Context-dependent knowledge is the form of knowing whose content cannot be abstracted from the particular situation in which it operates without catastrophic loss of meaning. Where episteme aspires to regularities that hold across contexts, context-dependent knowledge specifies what works in this situation, with these stakes, for these people, under these constraints. Flyvbjerg's career-long argument is that the most consequential forms of human knowledge — practical wisdom, expert judgment, institutional understanding — are constitutively context-dependent, and that the dominant intellectual traditions of modernity have systematically undervalued this form of knowing in favor of the context-independent aspiration inherited from the natural sciences.
The philosophical pedigree runs through Aristotle's distinction between phronesis and episteme, through Wittgenstein's analysis of rule-following, through Michael Polanyi's tacit knowledge, through Hubert Dreyfus's critique of rule-based AI. The common thread is the recognition that the knowledge that guides skilled action cannot be fully specified in propositions and therefore cannot be transmitted without loss by any system that operates purely through proposition-manipulation.
The distinction has operational consequences for AI. Large language models are extraordinarily effective at context-independent knowledge: they can retrieve, synthesize, and apply propositional content at speeds no human matches. They are structurally incapable of context-dependent knowledge in the full sense, because they possess no situation, no stakes, no embodied history, and no relationships within which context would acquire meaning. They can produce outputs that mimic context-dependent judgment by statistical extrapolation from training data — but the mimicry and the capacity are categorically different.
Segal's account of the Trivandrum transformation illustrates the principle. The twenty-fold productivity gain was not a universal finding but a context-dependent achievement — produced by the specific configuration of experienced practitioners, established trust, physical co-presence, and calibrated leadership. The general claim AI increases productivity survives the elimination of context. The phronetic finding — AI amplifies the judgment of experienced practitioners within high-trust teams under specific conditions — does not. The particularity is the value, because it tells the organizational leader what conditions must be created for the deployment to succeed.
The methodological consequence for research is that instruments designed to extract context-independent regularities systematically miss the most important features of the phenomena they study. A survey averages across contexts. A controlled experiment controls for context. A statistical analysis aggregates context into error terms. Each method produces findings that hold on average across situations — findings that are universally true and practically useless for any particular situation. Phronetic social science is the methodological response: research designed to preserve rather than eliminate the contextual features that constitute the phenomenon.
The philosophical concept traces to Aristotle. Polanyi's Personal Knowledge (1958) developed the modern formulation. Flyvbjerg synthesized the tradition with pragmatist and phenomenological resources in Making Social Science Matter (2001).
Constitutive, not incidental. Context is not noise to be abstracted away but the essential dimension of the knowledge itself.
Resistant to transmission. Context-dependent knowledge cannot be fully captured in propositions and therefore cannot be transmitted without loss through purely propositional systems.
Developmental. The knowledge is acquired through sustained engagement with situations over time, not through instruction or information transfer.
Politically consequential. Because context-dependent knowledge cannot be codified, it resists the centralization and automation that serve certain institutional interests.
Methodologically demanding. Research on context-dependent phenomena requires case-based, longitudinal, ethnographic approaches that the dominant paradigm has marginalized.