Classifiers as Knowledge Models
The Concept
A trained classifier can be understood as a crystallized representation of tacit knowledge, the implicit expertise that domain experts have built through experience but can’t easily articulate.
Why This Matters
We usually think of ML models as pattern-matching engines. Train on examples, find statistical regularities, apply to new data. This framing is technically accurate but misses something important.
When the training data consists of human judgments, the model is learning to approximate human decision-making. If those humans are experts making professional judgments, the model is approximating expertise.
This reframes what you’re building. You’re not just building a classifier. You’re building an operational representation of what experts know but can’t fully explain.
How It Works
Drawing from Nonaka’s knowledge creation theory:
Explicit knowledge can be expressed in formal language, data, specifications, manuals. It’s shareable and storable.
Tacit knowledge is personal and hard to formalize. Intuitions, hunches, pattern recognition built through years of experience. It’s demonstrated in practice but not fully expressible in words.
Market researchers reviewing video for highlights are exercising tacit knowledge. They know a valuable moment when they see it, but they can’t give you a checklist that fully captures what they’re looking for.
When you train a classifier on their selections, you’re aggregating the tacit models of multiple experts into an explicit, operational form. The classifier embodies their collective judgment.
Implications
For model development: The quality of your training labels matters differently. You’re not just looking for accuracy, you’re looking for labels that reflect genuine expertise. Noisy labels from non-experts produce models that crystallize noise, not knowledge.
For model interpretation: Attention weights, feature importance, and other interpretability techniques become tools for surfacing tacit knowledge, not just debugging models. What did experts implicitly care about that even they didn’t realize?
For organizational knowledge: This creates a path from tacit to explicit knowledge, one of the hardest knowledge management problems. The model doesn’t replace experts, but it makes their expertise portable, scalable, and examinable.
For model evaluation: Beyond accuracy metrics, ask: does this model make decisions the way an expert would? Would an expert look at its outputs and say “yes, that’s right”? Pure statistical evaluation may miss systematic departures from expert reasoning.
Limitations
The model captures statistical regularities in expert judgment, not the reasoning behind it. It can approximate what experts select without understanding why.
Tacit knowledge includes context-sensitivity that may not be present in training data. Experts might make different judgments in different situations in ways the model can’t adapt to.
If expert judgment is biased or inconsistent, the model crystallizes those biases and inconsistencies.