Concept · Chapter 1: What Is Artificial Intelligence?
The Knowledge-Acquisition Bottleneck
The knowledge-acquisition bottleneck is the discovery that the hardest part of rule-based AI is getting the knowledge in: much of what experts know is tacit, full of exceptions, and too vast to write down.
The problem
Symbolic systems are only as good as their rules, and real-world knowledge is enormous, uncertain, and largely unconscious.
The solution
None within symbolic AI itself — knowledge engineers could only interview experts and write more rules. The eventual answer was to learn knowledge from data instead.
The consequence
This bottleneck is the core reason AI turned from hand-written rules to machine learning — and it explains why data, not rules, is the raw material of modern AI.
You should understand first
- The Turing Test
- Symbolic AI
- Logic and Rules
- Expert Systems
- Knowledge Representation
- The Knowledge-Acquisition Bottleneck
Three reasons writing knowledge down fails
1. Experts can't tell you everything they know. An expert radiologist can't list the rules they use to spot a tumour; you can't explain how you recognize a friend's face. Much expertise is tacit — the philosopher Michael Polanyi put it as knowing more than we can tell.
2. Every rule has exceptions — and the exceptions have exceptions. Birds fly, except penguins, except… Each fix adds rules; rules begin to conflict; a change in one place breaks another. Systems become brittle: excellent inside their domain, failing abruptly just outside it.
3. Common sense is enormous. Real-world reasoning depends on millions of mundane facts nobody thinks to write down. Hand-encoding them turned out to be a decades-long project with no end in sight.
Feel it
In Write the Rules you build a spam filter by hand. You can get most of the messages you can see right — then new messages arrive, and your rules get many of them wrong. You could keep adding rules forever. That is the bottleneck in miniature.
Try it · toy model
Build a spam filter the 1980s way — by writing rules — then watch it meet new messages. Finally, let a learning algorithm work it out from examples instead.
The way out
If knowledge can't be written down, perhaps it can be learned: show the machine thousands of labelled examples and let it find the patterns itself. That idea — present since Samuel's checkers player and Rosenblatt's perceptron in the 1950s — became the dominant approach once enough data and computing power existed. See From rules to learning.
Why should I care?
As a researcher
It frames the central bet of modern AI: that knowledge can be learned from data at scale. Current work on data quality, synthetic data and data curation is the same bottleneck, moved from rules to examples.
As an engineer
Any system full of hand-coded business rules shows the same symptoms — growing exceptions, fragile interactions, costly maintenance. It's a strong signal to consider a learned component.
Modern systems that depend on it
- The shift to machine learning
- Data-centric AI
- Why pretraining on huge corpora works
Historical context
Before
Early confidence that intelligence came from general reasoning methods plus enough encoded knowledge.
After
Statistical machine learning and deep learning, which acquire knowledge from examples; today's LLMs absorb vast amounts of it from text.
Used today
As a design lens: when rules keep growing and breaking, learn from data instead — and when data is scarce, the bottleneck just moves to data collection.
What to remember
- The bottleneck is getting knowledge in, not reasoning with it.
- Experts know more than they can articulate (tacit knowledge).
- Rules multiply with exceptions and interact unpredictably (brittleness).
- Common sense is vast and mostly unwritten.
- Machine learning's answer: learn the knowledge from examples.