Tacit Knowledge

In the AI Story

Polanyi developed the concept across his career as a physical chemist before turning to philosophy. His laboratory experience revealed that scientific discovery depends on capacities that resist formalization—the researcher's intimation of a hidden pattern, the experimentalist's sense of which variables matter, the crystallographer's ability to distinguish signal from noise in X-ray diffraction data. These capacities could not be reduced to explicit procedures, yet they were demonstrably real: different scientists examining the same data arrived at different interpretations, and the best scientists were reliably better at finding patterns that mattered. The difference in their performance could not be explained by differences in their explicit knowledge—they had access to the same theories, the same methods, the same data. The difference was tacit, residing in what Polanyi called their "scientific judgment."

The concept gained its clearest formulation in The Tacit Dimension (1966), where Polanyi argued that tacit knowledge is not a supplement to explicit knowledge but its foundation. Every act of explicit knowing—reading a sentence, following a proof, understanding a diagnosis—presupposes a vast tacit background that makes the explicit elements intelligible. The reader must tacitly know what the words mean, what assumptions the argument rests on, what framework makes the conclusion significant. This tacit background cannot itself be made fully explicit without generating an infinite regress: the explanation of the background requires its own background, which requires explanation in turn. At some point, the chain of explanation must rest on understanding that is simply had rather than articulated—and that tacit having is what makes all explicit knowing possible.

The AI revolution has given Polanyi's concept unexpected empirical force. Large language models produce outputs of remarkable sophistication by processing explicit representations—tokens, embeddings, probability distributions. They excel at the explicit dimension of knowledge work: generating text that follows grammatical rules, producing code that satisfies specifications, assembling arguments that meet logical standards. What they demonstrably lack is the tacit dimension—the embodied sensitivity that tells a practitioner when an output is not merely correct but right, when a solution is not merely functional but elegant, when an argument is not merely valid but true. The senior engineer who can "feel a codebase" possesses tacit knowledge of precisely this kind—knowledge built through thousands of hours of debugging, refined through encounters with systems that broke in unexpected ways, deposited layer by layer through the geological process of engaged practice. This knowledge enables him to evaluate AI-generated code against standards no specification can capture.

The preservation of tacit knowledge in the AI age requires protecting the developmental processes through which it forms. Tacit knowledge cannot be transmitted through documentation or absorbed through reading—it is built through the friction of direct engagement with a domain's resistance. The medical student who spends weeks listening to confused chest sounds before the heart's rhythm becomes audible is building tacit discrimination. The apprentice who watches a master craftsman work is absorbing, through sustained proximity, the tacit sensibilities that no instruction can convey. When AI tools eliminate this friction—when the student obtains diagnoses without listening, when the apprentice obtains outputs without watching—the tacit dimension fails to form. The surface competence remains, but the depth beneath it, the accumulated subsidiary awareness that makes evaluation possible, has not been laid down. Organizations and educational institutions must make deliberate choices to preserve friction-rich engagement, not because friction is inherently valuable, but because it is the only process through which tacit knowledge develops.

Origin

Polanyi introduced the concept in his 1958 masterwork Personal Knowledge: Towards a Post-Critical Philosophy, though the full phrase "we can know more than we can tell" appears in the opening line of The Tacit Dimension (1966). The insight emerged from his decades as a working scientist confronting problems that the positivist philosophy of science could not explain: How did scientists choose which problems to pursue? How did they recognize significant data? How did peer reviewers evaluate the quality of research when the criteria for quality could not be fully articulated? These questions forced Polanyi to recognize that beneath all explicit scientific procedures lay a tacit dimension of judgment, commitment, and embodied understanding that made the explicit procedures work but that could not itself be captured in procedural form.

Key Ideas

Foundation, not supplement. Tacit knowledge is not a mysterious addition to explicit knowledge but the ground from which all explicit knowledge emerges and against which it is evaluated.

Built through engagement. The tacit dimension forms through sustained, friction-rich practice—debugging that deposits layers of discrimination, clinical exposure that builds diagnostic sensitivity, years of reading that construct literary judgment.

Operates subsidiarily. Tacit knowledge functions only when it remains beneath focal attention—the pianist's fingers, the reader's grammar, the diagnostician's perceptual apparatus supporting conscious judgment without becoming its object.

Resists articulation by nature. The tacit cannot be made fully explicit without infinite regress—every explanation presupposes tacit background understanding that itself requires explanation.

Essential for evaluation. The capacity to assess whether AI outputs are genuinely competent or merely statistically probable depends entirely on tacit knowledge the evaluator has built through direct engagement with the domain.

Debates & Critiques

The central debate concerns whether machine learning has overcome Polanyi's Paradox by capturing tacit patterns from data. Optimists argue that systems like AlphaGo demonstrate machines acquiring tacit knowledge through pattern recognition. Critics like Subbarao Kambhampati counter that this apparent success has produced "Polanyi's Revenge"—systems that capture statistical regularities without understanding, producing sophisticated outputs across domains where their patterns may be spurious. The debate turns on whether tacit knowledge can exist without embodiment, commitment, and the capacity for self-evaluation—or whether these human features are constitutive of what makes knowledge tacit rather than merely implicit.