Most professionals I know—whatever their field—appear to me today to be in a kind of collective flinch, as if anticipating a painful blow. Even as increasing numbers are experimenting with the use of artificial intelligence in their work, some with genuine enthusiasm, there is a premonition that the real changes to come, including in their own work lives, will be systemic and seismic, and remain unpredictable and undefined. While levels of optimism for the ultimate impacts on a given field and its practitioners may differ among journalists, lawyers, professors, public servants, or financiers, the expectation that a disruptive tipping point is coming seems relatively consistent.

My colleagues and I at the Carnegie Endowment for International Peace are party to the ripples of reflection reverberating through the professional class. Several among us are scholars examining the unfurling impacts of AI on international politics, trade, regulatory regimes, political economy, security, and economic development. All of us are implicated in thinking about how AI will change the nature of the work we do and how we do it. We are learning, experimenting, and feeling our way through what is likely—at most—the end of the beginning of the coming transformations.

Three years ago, I often described the public discourse around AI as “frothy.” It was a feeding frenzy of punditry, with a surplus of confident (sometimes contradictory) predictions and a deficit of apparent experts. Now, the discourse is maturing. Predictions these days feel both more humble and better substantiated and contextualized. Debates between analysts are more clearly rooted in an identifiable epistemic community and have accordingly become less polemical and more nuanced. In my own interactions with Chinese experts at several track 2 events over the past twenty-four months, I have noticed that what once felt like a bifurcated conversation increasingly has some areas of overlap, albeit populated by different perspectives on fundamental questions.

I have tried to learn and think about what AI will mean for the community of nearly 300 researchers around the world that I am fortunate to be a part of and steward. As I endeavor to look into the future of our field, I am not confident in my vision. My hypotheses are weakly held and piecemeal; my expectation for a future state is not comprehensive. And I think I very likely underestimate how changes elsewhere in politics and society will have second- and third-order effects that impact how we do what we do. But I believe that having some sort of vision, even as a placeholder, a “vision subject to revision,” is useful to pull out of the flinch and to increase one’s odds of more successfully adapting. Waiting for AI to happen to us can’t be the most effective way to maintain our emotional well-being, and (for that reason if no other) it probably isn’t the best way to support our impact as scholars and civic contributors either.

Here are five—what should I call them? Not predictions—“emergent intuitions” about how AI may change Carnegie and other research institutions. These are meant to start conversations, not to finish them. Some are more provocative than others, and none are meant to be confidently prescriptive.

1) We might produce less, not more.

My colleagues and I have already seen early signs that external funders of our work will increasingly expect us to demonstrate productivity gains by incorporating AI into our research process. If it used to take six months of researcher time to produce a report, the advent of AI to help with finding and summarizing sources, outlining ideas, generating graphics, and so on should mean that the same report takes much less time, the thinking goes.

My intuition is that this expectation is incorrect, and that in the coming years, we may produce fewer publications, not more (and still be doing our job right).

If the internet democratized the distribution of content, AI has democratized the creation of content. The combination of these two phenomena has changed the backdrop against which we work. Think tanks were created in a time when research content on specialized topics was scarce. Our model is premised on there being unmet demand (maybe not from a large share of the general public, but from a significant number of people who need it for their work or learning) for a wide range of research content.

There was a time when Carnegie could produce a report on a policy question—say, the proliferation of nuclear weapons—and expect that the New York Times or Washington Post might cover its release. We printed copies of these reports and dropped them off at Senate office buildings or mailed them to the Pentagon knowing that policymakers would eagerly consume them. High-quality written overviews of complex and consequential policy problems were scarce; when offered, they met a persistent demand. Even in the first decade of this century, if we put a primer on a policy topic—say, the enduring challenges to peace in the Balkans—on our website, students, journalists, and diplomats from around the world might have happened upon it.

For decades, Carnegie’s name has been a mark of quality that meant: This information and analysis is unbiased, rigorous, and fact-based. In the future, our mark of quality might also need to reliably say: This information and analysis is new, novel, and different from what other sources, including LLMs, provide. Of course, we already provide innovative and new insights in much of our work. But historically, much of our written product has been synthetic of other work, and has functioned as explainers of particular issues. These products had value—when it was hard to find an explanation of the ongoing challenges to peace in the Balkans, an explainer of some of the key political developments and their historical roots was a real contribution to understanding that could lead to better policy choices. These kinds of explainers and overviews are not likely to be how we distinguish ourselves as researchers going forward—or, rather, they are the kind of research outputs that can or will most obviously be accelerated by and in some cases replaced by AI.

Our future research must take account of the shift from a world with a research content shortage to a world with a research (or ersatz research) content surplus, and address the distinct needs of that world. My hypothesis is that we should focus more on publishing written research and analysis that directly challenges prevailing assessments or conventional wisdom, leverages primary research and fieldwork, and crystallizes or reframes large amounts of information into heuristics or frameworks that orient human readers within complex problem sets.

If one imagines a pyramid of the research content of Carnegie in the past, at the top there might be a smaller slice of “genuinely novel” written research. Underneath might be a broader slice of research that brings together information from disparate sources to organize it and simplify it for policy audiences, roughly “synthesizing research.” And under that, an even broader slice of expert commentaries and short pieces that helped fill understanding gaps in more niche areas, “expository research.” In the coming years, I expect the bottom two layers of the pyramid will significantly decrease in size, and the top slice will become the primary focus of research effort. We may publish fewer, but more genuinely novel pieces of research.

Importantly, this focus on “genuinely novel” is not only or even primarily about how to preserve a Carnegie comparative advantage or distinctive product in a disrupted content market. Rather, it’s that the other types of research product have a less obvious route to impact in a changed world. Genuinely novel research will be distinctly helpful to policymakers trying to make better choices in a world of research content surplus. The consumers of our content used to be thirsty for knowledge; now they are swimming in it. As I’ll explain below, I think we will increasingly have to become sherpas for our research products, creating derivative products from them and shepherding them to various audiences, in order for them to break through and make impact in a content-rich world.

2) As researchers, our core skills will be curation, refinement, and creative improvement of ideas, often in collaboration with others.

Of course, these skills are already core to a good scholar, but they are likely to become an even greater advantage. Being a clever, astute writer has in the past often been enough—enough even to sustain a whole career. The scholars who got by because they were good writers rather than truly innovative thinkers weren’t frauds—ideas needed human artisans to translate them into prose—just as, before photography, a human face with all its subtleties of expression needed a painter to translate it into a portrait. The artistic photographer distinguishes herself by the ability not just to render a likeness, but rather to use perspective, light, a focus on a particular detail, even a distortion or manipulation, to generate an image that compels an audience to see differently.

Just as I expect a shift in the focus or type of Carnegie’s research output, there will be a correlative shift in the allocation of research effort by Carnegie scholars, and in the skill set that distinguishes exceptional scholars. Carnegie’s scholars will need to spend less of their time gathering and organizing facts—AI will be ever more capable at this—and more time thinking deeply about the dynamics at play in the policy problems these facts describe.

When AI can summarize the history and intricacies of a particular problem, like how NATO protects European security, the human scholar (who used to spend time laying out historical context and approach) will need to focus more on what reimagining European security might look like. It’s not that our articles have to become kooky creative provocations attacking the bounds of the Overton window, but that our writing will need to be more focused on refining or adding nuance to existing analysis (and explaining why that nuance might shape different choices) or lifting up certain insights among many. As AI becomes ever more capable of (and ever more depended on for) summaries of conventional wisdom, brilliant humans may be able to focus more of their time on pursuing well-founded challenges and updates to that conventional wisdom and Kuhnian paradigm shifts.

A corollary hypothesis to my first intuition in the section above is that we may spend more time in the old-fashioned writing process than we do now and produce fewer published words: It takes longer to work through thorny intellectual problems and to generate genuinely novel conclusions than it does to hammer out a summary of what we know. The kind of writing that we do has always had two products. Of course, writing produces reports and other research outputs for audiences to consume; but the process of writing also produces the expert. Indeed, for most of us, the act of writing is the most important ongoing training we have for wrestling with ideas. It is the way we understand and explain the world to ourselves. Writing is at least how we capture, and often how we come up with, something new. The idea of being an expert on a particular topic without having written about it would feel impossible to most of us. That experiential truth will remain.

In addition to the individual pursuit of fresh and rigorous thinking, we will find new appreciation for irreplicable value in the dialectic creation of knowledge. While many scholars are already learning to use AI effectively as a sounding board and thought partner, I suspect that these exchanges will not replace the value of human-to-human exchange, and that the latter will actually be even more essential to the research process in the years ahead. Convenings, including conferences, seminars, university lectures, and the like, have long been ways of disseminating and exchanging information. The very best of these are organized and curated in ways that facilitate the exchange of ideas around a set of problems in a way that contributes to new ideas and approaches. A talent for designing convenings, especially of smaller discursive groups in workshop rather than lecture or speakers’ panel formats, and for facilitating these gatherings, will become ever more valuable and distinctively additive as the dissemination and packaging of information becomes more automated. Being able to attract high-quality participants with a diversity of views will distinguish institutions and individual researchers as leaders of epistemic communities. Refining and improving existing ideas, thereby unlocking consequential subtleties, can be a productive and fulfilling team effort in parallel with individual scholarship. And I suspect we will see convening as progressively more essential to knowledge creation, rather than to dissemination. Research organizations that don’t already have a focused convening strategy in their knowledge production approach will find themselves drawn to developing one.

3) In our post-expert world, experts may yet come back into favor.

Carnegie exists to compensate for a market failure: In order for those in positions of power and authority to make choices that promote peace and human flourishing, they need information that the market (or the bureaucracy) alone will not provide. We will continue to do original research to produce insights that help shape policy in constructive ways. At the same time, we will find our greatest impact derives more from our ability to help key audiences sort through an abundance than from contributing further to it. Where Carnegie’s research staff were once almost entirely information producers, we will increasingly function as information guides helping audiences navigate information surplus. To benefit the world, we will have to mediate the intersection between information and structures of power—in politics, in media, in the private sector, and in culture more broadly. While the digital age has made it much easier for anyone to acquire a baseline of knowledge in a given topic, information overload will create a new premium on genuine expertise that can help policymakers and others who use specialized knowledge for decisionmaking.

In its first two decades, the internet exploded access to information. It also enhanced the productivity of knowledge workers, and their ability to distribute their content, but content itself still required the cognitive abilities exclusive to human labor to produce. In its first years, commercially available AI has exploded the production of content for humans to consume. There has been a persistent public scrutiny of AI hallucinations, where LLMs generate bibliographic records of books, data, or historical events that aren’t real. The models will get better; hallucinations may not be eliminated in the near term, but they will become more subtle. In some ways, the focus on hallucinations, while reasonable and important, has been a mechanism for collective self-soothing about the enduring supremacy of human-produced content. It may have caused us to focus rather less on the reality that AI is actually extremely good at producing moderately sophisticated primers on a wide range of topics—it has solved the inefficiency of gathering collective observations and insights in everything that’s been written on a particular topic, and it has done that to good effect for a majority of use cases.

When I was growing up, we had a thirty-volume encyclopedia in our house. (In 1986, my parents invested serious money into buying one—a feature of a middle-class aspirational family.) For grade school and middle school homework, the encyclopedia was very useful. But even by the end of high school, it was hard to write a good essay relying only on the encyclopedia. It could provide a start, but its pretensions to completeness were part of what made the writing flat and the nuance thin. AI means that the world has access to literally infinite encyclopedia articles.

However, people will still need assistance to confidently navigate nuance, and to identify what’s missing, what’s oversimplified, or what are the unanswered questions.

The ability for various actors to generate authoritative and (seemingly) data-driven content that supports a particular set of empirical conclusions or policy choices will complicate the task of those in positions of power attempting to sort through information. Senate staffers and analysts at major investment banks don’t just read industry publications and name-brand newspapers—they are affected by the videos that they see scrolling social media or the conversations they have with family members who saw a revealing graph on X. We are all already swimming in AI slop. Some may be happy to swim, but others will increasingly demand a life raft, especially those whose decisions have economic and political consequence. They will want to know that the content generated by a model passes muster with a human expert, and they will want to know that it hasn’t been distorted by the national, commercial, or political biases of the models’ owners or creators. In this context, the personal and institutional reputation—for quality and independence—of experts will matter as much as ever: When lost at sea, you want not just a life raft but a competent and trusted navigator.

Audiences (and experts themselves) will increasingly learn that we are in an era when facts do not, in fact, speak for themselves. Facts need experts to speak for them, and to elevate rigorous ideas from the rest. Every piece of research we do may require a suite of derivative products that essentially function as marketing materials for the core insights. A significant paper or report will almost always need an op-ed version to advocate for the ideas it contains in a public square awash in content. And while op-ed–style pieces are likely to remain a mechanism for reaching policy elites, they are less and less effective at shaping the political context in which policy elites deliberate (and elites themselves consume a broader range of policy-relevant information than they once did). Op-eds for the non–op-ed audience, whether short-form video or expert appearances in media, will be increasingly important in moving the Overton window of policy debates and credentialing experts in the public eye. One can be an expert without being recognized as such, but one cannot be an effective shaper of the contemporary public square.

Mediating the intersection between information and structures of power demands an additional skill set of experts—they need to not only create and articulate the knowledge that is relevant to public decisionmaking, but identify the interlocutors who can use that expertise, reach them, and interact and communicate with them effectively. In a post-AI world, it’s not enough to put knowledge “out there”; one must act as a sherpa helping it find its intended audience. My hunch is that as a consequence of this reality, we will have to reckon with the fact that the profile of successful experts will feel more like what we might call “charismatic intellectuals,” rather than the (always a bit of a caricature) bookish solitary scholar. As AI-generated content proliferates and the tone and tenor of that content regresses to a synthetic mean, the aesthetic attraction of a beautiful human mind at work and its distinctive way of communicating ideas may compete favorably for attention. The best of these charismatic intellectuals are and will be highly impactful by connecting people with good ideas (the worst will always be charlatans). We haven’t historically prioritized this skill set when hiring—if the profile of compelling human experts evolves, we may have to learn how to attract and develop that kind of talent in a way that doesn’t forsake the essential foundation of intellectual strength.

4) We have focused too much on how AI might replicate the things we can do, rather than how it might help us do the things that we can’t . . . or couldn’t.

Most concretely, this intuition is that, as researchers, most of us think more about how AI might affect—even contaminate—the process of writing than we do about the way that it might unlock the potential to do much more research that is data-driven or how we might use data in new ways to support our more theoretical research. The bias to be more concerned about AI and writing reflects our typical background—most of us were trained as social scientists, political scientists, and political theorists. For many of us, empirical research has been based on some kind of field work—we are anthropologists of the policymaking and policy-shaping classes in so many places. When we have produced reports based on data, we have often depended on other economists, pollsters, or statistics-keeping actors to provide the data and sometimes the quantitative analyses from which we have drawn our research hypotheses and buttressed our conclusions. AI tools today have tremendously expanded the ability of a single researcher to collect, assemble, manipulate, and analyze data. The progress of the technology, and, equally importantly, our facility with using it, will make it possible for theorists and political scientists to leverage more robust data analytics to derive empirical insights and support their research claims. We have spent increasing amounts of time experimenting with how we can prompt the models to do some of what we do; we should spend more time thinking about how the existence of the models might prompt us to do things we don’t.

There are at least three ways AI empowers policy researchers to do work that we could not do five years ago. First, AI has democratized the “beyond Excel” frontier, making it possible to do the kind of data analysis that was until recently the province of advanced computer programmers or those with large teams of data analytics experts. We can test research hypotheses, look for patterns, and derive insights from more data. Second, we can bring data sets together in new ways and assemble new data sets of our own. Using AI to scrape data from the internet or from digital scans of analog sources unlocks data sets that were theoretically possible before (a historical archive of the statements of all EU countries’ foreign ministries from 1998 to 2024, coded and searchable by any one of 200 key terms, translated into English) but economically unfeasible for a policy research organization. And it allows us to marry two or more existing data sets (for example, organizing World Bank macroeconomic data, Intergovernmental Panel on Climate Change data, and Transparency International corruption data by country for the fifty largest economies in the world). Third, the computing power and vibe coding capabilities that we now possess allow us to experiment with new forms of data analysis. Economists and medical scientists have used Monte Carlo analyses—that simulate a scenario thousands or millions of times based on multiple variables—to arrive at probabilistic assessments of likely outcomes of phenomena that are complicated to model. Policy researchers might increasingly use such analyses to complement more qualitative observations drawn from, say, war game simulations of escalation toward nuclear conflict.

Of course, other researchers will have access to these expanded data analytics capabilities too. Field work—good old-fashioned human intelligence-gathering—is likely to remain a comparative advantage of excellent researchers. Curating closed-door spaces and relationships of trust so that people share reflections and insights in person that they wouldn’t share in written communication will remain essential. But the ability to analyze large amounts of data fundamentally changes the nature of policy research by making it possible to marshal empirical validation for the kind of research hypotheses that we once had to interrogate with only anecdotal or qualitative evidence. We can answer questions with moderate confidence that we couldn’t have answered before, and reject hunches more quickly in which we might have previously invested precious time in vain.

It’s exciting to think about the possibilities—but it’s also daunting. A vast expansion of data-driven and quantitative research is not a small change in our work! It is not a matter of adding a few extra graphs to a research paper—it’s a methodological shift that would affect all stages of the research process, from the framing of research questions to analysis to editing and fact-checking before publication. In order to responsibly take advantage of the opportunity, we would have to be willing to consider significant training and upskilling of those who produce our research products.

5) We will wrestle with the question of how much of our knowledge we should make public.

As with most technological changes, we tend toward a non-malicious narcissism in our consideration of their likely impact. We think more about how they affect us than about how they affect others. In addition to a tendency to be self-centered, we also reasonably focus on the effects we can see in our own lives because of the daunting nature of the sheer complexity of trying to account for so many second- and third-order effects on the wider world. And yet, if we are to be effective in navigating technological change, we need to spend some time thinking through not only how AI changes what we do and how we do it, but also how it will change the behaviors of other actors.

Like some other nonprofit research organizations, Carnegie has long maintained a default approach of making our research publicly available for free. For both mission-driven reasons and existential considerations relating to what, for lack of a better term, I’ll call our “business model,” I wonder if we will consider modifying that approach and keeping more of our research product private. I realize that this may be among the most provocative of the intuitions shared here. I am certainly not recommending that we should change our approach, only observing that there may be pressures that cause us to consider its impact against the background of broader changes in the information ecosystem.

In terms of our mission—to inform policymakers and the general public so that public decisionmaking more consistently tends toward peace—the curation role that is increasingly essential to our research is something we cannot control if our research is mainly disseminated in regurgitated form through LLMs.

To be sure, the models are already effectively distributing some of the ideas in our research farther and wider, deepening the impact of our work by reaching beyond a relatively small and elite audience but also by making it part of the basis for societal understanding of key historical events and socio-political phenomena. (I admit to a small shot of pride each time I ask a model to create a briefing note before a TV appearance and see a Carnegie article in the citations.) If the models become primary widespread sources of human understanding, our research transmitted through them could shape epistemic communities and general understanding.

But on the other hand, like a quote chopped up by a journalist so that it includes the caveat and not the statement—or vice versa—our ability to curate, contextualize, and appropriately nuance the content that we put into the world is diminished by the digestive processes of these models. If our value-add is providing deep research and nuanced insights, a repackaged executive summary that isn’t traceable to the foundational research is not achieving our objectives. And if the future is one in which our scholars’ impact will be derived not only from doing the research but from seeking and being invited to engage interactively with those in positions of power, we will miss out on the credentialing effect of research products that stand on their own with a byline. A Claude query is far less likely to lead to a congressional hearing or a call from the foreign ministry than a well-written and nuanced report or op-ed.

In some sense, we will need to decide whether to make a bet on ourselves; specifically, we have to be confident that even in an era of exceptionally powerful models that are very good at providing a baseline of knowledge about particular policy challenges, we can continue—through a mix of research, dialogue, and fieldwork—to produce better and fresher insights than the models. If we make that bet, we might choose to withhold some of our work product so that we can choose to deliver it outside the models. We might say: You can’t put our intellectual caviar in your AI soup, but you can come to us and eat it on its own.  We know it’s better, and you should recognize that too. Is this hubris? Perhaps. Time will tell. But I think the question of whether to make that bet is an essential one in the years ahead.

And this choice will be an uncomfortable one, not only because it requires us to pick a side in the “man against the machine” question of our time (though really it is not an assertion that the machines aren’t superior at some kinds of intelligence, only that as human experts we will continue to have distinctive value by encompassing a broader set of skills), but also because it will change something essential about how we understand ourselves. The public nature of our work, offered for free to the marketplace of ideas, has not been just an impact strategy for us; it has been core to who we are. If we were to shift in some meaningful ways to being something more like a private intelligence organization that is dedicated to the public good, we would have to adjust both what we do and how we see ourselves.

Whatever our view on the wisdom of such a shift in terms of how we pursue our mission, there may be pressures on our funding model that push us toward it. Our philanthropic funders support the independence of our research, but they reasonably want to know how their contributions are making a difference. If our research becomes mostly grist for the model mills, even if that produces better general understanding, it will be hard to point to the ways in which it does so concretely. Driving direct consultation with our experts and engagement with bespoke research products might become a more important part of how we demonstrate our ongoing impact on public decisionmaking.

If we do evolve into a collection of scholars who have unique and non-public insights that can help decisionmakers filter and clarify the era of content surplus, AI offers new opportunities to develop a cache of proprietary knowledge. Those data—more than the packaging and distribution of them—have the potential to be a comparative advantage. Research-informed syntheses of existing data sets—like recipes that combine multiple ingredients to produce a new taste—can be used to derive unique, empirically backed insights. We have not historically distinguished ourselves in the ideas space by developing a library of proprietary data sets, but AI makes it cost-efficient to develop data both to inform our thinking and research and to draw relevant audiences interested in those data.

AI also expands the possibilities for internal knowledge management in an organization like ours. I would be very interested in having easy access to the takeaways from conversations that colleagues of mine have had with German defense manufacturers or Ukrainian frontline soldiers, or the discussions at a closed-door convening on critical minerals. Many of the people we talk to offer private insights that need to be treated with confidentiality and discretion. The content collected in our research process is inefficiently used—usually only by the researcher or researchers who collected it—but it could inform the work of other colleagues, and AI could enhance our ability to responsibly and efficiently share across the organization.

Conclusion

While I feel relatively more confident about the broader transformation of research experts from a primary focus on being information producers and distributors to being (at least as importantly) skilled sherpas guiding targeted audiences through information surplus, the modalities and specifics of that new world are less clear. This essay is bound to be unsatisfying, especially to those who are persuaded by some portion of the thoughts above, because it leaves unanswered the question “what should we do about it?”

I don’t know. “For now we see through a glass, darkly,” indeed. As I’ve hinted, I suspect that we will have to adjust and adapt our approach to everything from fundraising to communications and distribution of research to research design and execution itself. And it seems likely that AI will significantly change our approach to recruiting, training, and managing human intelligence, too. Part of what we can do now is work to be intellectually and emotionally open to the likelihood of significant shifts in how and what we do. We cannot accept such shifts uncritically, but we need to ensure that we are not blind to the ones that, however uncomfortable, have merit. And we will need to think ahead about the resources and talent that will be required as we adapt.

Finally, a conviction, rather than an intuition: One of the attractions of the organization that I work for, the oldest foreign policy and international affairs think tank, is that it has its—beautifully idealistic—mission right there in its name. We are the Carnegie Endowment for International Peace. Our commitment is to the mission, not the model.

If history can be any guide, and the coming transformation is as significant as other dramatic changes in technology and political economy, the challenges to peace are likely to increase. As AI, together with other technological innovations, transforms our world and the relations between polities and people, it should not transform our values. An ethical commitment to pursuing peace is rooted in a moral commitment to the dignity of human beings.

Increasing the likelihood that those in positions of power and authority make intelligent and informed choices that tend toward peace will remain a worthy objective. How we contribute to it will change. There will inevitably be parochial temptations to preserve almost everything about our current way of doing things. We may welcome technology’s ability to solve what we currently assess to be nuisances, but we will be more resistant to changing the parts of our work that we currently find comfortable and familiar. We need to be clever conservatives or selective revolutionaries—preserving what we ought and changing what we must. Old values can help us solve new problems, and if we need a guide for how to adapt, what to preserve, and what to innovate in our mode of interaction with the world, we should start not with our vocation but with our mission.

Editorial note: This article was written without any use of AI.

Acknowledgments

I would like to thank my Carnegie colleagues for many conversations in recent months that have helped shape my thinking on this topic. Special thanks to Jon Bateman, Sophia Besch, Tom Carothers, Thom Crockett, Tino Cuéllar, Sarah Labowitz, Alison Markovitz, Arthur Nelson, and Katelynn Vogt.