Does refusing explicit knowledge harm AI system performance?

The AI field's romance with tacit knowledge — learning everything from data, refusing to incorporate explicit causal models, domain rules, or codified expertise — is creating avoidable problems. Polanyi's Revenge names the irony: Polanyi's paradox said we know more than we can tell. AI's tacit-learning agenda inverts this: we can tell (we have explicit domain knowledge) but AI refuses to hear it.

The pattern is pervasive. Researchers build Rubik's Cube solvers from billions of examples rather than accepting the eight simple rules governing the puzzle. Industry practitioners convert doctrine and standard operating procedures into "data" only to have the knowledge "learned back" from that data at enormous cost. Policy infrastructure for AI relies exclusively on massive datasets even when hard-won explicit knowledge exists.

The costs are direct:

Interpretability: When systems learn their own representations from raw data, there is no reason to believe their reasoning will be interpretable to humans. Explicit knowledge, by contrast, provides the structural vocabulary for explaining decisions — "the system applied rule X to fact Y." Pure tacit learning produces weights that serve no interpretive function.

Bias: Systems that learn from data inherit whatever is statistically dominant in that data, with no explicit signal to override it. Explicit knowledge can include normative corrections that data alone cannot supply ("do not discriminate by X, regardless of correlation").

Robustness: Tacit learners generalize in the direction of training distribution. Explicit rules can enforce invariances that data doesn't adequately represent. A system that "knows" a causal rule can maintain it when data is sparse or adversarially constructed.

The civilizational argument is pointed: human progress has been built on codification — approximate, aspirational, but explicit. The current AI agenda is running the opposite direction, learning to distrust codification in favor of raw statistical patterns. This is a historically unusual choice with consequences.

Connects directly to domain specialization: Can prompt optimization teach models knowledge they lack? confirms this is not a prompting problem — the explicit knowledge must enter at training time. Can organizing knowledge structures beat raw training data volume? shows that structured explicit knowledge injection at 0.3% of corpus size substantially closes the gap.