Cumulative Knowledge

In the AI Story

Eisenstein's central claim was that the conventional attribution of the Scientific Revolution to individual genius — Copernicus, Galileo, Newton — missed the structural conditions without which those individuals' work could not have become science. A scientific method required standardized observations, cumulative building, systematic comparison across distances and generations. None of these practices existed before print, because none of them was possible. The press did not produce the scientific method; it created the conditions under which the method could emerge.

The mechanisms by which print enabled cumulation were specific. Fixity meant that two scholars in different cities could be certain they were discussing the same text. Dissemination meant that a new work reached the scholarly community quickly enough for criticism and extension to occur within the author's lifetime. Preservation meant that the foundations on which any particular work was built would survive long enough for the next generation to continue the inquiry. Each mechanism contributed something distinct; their combination produced the cumulative capacity that characterizes modern knowledge.

The AI transition presents an ambiguous relationship to cumulative knowledge. On one hand, AI systems represent an extraordinary compression of accumulated human knowledge — a large language model has absorbed vastly more text than any human scholar could read in a lifetime. In this sense, AI is the culmination of print's cumulative project: an instrument that can synthesize across the entire written corpus in ways no previous tool could approach.

On the other hand, the properties of AI-mediated knowledge — opaque provenance, generative variability, centralized training — compromise the conditions that make cumulative building possible. If an AI's assertions cannot be traced to sources, the next generation cannot evaluate them. If the model's outputs vary across invocations, no stable text exists for scholarly engagement. If the training corpus is controlled by corporations whose decisions are opaque, the knowledge base can shift without notice, undermining work built on earlier versions. The AI transition may turn out to enable new forms of cumulative inquiry while simultaneously undermining the conditions on which print-era cumulation depended.

Origin

The concept of cumulative knowledge as a structural achievement rather than a natural property was developed by Eisenstein in response to what she saw as the naïve triumphalism of traditional histories of science. Earlier work — notably George Sarton's monumental history of science and the logical-positivist tradition of Hans Reichenbach and Carl Hempel — had treated cumulative progress as the natural trajectory of rational inquiry. Eisenstein showed that this 'natural' trajectory was historically specific and technologically conditioned.

The concept has been extended by historians of science including Steven Shapin and Simon Schaffer, whose Leviathan and the Air-Pump (1985) examined the specific practices through which experimental knowledge became stable and transmissible. Bruno Latour's work on 'immutable mobiles' develops a closely related concept from actor-network theory.

Key Ideas

Not natural, but structural. Cumulative knowledge is not the automatic trajectory of rational inquiry; it requires specific technological and institutional conditions.

Print created the conditions. Fixity, dissemination, and preservation combined to make cumulative building possible for the first time at scale.

Traceability is essential. Cumulative inquiry requires the ability to evaluate, correct, and extend prior work — which requires access to the evidence on which prior work was based.

AI's ambiguous relationship. AI synthesizes across the entire written corpus (the culmination of print's project) while simultaneously compromising the traceability that cumulative building requires.

Institutional work remains. Whether AI supports or undermines cumulative knowledge depends on institutional choices that have not yet been made: about provenance, transparency, and the distribution of control over training data.

Debates & Critiques

The question of whether AI extends or undermines cumulative knowledge is fundamentally about institutional choice, not technological property. The same technology can support cumulation — if it preserves provenance, enables verification, and distributes control — or undermine it, if it does not. The historical record from print suggests that the institutions that made cumulation work took centuries to develop and were never complete. The AI transition raises the question of whether compressed timescales will produce compressed institutional development, or whether the institutions will lag so far behind the technology that cumulation breaks down.