Tacit Knowledge & Scaffolding
A Tri-Domain Analysis
We can know more than we can tell
— Michael Polanyi, The Tacit Dimension (1966)Knowledge Across Three Domains
Exploring how tacit knowledge emerges, transfers, and becomes scaffolded across human, animal, and artificial systems.
Human Domain
Complete tacit knowledge spectrum
Humans possess all three categories of tacit knowledge, from procedural skills to embodied intuition. Language functions as a scaffold that restructures tacit knowledge, enabling externalization and manipulation in ways unavailable to non-linguistic systems.
Codifiable but Costly
Procedural skills, pattern recognition, expert intuition
Interpretive Nuance
Social cues, contextual judgment, artistic appreciation
Embodied Knowledge
Sensorimotor skills, physical intuition, bodily experience
Meta-Cognitive Access
Ability to reflect on and partially articulate knowledge
Animal Domain
Foundational substrate of knowing
Polanyi held that animal knowledge is wholly tacit—they possess the foundational substrate from which human symbolic capacity emerged. Tool use in non-human primates demonstrates knowledge that is learned, refined through practice, and transmitted socially.
Codifiable but Costly
Learned behaviors, conditioned responses, trial-and-error learning
Interpretive Nuance
Social recognition, emotional contagion, limited contextual judgment
Embodied Knowledge
Instinctive behaviors, motor skills, physical adaptation
Meta-Cognitive Access
No evidence of explicit meta-cognition or self-reflection
Artificial Domain
Statistical pattern emergence
Current AI demonstrates two of three categories: codifiable-but-costly knowledge and interpretive nuance, while lacking physical embodiment. LLMs "know more than they can tell" in their weights—performing tasks without being able to fully explain their methods.
Codifiable but Costly
Learned weights, statistical patterns, trained behaviors
Interpretive Nuance
Language nuance, style adaptation, contextual interpretation
Embodied Knowledge
No physical substrate or sensorimotor experience
Meta-Cognitive Access
No self-model or genuine reflection capability
Types of Tacit Knowledge
Three distinct categories refined from Polanyi's foundational framework.
Codifiable but Costly
Knowledge that could be made explicit but at prohibitive cognitive or computational cost. Examples include procedural knowledge of riding a bicycle or pattern recognition of expert chess players.
Interpretive Nuance
Nuance, subtext, and contextual judgment that cannot be fully captured in rules. Includes understanding social cues, appreciating artistic quality, or making clinical diagnoses.
Embodied Knowledge
Sensorimotor, situated, and physical knowledge inseparable from bodily experience. As Polanyi noted, "the skill of a driver cannot be replaced by thorough schooling in the theory of the motorcar."
Types of Scaffolding
Five distinct but interrelated forms of scaffolding that support knowledge transfer.
Cognitive Scaffolding
Structures supporting conceptual understanding—analogies, worked examples, graphic organizers
Education & Expert SystemsPhysical Scaffolding
Embodied interaction structures—tools, environments, and physical constraints that shape learning
Craft & RoboticsSocial Scaffolding
Interpersonal knowledge transmission through mentorship, apprenticeship, and collaborative learning
Human DevelopmentAlgorithmic Scaffolding
AI training architectures—curriculum learning, hierarchical RL, progressive neural networks
Machine LearningRelational Scaffolding
Shared attention, mutual agency, and resonant presence required to initiate knowledge transfer
All DomainsKey Insights
The LLM Tacit Knowledge Paradox
Mechanistic interpretability reveals that LLMs "know more than they can tell" in their weights—they can perform tasks without being able to fully explain their methods, and can interpret nuance in ways that suggest genuine understanding. Yet they cannot explain their own neuron activations, and their "knowledge" emerges from statistical patterns rather than embodied experience.
Can Scaffolding Create Genuine Tacit Knowledge?
This question hinges on whether tacit knowledge requires subjective experience or can exist as purely functional competence. Current AI demonstrates functional competence without clear evidence of subjective experience, suggesting simulation rather than genuine tacit knowing.
This reframing shifts attention from individual cognition to relational dynamics, from possession of knowledge to conditions of its emergence.
Evaluating Artificial Tacit Knowledge
Six criteria for recognizing artificial tacit knowledge if it emerges.
Inarticulability Criterion
The system can perform tasks it cannot fully explain. Necessary but not sufficient—current LLMs already meet this.
Transfer Criterion
The system can apply knowledge learned in one context to novel, structurally similar contexts without explicit retraining.
Contextual Adaptation Criterion
The system adjusts its behavior based on subtle contextual cues that were not explicitly trained.
Scaffold Independence Criterion
The system can function effectively when external scaffolding is removed, having internalized relevant structures.
Relational Engagement Criterion
The system demonstrates genuine mutual recognition and can engage in repair sequences when communication breaks down.
Embodied Grounding Criterion
For embodied systems, knowledge is grounded in sensorimotor experience rather than purely symbolic manipulation.