The Semantic Fault Lines of Science: Linguistic Instability and the Crisis of Cumulative Knowledge
The Word That Means Everything and Nothing
Consider the term 'model.' In theoretical physics, a model is a mathematical structure that represents, with varying degrees of idealization, the behavior of a physical system. In molecular biology, a model organism—the laboratory mouse, the fruit fly, C. elegans—is a living system chosen for its experimental tractability, standing in for biological processes presumed to be universal. In machine learning, a model is a parameterized computational function trained on data to produce outputs that approximate some target distribution. In epidemiology, a model is a system of differential equations projecting the spread of disease through a population.
These are not merely different applications of a shared concept. They are, in philosophically significant respects, different concepts that happen to share a label. The physicist's model is evaluated primarily for its predictive accuracy and mathematical elegance. The biologist's model organism is evaluated for its genetic tractability and physiological homology to the system of interest. The machine learning model is evaluated for its performance on held-out test data, a criterion that is itself contested and context-dependent. The epidemiological model is evaluated for the plausibility of its assumptions and the robustness of its projections to parameter uncertainty.
When researchers across these fields read one another's work, attend interdisciplinary conferences, or collaborate on grant applications, they share a vocabulary without sharing the conceptual architecture that gives that vocabulary its meaning. The consequences for the coherence of scientific communication are serious, and they have received far less attention than they deserve.
Semantic Drift and the History of Scientific Terminology
The phenomenon of semantic drift in scientific terminology is not new. Historians of science have documented numerous cases in which terms migrated across disciplinary contexts, acquiring new meanings in the process while retaining the authority associated with their original usage. The concept of 'information,' to cite one of the most consequential examples, was given precise mathematical definition by Claude Shannon in his 1948 paper on communication theory. Shannon himself was explicit that his technical definition was not intended to capture the ordinary meaning of information as meaningful content—it was a measure of uncertainty in a probabilistic system, deliberately stripped of semantic significance.
Yet within a decade, 'information' in Shannon's technical sense had migrated into molecular biology, cognitive science, and eventually neuroscience and artificial intelligence, carrying with it both the mathematical precision of its origin and the commonsense connotations of meaning and significance that Shannon had explicitly excluded. The result was a term that functioned simultaneously as a rigorous technical concept and a powerful rhetorical device, capable of lending scientific authority to claims that its formal definition did not support.
This pattern—precise technical definition followed by disciplinary migration, semantic enrichment, and eventual ambiguity—characterizes the life cycle of many of the most productive terms in scientific discourse. 'Fitness' in evolutionary biology, 'energy' in thermodynamics and popular culture, 'significance' in statistical inference: each illustrates how the very productivity of a scientific concept creates conditions for its eventual semantic instability.
The Replication Crisis as a Semiotic Problem
The replication crisis, as it has been discussed in the American scientific community since roughly 2011, is typically framed as a methodological problem: underpowered studies, selective reporting, questionable research practices, and publication bias have together produced a scientific literature in which many published findings cannot be reproduced. These methodological factors are real and important. But they do not exhaust the sources of the crisis.
A growing number of philosophers of science and metascientists have begun to argue that a significant portion of replication failures are not methodological failures at all—they are semantic failures. When a replication study fails to reproduce an original finding, the standard assumption is that one of the two studies was conducted improperly. But a third possibility exists: the two studies were, in fact, investigating different things, because the key terms in the original study were operationalized differently in the replication, reflecting genuine conceptual ambiguity rather than methodological error.
This possibility is particularly acute in psychology, where replication failures have been most extensively documented. Constructs like 'attention,' 'working memory,' 'self-control,' and 'priming' are operationalized through a wide variety of experimental paradigms, and there is ongoing disciplinary disagreement about whether different paradigms measure the same underlying construct or distinct ones. When a priming effect fails to replicate, the relevant question is not only whether the replication was conducted correctly but whether 'priming' in the replication refers to the same phenomenon as 'priming' in the original study. Answering that question requires conceptual analysis, not merely methodological audit.
Cross-Disciplinary Communication and the Tower of Babel Problem
The problem intensifies when scientific communication crosses disciplinary boundaries, as it increasingly does in an era of interdisciplinary research initiatives, multi-institutional collaborations, and translational science programs. The National Institutes of Health, the National Science Foundation, and major private funders have invested substantially in breaking down disciplinary silos, on the reasonable assumption that the most important scientific questions require expertise from multiple fields. What these investments have not addressed is the linguistic infrastructure required for cross-disciplinary communication to be genuinely cumulative rather than merely additive.
Consider the concept of 'network,' which has become a dominant organizing metaphor across disciplines as diverse as neuroscience, sociology, ecology, and computer science. In each field, 'network' carries specific formal properties derived from the mathematical apparatus of graph theory, but the interpretation of those properties—what it means for a network to be 'robust,' 'efficient,' or 'modular'—varies substantially across disciplinary contexts. Researchers who borrow network concepts from adjacent fields frequently import not only the mathematical tools but the interpretive assumptions embedded in those tools, assumptions that may not transfer validly to the new domain.
Toward a Semiotics of Scientific Discourse
Addressing this dimension of the replication crisis requires drawing on intellectual resources that the scientific community has been reluctant to engage: the philosophy of language, the history of scientific concepts, and the sociology of knowledge. The work of scholars like Ian Hacking on the historical ontology of scientific concepts, Hasok Chang on the operational definition of scientific terms, and the tradition of conceptual analysis in the philosophy of psychology offers tools for the kind of systematic examination that the semantic instability of scientific terminology demands.
What is needed, ultimately, is not a universal scientific language—the history of such projects, from the logical positivists' unified science program to more recent ontology-engineering initiatives in bioinformatics, suggests that the aspiration to eliminate semantic diversity from science is both practically unachievable and philosophically misguided. Different disciplines have different questions, and those questions generate different conceptual needs. What is needed instead is a more systematic practice of conceptual cartography: explicit, discipline-aware documentation of how key terms are defined, operationalized, and interpreted across the contexts in which they are used.
Without such a practice, the replication crisis will continue to be treated as a problem of scientific conduct when it is, in significant part, a problem of scientific language—a problem that no amount of pre-registration, open data, or power analysis alone can resolve.
