What is symbolic AI?
Symbolic AI is the approach to artificial intelligence in which knowledge is represented as human-readable symbols (concepts, facts, rules, and relationships) and intelligence comes from manipulating those symbols with logic. A symbolic system might hold the fact “pump P-204 feeds boiler B-1” and the rule “if a pump feeding a boiler fails, the boiler must be shut down,” and derive the shutdown without ever having seen a failure. The reasoning is explicit, and each conclusion can be traced to the facts and rules that produced it.
Symbolic AI was the dominant paradigm from the 1950s through the 1980s and produced expert systems, automated planners, theorem provers, and the field of knowledge representation. It is sometimes called “good old-fashioned AI.” Its counterpart is sub-symbolic or connectionist AI, in which knowledge is spread across the numeric weights of a neural network and cannot be read directly.
Why symbolic AI matters
Symbolic methods remain the right tool where correctness, consistency, and explanation matter more than flexibility. Rules engines run eligibility, pricing, and compliance logic in banks and insurers. Constraint solvers schedule production lines and crew rotations. Ontologies and knowledge graphs give enterprise data a shared meaning. Planning algorithms sequence the actions of robots and software agents. None of these learn from data, and that is their strength: they do exactly what they were told and can show their work.
Their weakness is the knowledge acquisition bottleneck. Someone has to write the rules, and the world keeps producing cases the rules did not foresee. Symbolic systems also struggle with raw perception, which is why reading an image or a free-text email is handled by neural models.
Symbolic AI in practice
A city permitting office runs a rules-based system that checks building applications against zoning code, setback requirements, and fire regulations. Each rejection lists the specific rule violated and the code section behind it, which applicants and appeals boards can verify. A language model has since been added to read the submitted plans and populate the facts the rules engine needs, forming a neuro-symbolic pairing in which perception is learned and judgment is coded.
What to watch for
- Treat the rule base as a governed asset with owners, tests, and version history.
- Watch for rule interactions as the base grows, since conflicting rules fail quietly.
- Do not expect a symbolic system to handle inputs outside its vocabulary; pair it with a learned model for perception.
- Use symbolic components to enforce the constraints that a neural model cannot be trusted to respect on its own.