individual has understood it. It also becomes something that external systems can continuously process, filter, judge and call upon. Human beings do not disappear from this picture, but knowledge no longer has to wait for a person to perform every intermediate cognitive step before it can continue to operate. As AI enters search, research planning, experimental work and knowledge synthesis, this claim is acquiring increasingly concrete forms. What first appears to be an epistemological transformation is becoming a practical issue of research organisation, knowledge distribution and institutional structure. Book link: https://read.amazon.com/kp/embed/?asin=B0GJ3RV4Z7
In this environment, the competitive position of experts also needs to be understood differently.
The three qualities I regard as most important are originality, foundational significance and discoverability. Originality means offering a genuinely different question, explanation or method. Foundational significance means that the originality does not merely add a small detail within an inherited framework, but reaches the concepts, assumptions, methods or even the formulation of the problem itself.
By “foundational” I do not mean that greater abstraction is automatically better. I mean work capable of changing the conditions under which later research proceeds. Reorganising the way a problem is posed is a different level of contribution from finding an answer more quickly inside an already established problem space. Incremental improvement certainly has value. But as search, calculation, comparison and expression within existing frameworks become increasingly automatable, I place greater weight on work that can reorganise the direction of inquiry itself.
Nor should we define foundational innovation in advance as an area that humans will always monopolise. If AI systems become more capable of participating in hypothesis formation, method design and theoretical exploration, the division of capabilities between humans and machines will continue to shift. Experts cannot secure their future role simply by asserting that machines can never truly innovate. Their position will have to be established through actual contribution.
The third dimension, discoverability, matters because the value of an idea and the opportunity for that idea to receive attention are not the same thing.
This problem predates generative AI. A 2021 study in the Proceedings of the National Academy of Sciences analysed roughly 90 million papers across 241 fields and 1.8 billion citation links. It found that dramatic growth in the number of papers does not necessarily refresh a field’s core ideas. In larger fields, attention can become more concentrated on already established work, making it harder for new contributions to spread gradually across the field. Source: https://doi.org/10.1073/pnas.2021636118
That study predates the recent expansion of generative AI, but it exposes an important mechanism. When the supply of knowledge grows faster than the capacity to read and evaluate it, producing more does not necessarily make new ideas easier to discover. If AI further expands the supply of knowledge, the tension between production and attention may become even stronger.
Communication, then, is not a decorative layer added after the research is finished. For work that seeks to enter public or professional discussion, communication affects whether it encounters the readers, critics, collaborators and later users for whom it matters.
But that does not mean scholars should simply maximise traffic.
There is a mismatch that is often ignored. The audience reached by maximum traffic is not necessarily the audience an article most needs to reach. A foundational argument in epistemology may matter most to relevant researchers, technology developers, educators and members of the public willing to think seriously about the issue. It does not need every possible person to click on it, but it does need as many of the right people as possible to have a realistic chance of encountering it.
The proper goal is therefore not undifferentiated exposure but effective expansion of the relevant audience.
Nor should this be reduced to the claim that “the public has a low level of understanding, so serious work cannot attract attention”. The same person may be highly sophisticated in one domain and need a clear entry point in another. Some people do not lack the capacity to understand; they lack the background. Others are willing to go deeper but do not know that a short piece sits above a fuller argument.
The task of communication is precisely to reduce those barriers to entry, not to compress every idea until only emotion and conclusion remain.
A 2024 systematic review and meta-analysis in the Journal of Medical Internet Research included 50 studies, nine of them randomised controlled trials. It supported the conclusion that social media can improve the dissemination of research, especially measures such as access, reading and downloads. But reach, engagement, citation and practical impact are different levels of effect and should not be treated as interchangeable. The review focused mainly on healthcare professionals, so its conclusions cannot simply be generalised to every field. Still, it highlights an important evaluative principle: whether dissemination occurred and what that dissemination actually changed are separate questions. Source: https://www.jmir.org/2024/1/e51418
For an expert, whether ten thousand views are more valuable than one thousand serious reads by highly relevant people depends on what the article is meant to achieve. And successful communication is not only about reaching those who already understand the subject. It can also help some readers who initially lack the background gradually acquire the conditions needed to understand it.
That requires a specialised capability.
Researchers may be skilled at formulating problems, building frameworks and organising evidence, but that does not mean they are equally skilled at audience analysis, media selection, layered explanation and distribution strategy. The growing importance of communication does not imply that every scholar should independently master every related skill.
In a 2025 Elsevier survey of more than 3,200 researchers, only 45 per cent said they had enough time to conduct research. That finding is a useful warning: taking communication more seriously cannot simply mean adding an unsupported new obligation to already overloaded researchers. Source: https://www.elsevier.com/about/press-releases/elseviers-global-survey-of-3-000-researchers-reveals-less-than-half-have
A more plausible direction is a new division of labour. Experts determine the intellectual structure, intended audience, limits of simplification and evidence that must be preserved. Editors and communication specialists handle translation and presentation. Technical staff support accessibility and digital maintenance. AI can assist with part of the process.
An expert does not have to personally operate every account. But the expert should participate in deciding how the work enters public space. Execution can be distributed. Control over how the underlying thought is represented cannot be abandoned completely.
Today this communication architecture also has to deal with a new kind of recipient: AI systems.
Calling AI an “audience” is primarily a functional description. An AI system may read research, select sources, reorganise arguments, generate answers and pass processed knowledge to other people. It can be both a user of knowledge and an intermediary in its further distribution.
But AI is not a single, homogeneous recipient.
Different systems can use different training pipelines, retrieval systems, generation methods and source-selection mechanisms. Even within one platform, separate features need not rely on identical models or processing. Google’s own documentation notes that AI Mode and AI Overviews can use different models and techniques, which helps explain why they may produce different answers and supporting links. Source: https://developers.google.com/search/docs/appearance/ai-features
User preferences can also enter these systems, but it would be a mistake to treat AI capability as nothing more than a projection of the average user. Research presented at ICLR 2024 found that human preference feedback can, under some conditions, reward answers that accommodate a user’s beliefs rather than answers that are more truthful. Feedback mechanisms therefore do influence model behaviour. But this does not mean that every user query directly enters training, or that a model can only reproduce the level of understanding already possessed by its users. Source: https://proceedings.iclr.cc/paper_files/paper/2024/hash/0105f7972202c1d4fb817da9f21a9663-Abstract-Conference.html
For knowledge distribution, the more immediate issue is what material a system can find at answer time, how it filters that material, whether it preserves relationships among concepts and whether it can trace claims back to their original sources.
The OpenScholar work published in Nature in February 2026 provides a useful example. The researchers built a retrieval corpus containing roughly 45 million open-access papers and combined retrieval, reranking and feedback to enable language models to synthesise scientific literature. The key point is not simply that the model became larger. What mattered was the knowledge system to which the model was connected and how it could retrieve and verify material from that system. Source: https://www.nature.com/articles/s41586-025-10072-4
Researchers therefore do not have to understand “making AI know my work” only as waiting for some future training run. In systems that can consult external sources, whether a research output is indexed, retrievable and clearly associated with the correct author and version can matter just as much.
As AI participates in distribution, the use of knowledge may also become increasingly separated from clicks on the original website.
In 2025, Pew Research Center analysed browsing data from 900 adults in the United States, covering roughly 69,000 distinct Google searches. When a search page contained an AI summary, users clicked a traditional search-result link 8 per cent of the time, compared with 15 per cent when no AI summary appeared. Links cited inside the AI summary were clicked only about 1 per cent of the time. This was not a randomised experiment and reflects a specific sample and time period, but it shows a pattern worth taking seriously: users may obtain an answer at the summary layer without ever visiting the source page. Source: https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
It follows that future measures of knowledge distribution cannot rely on website traffic alone. A contribution may enter AI-generated answers without producing corresponding visits for the author. Conversely, a site can attract traffic without ensuring that its ideas are understood or cited accurately.
Authors increasingly need to ask whether their ideas appear in relevant answers, whether necessary qualifications are preserved, whether claims are misattributed and whether a traceable route back to the original argument remains available. This is more complicated than counting page views, but it is much closer to the real effects of knowledge distribution.
For that reason, I believe experts and scholars need to build a “knowledge interface” for both people and AI.
The interface need not be a complicated software product, and it does not mean everyone needs to build an application. It is first of all a way of organising research outputs so that external readers and machine systems can tell who proposed a claim, where the full argument can be found, what evidence supports it, how it relates to existing research and what has changed in later revisions.
A researcher can maintain stable entry points that connect papers, books, concept pages and public essays. Core concepts should have relatively stable definitions. Important claims should point back to their original arguments. Different versions should be distinguishable. Specialist readers can go straight to the full material, while general readers can begin with explanatory essays or interviews.
The same body of thought can also exist at different depths. Books preserve full structures, essays address specific questions, podcasts expose the reasoning process and short-form material provides an entry point. The goal is not to repeat the same paragraph mechanically, but to connect these expressions so that ideas do not dissolve into thousands of untraceable fragments.
None of this should be sold as a mysterious set of “AI recommendation tricks”. Google explicitly states that appearing in its AI search features does not require special optimisation and that there is no special method that guarantees inclusion. Accessibility, clear written content, sensible internal linking and information that accurately reflects the page itself remain basic conditions. Source: https://developers.google.com/search/docs/appearance/ai-features
The real task is not to flatter the short-term preference of a particular model. It is to reduce avoidable loss as an idea moves into larger knowledge systems. What is made public and under what permissions should still be decided by authors and rights holders. Building an interface does not mean abandoning control over the work.
In Sustenesis, I describe a related development as the migration of knowledge systems. Writing, libraries, databases and the internet progressively externalised the storage and retrieval of knowledge. AI adds the external processing, recombination and operation of knowledge. The relevant question is therefore no longer only what an individual knows, but how the larger system preserves, retrieves, verifies and uses knowledge.
The need for experts to build their own knowledge interfaces can be understood as a practical extension of that claim. As knowledge increasingly operates through distributed systems, the way research enters those systems is no longer merely a matter of publicity. It affects whether the knowledge can participate in what happens next.
This is not to say that a few technical developments have conclusively proved the philosophical arguments in these books. The more measured claim is that the structural changes described there are becoming increasingly concrete. The position of the human knower, the mediating role of institutions and the ways research is called upon are beginning to shift together.
Traditional systems of academic evaluation also need to be reconsidered in this context.
I do not think journals or peer review should disappear. On the contrary, as the cost of producing content falls and the volume of material grows, verifying evidence, detecting error and distinguishing reliable conclusions become even more important. The real question is whether quality assessment, academic reputation and knowledge distribution must remain as tightly bundled together as they have been.
Traditional evaluation has structural limitations of its own. Research by Wang, Veugelers and Stephan in Research Policy found that highly novel work often faces delayed recognition and that its value may not be captured well by short citation windows or journal-level indicators. Such work can have large downstream effects but also carries greater uncertainty, which means evaluation systems that privilege near-term visible outcomes may systematically disadvantage it. Source: https://www.sciencedirect.com/science/article/pii/S0048733317301038
This does not show that rejected work is necessarily excellent. It does show that institutional recognition and research value do not correspond instantaneously or without bias.
If the scale and form of knowledge production change while evaluation continues to depend too heavily on inherited classifications, short-term metrics and fixed procedures, then in some circumstances a mechanism designed to help knowledge enter society may become a mechanism that delays or obstructs the circulation of new ideas.
That conclusion requires an important qualification. The issue is not that traditional institutions are destined to fail. The question is whether they can reorganise their functions.
One experiment has been to connect public dissemination more explicitly with open evaluation. In 2022, eLife announced that from 2023 it would publish manuscripts selected for peer review as Reviewed Preprints accompanied by editorial assessments and public review reports, rather than issuing a conventional accept-or-reject decision after review. The journal still screens which manuscripts enter review, so this is not the removal of every gate. But it demonstrates that dissemination and evaluation can be organised differently from the conventional journal sequence. Source: https://elifesciences.org/inside-elife/54d63486/elife-s-new-model-changing-the-way-you-share-your-research
That model will not suit every discipline and should not be treated as a settled solution. What it shows is that rigorous evaluation and timely distribution do not have to be tied to a single institutional sequence.
A better future system would make research discoverable while keeping its evidential status, review history, disagreements and revisions visible. Material that has not been sufficiently verified should not be presented as settled fact. But material that has not yet received traditional recognition need not be made invisible to discussion.
For experts themselves, this implies a different allocation of capability.
Research still needs depth, but communication cannot always wait until “there is time”. It needs stable time, appropriate collaboration and clear responsibility. How much effort this requires will differ across disciplines, career stages and goals. But for researchers who want their work to have lasting reach, communication has to occupy a real place in the research plan rather than being treated as costless peripheral labour.
Universities, research organisations and publishers also need to participate in this adjustment. A new knowledge interface cannot depend entirely on individuals paying for their own websites, editing media in their spare time and continuously operating accounts. If institutions reward production but do not support the accurate discovery, explanation and use of research, they have not fully responded to the new knowledge environment.
A scholar’s personal knowledge brand should not be understood as the manufacture of a recognisable persona. For a researcher, it should mean a traceable relationship: what this person has worked on over time, which concepts they have proposed, where the arguments are located, which claims have been revised and which questions remain open.
Such a public presence helps others find the work, but it also makes criticism, verification and further development easier. It does not replace academic evaluation with fame. It gives evaluation and exchange a clearer object.
I therefore continue to think that originality, foundational significance and discoverability will jointly shape the future position of experts and scholars. The first two concern what new possibility a contribution creates. The third concerns whether that possibility can actually enter the activity of a knowledge system.
Traffic cannot determine whether an idea is true. But an idea does not automatically gain the opportunity to be tested, adopted and developed simply because it is valuable.
Experts and scholars in the AI age do not all need to become professional communicators, but they increasingly need to participate in how their knowledge is distributed. This is true in relation to public audiences, and it is increasingly true in relation to AI systems.
This is not a demand that scholarship submit to traffic. It is a demand that genuine knowledge creators re-enter the process through which knowledge reaches society. The creation of new knowledge and the effective distribution of that knowledge can still be divided among different people in practice, but they can no longer be treated as unrelated tasks.
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.