Tacit knowledge and Polanyi’s paradox

Working Definition

Tacit knowledge is the dimension of human cognition comprising embodied skills, diagnostic connoisseurship, rules of taste, and perceptual orientation that an agent deploys effectively in action but cannot fully articulate in explicit propositions, formal definitions, or step-by-step instructions. For Michael Polanyi, tacit knowing is not a defective, pre-scientific residue awaiting formalization, but the indispensable foundation of all explicit knowledge: every mathematical formula, legal statute, or laboratory manual requires an uncodified skill to read, interpret, and apply it to a concrete case. Derived from this insight by economist David H. Autor, Polanyi’s paradox states that our tacit mastery of how the world works vastly exceeds our ability to state the explicit rules governing our own performance—a gap that for decades blocked the computerization of seemingly effortless human tasks.

Genesis and Global Lineage

The concept was forged by the Hungarian-British physical chemist and philosopher of science Michael Polanyi in his magnum opus Personal Knowledge (1958) and distilled in The Tacit Dimension (1966), building on Gilbert Ryle’s distinction between “knowing how” and “knowing that” (1949) as well as Gestalt psychology. Polanyi—for decades a world-renowned experimentalist in chemical kinetics and X-ray diffraction before turning to philosophy—wrote in opposition both to logical positivism’s ideal of wholly impersonal scientific detachment and to the Soviet program of centrally planned science (which he campaigned against alongside Friedrich Hayek in the Society for Freedom in Science). He analyzed the functional anatomy of tacit knowing as a triad linking two kinds of attention in a “from–to” relation: when a blind person explores a street with a cane or a surgeon probes a cavity, they do not direct “focal awareness” to the pressure of the handle against their palm, but rely on those tactile impacts as “subsidiary awareness,” attending from them to the object at the tip of the probe. In the 1990s, the concept was co-opted by corporate knowledge-management theory (most famously Ikujiro Nonaka and Hirotaka Takeuchi’s 1995 SECI model), which routinely flattened it into not-yet-documented employee know-how that HR departments could “extract” into databases—directly inverting Polanyi’s thesis that the tacit root can never be eliminated.

Central European Lineage

Michael Polanyi’s epistemology grew out of the extraordinary intellectual ferment of fin-de-siècle Budapest in the final decades of the Austro-Hungarian Empire. Born in Budapest in 1891 as Mihály Polányi (younger brother of the economic historian Karl Polanyi), he earned his medical degree at the University of Budapest in 1913, served as a medical officer in the Austro-Hungarian army during the First World War, and completed a doctorate in physical chemistry there in 1917 under Gusztáv Buchböck. In Budapest he was an active member of the progressive student Galileo Circle (Galilei Kör, founded in 1908 with his brother Karl as first president) and frequented the Sunday Circle (Vasárnapi Kör, 1915–1919), the philosophical salon around György Lukács, Béla Balázs, and Karl Mannheim where participants debated the crisis of European positivism and the existential grounding of culture. Following the collapse of the Hungarian Republic and the rise of Miklós Horthy’s regime in 1919, Polanyi emigrated first to Karlsruhe and Berlin—where at the Kaiser Wilhelm Institute he worked alongside Fritz Haber, Albert Einstein, and fellow Budapest native Eugene Wigner—and in 1933, fleeing Nazi Germany, accepted a chair at the University of Manchester. Almost five decades after The Tacit Dimension (1966), MIT economist David H. Autor elevated Polanyi’s Budapest-to-Manchester maxim—“we can know more than we can tell”—into “Polanyi’s paradox” (2014), making it a cornerstone of the global debate on automation and artificial intelligence.

Why Now

Between 2018 and 2026, Polanyi’s paradox stood at the epicenter of the deep learning revolution and the rise of large language models (LLMs). For half a century, classical symbolic artificial intelligence (“Good Old-Fashioned AI”) repeatedly crashed against Polanyi’s barrier: engineers tried to interview human experts and hand-code explicit rules for recognizing faces, driving in traffic, or translating idiomatic prose, yielding brittle decision trees. The breakthrough of deep neural networks and the Transformer architecture after 2017 succeeded precisely by circumventing Polanyi’s paradox rather than solving it directly: instead of asking humans to state rules they could not articulate, machine learning inferred statistical patterns tacitly from billions of recorded examples of human output. As a result, between 2022 and 2026 Polanyi’s paradox flipped on its head: artificial neural networks generated their own opaque, machine counterpart to tacit knowledge—matrices of billions of floating-point weights capable of writing functional code or spotting tumors on a scan without being able to offer a faithful chain of explicit rules for how they arrived at the answer (hence the crisis of explainable AI, or XAI). At the same time, in skilled trades, wet-lab experimentation, and physical caregiving, embodied somatic knowledge continued to resist robotic automation far more stubbornly than white-collar text processing.

Explanatory Power

Polanyi’s distinction between subsidiary and focal awareness explains why attempts to hyper-rationalize skilled performance and decompose it into explicit rules often paralyze the very practice they aim to improve. When a concert pianist mid-performance or a tennis player mid-serve shifts focal attention from the musical phrase or the trajectory of the ball to the individual muscles of their fingers and wrist, they immediately stumble and lose fluency. The exact same pathology operates at the scale of institutions: when a hospital, university, or editorial desk replaces apprenticeship and clinical connoisseurship with exhaustive checklists and matrices of key performance indicators (KPIs), practitioners are forced to abandon holistic pattern recognition in favor of ticking isolated boxes. At the macroeconomic level, tacit knowledge explains why transferring cutting-edge industrial capabilities—such as extreme ultraviolet (EUV) semiconductor lithography or mRNA vaccine manufacturing—never succeeds through patent licenses and blueprints alone: without moving teams of engineers who carry the calibration of instruments “in their fingertips” (indwelling), a factory built from identical schematics can struggle for years to achieve viable yields.

Axes of Debate

Underappreciated Angle

In the era of generative AI, the deepest risk is not that machines will appropriate the tacit knowledge of today’s senior experts, but that they will saw off the bottom rungs of the ladder by which tomorrow’s humans acquire tacit knowledge in the first place. Expertise in law, software architecture, medicine, or editing is built not by reading finished summaries, but through thousands of hours of friction with raw material in junior roles. While manually debugging code, drafting briefs, or reading routine scans, a novice accumulates the dense network of subsidiary clues that later triggers an instant intuitive alarm when something is wrong. If firms replace junior work with generative models and keep seniors merely as reviewers, once the current cohort of masters retires there will be no one left with the tacit grounding required to spot a subtle machine hallucination.