Beyond Cherry-Picking: Scaling Historical Arguments

Why Historians Must Enter the Age of Big Claims

Chris Phillips is absolutely right: statistics help historians think about how unique any given example is. They push us against cherry-picking, against the temptation to elevate the exceptional case, and against a subtler form of presentism in which we search the archive for a shiny precursor that mirrors today’s mood or movement. Those habits all have their place. But if we take seriously the older ambition, historia magistra vitae, then history also aspires to say something about what is normal.

In that respect, historians are no longer alone. Political scientists like Erica Chenoweth have assembled longue durée datasets of nonviolent movements in order to generalize about their effectiveness over time. (Her conclusion: nonviolence wins more often than violence, though with important exceptions.) Peter Turchin, working with historians of the ancient world, has built the Seshat: Global History Databank, comparing technologies, wars, empires, and religions across millennia to produce arguments about the origins of cities, kingship, and even monotheism. Nassim Nicholas Taleb, whom I dined with this week, has compiled his own longue durée dataset of wars and casualties in order to test claims about whether violence is declining over time.

These are not my methods. They may not be yours either. As a historian trained in social and cultural approaches, I want to know about individual lives, about moods and propaganda, about the lived experience and interpretation of war, not merely the number of dead or the count of technologies. There is, in these datasets, often too little of the texture that historians are trained to value.

And yet I admire these projects deeply. I admire their scale. I admire their willingness to enter public debate on questions that matter right now. Taleb began counting war dead in order to argue with Steven Pinker’s The Better Angels of Our Nature. That conversation has drawn in historians as well, my coauthor David Armitage among them, though from a different angle, asking how categories like “civil war” themselves have been historically constructed and deployed rather than simply counted.

What unites Chenoweth, Turchin, and Taleb is their determination to generalize about war and peace, violence and nonviolence, questions that have always been central to historical inquiry. They offer clear, even binary answers to contested claims. Is nonviolence increasing? Chenoweth says yes, particularly in the modern period. Are wars becoming less stochastic? Taleb says no. If this is where our collective understanding of society is being formed, historians cannot afford to be absent.

What, then, makes historians different?

It is not a lack of rigor. It is not an allergy to numbers. It is our omnivorousness. Other disciplines pride themselves on agility, on the mastery of mathematics or philosophy. (I think, for instance, of Tyler Cowen celebrating the reach of economics.) But no discipline has the breadth of history at its best: statistics and mathematics alongside philosophy, the cultural turn’s engagement with art and meaning, social history’s grounding in linguistics and lived experience.

In my own work on the history of political economy through text mining, I have tried to build on questions from political economy and social history alike, using a sensitivity to memory encoded in language that depends on both historical linguistics and corpus linguistics. If the last wave of digital humanities was built on NLP, the LLM promises something more ambitious: the possibility of merging datasets like wages (as Louis Hyman has been documenting) with the arguments, experiences, and nuances embedded in historical texts.

The challenge is how to do that without losing sensitivity to individual lives, how to move from large-scale datasets to arguments about trends like the rise of nonviolence without flattening culture, decision, and imagination into mere counts.

Historians will not become economic historians overnight. But many are already motivated by questions, about capitalism, conflict, governance, that implicitly place them in dialogue with Pinker, Chenoweth, and Turchin. The question is whether we can engage those datasets without surrendering what we do best.

One answer lies in what I have called Critical Search (developed in The Dangerous Art of Text Mining). In practice, Critical Search treats datasets not as endpoints but as indices of change. Peaks, anomalies, and concentrations become invitations to investigate. The historian identifies a pattern, then zooms in, modeling individuals or events through text mining, and finally returning to close reading at the moment when everything begins to look different.

A concrete example: in my work on environmental rhetoric in Congress, I found that members of Congress routinely referred to environmentalists as “zealots,” “academicians,” and “radicals.” Counting these phrases allowed me to answer Phillips’ question: how widespread were these denunciations? The answer was: not very. Roughly 90% of such attacks were produced by just six members of Congress, and overwhelmingly by one figure, Ted Stevens.

At that point, the project pivoted. The data identified Stevens not as a cherry-picked case, but as a statistically grounded exemplar. The next step was to zoom in: to read his speeches, trace his career, and understand his role as a defender of oil pipelines and a central voice opposing environmentalism from the 1970s through the 1990s and beyond. The result was a history that could sustain a general claim, environmentalism faced sustained attack in Congress, while grounding that claim in the detailed study of a particular actor and his evolving rhetoric.

This is how political and social history scale. Not by abandoning the case study, but by selecting it rigorously.

LLMs now make it possible to extend this method far beyond parliamentary debates. They allow historians to draw on longue durée datasets, wars, casualties, nonviolent movements, wages, and to model change over time in ways that engage directly with the Pinkers, Turchins, and Talebs of the world. But crucially, they also allow us to use those datasets to identify moments of exception: the unusually violent, the unexpectedly peaceful, the early adopters of nonviolence.

Those moments, in turn, become the basis for historical explanation. They let us test the limits of quantitative claims by showing what is hidden until we zoom in. They offer a way to tell large-scale stories without surrendering detail. And they provide a more rigorous alternative to cherry-picking, not by abandoning selection, but by disciplining it.

That, I think, is the historian’s reply to statistics: not resistance, but integration on our own terms.

---