The Tacit Knowledge Dilemma in AI

The Tacit Knowledge Dilemma: Transferring What We Know But Cannot State

In 1966, the philosopher of science Michael Polanyi formulated a deceptively simple observation that remains the defining puzzle of modern artificial intelligence:

"We know more than we can tell."

Polanyi called this tacit knowledge. It is the implicit, experiential wisdom that humans possess but struggle to codify into words. You can ride a bicycle without being able to calculate the physics of balance on paper. You can recognize a friend's face in a crowd of thousands without being able to specify the exact millimeter coordinates of their cheekbones.

In software engineering, tacit knowledge is everything that separates a seasoned architect from a syntax reference book.

The Iceberg of Engineering Discernment

If you look at any codebase, explicit knowledge is merely the ten percent above the waterline: The language syntax and compiler rules. The API endpoints and schema types. * The functional tickets and unit tests.

Submerged beneath the waterline is the vast, invisible mass of tacit wisdom: Architectural Taste: Knowing intuitively when an abstraction is premature and will become technical debt. Defensive Instincts: Smelling a race condition in a concurrent loop before running a single benchmark. Contextual Pragmatism: Knowing when to adhere strictly to theoretical design patterns and when to cut through them for production reliability. Hard-Earned Scars: The visceral memory of past outages that guides your choice of data structures.

Conventional AI models are trained exclusively on what is written down—the explicit tip of the iceberg. They ingest millions of open-source repositories, but they cannot observe the hesitation in a developer's hand before they refactor a dangerous function.

Can Machines Absorb Tacit Wisdom?

The common industry approach to "training" AI assistants has been prompt engineering: writing longer system instructions telling the assistant to "write clean code" or "be an expert architect."

This approach fundamentally misunderstands tacit knowledge. You cannot tell an assistant what you cannot formulate into rules yourself.

At Autobahn Bot, we realized that tacit knowledge cannot be taught through lectures; it must be absorbed through co-presence and correction: 1. The Power of the Veto: When an engineer pairs with an assistant and rejects a suggestion, that rejection is a goldmine of tacit discernment. Why did the human reject it? Because it felt brittle? Because it broke an unspoken layout harmony? 2. Session Reflection: Rather than relying on generic rules, External Brain records which proposed files the developer accepted or rejected, and keeps the corrections the assistant hands over when the session closes. That gap is where tacit taste shows. 3. Crystallization into Reusable Instincts: Once an unstated preference is repeatedly demonstrated across several tasks, the memory substrate crystallizes it into an active rule.

The Future of Pair Programming

Artificial intelligence will never replace human taste, because taste is grounded in human experience, culture, and empathy. But by designing tools that observe how we work rather than just what we say, we bridge Polanyi's chasm—building machines that slowly learn to see the world through the creator's eyes.