FIELD NOTE / 2026.09.134 MIN READ / 5 SOURCES

Cyc and the Attempt to Encode Common Sense by Hand

Cyc pursued one of AI's most ambitious symbolic projects: manually encoding the background knowledge and exceptions that people rely on without stating them.

Cyc began from the belief that AI lacked ordinary background knowledge

Douglas Lenat launched Cyc in 1984 around a provocative diagnosis: many AI systems failed not because their reasoning algorithms were weak, but because they lacked the enormous body of unstated common sense that humans bring to every problem. A program may know formal facts about restaurants yet still fail to understand that people normally eat after ordering or that objects cannot occupy arbitrary places simultaneously. Cyc therefore set out to encode broad consensus knowledge explicitly rather than waiting for each application to rediscover it.[1]

The missing knowledge was often too obvious for textbooks

Commonsense facts are precisely the assumptions humans omit from instructions because everyone is expected to know them already.

The project chose manual knowledge engineering at extraordinary scale

Lenat and R. V. Guha’s book Building Large Knowledge-Based Systems described Cyc as an attempt to construct a very large knowledge base and inference system spanning everyday domains.[2] Knowledge engineers entered concepts, relations, rules, and exceptions by hand. The strategy was intentionally different from statistical learning. The project assumed that careful symbolic representation could accumulate a reusable substrate of world knowledge whose cost would be amortized across many applications.

Scale was not an accidental side effect

The premise of Cyc was that useful common sense requires breadth. A tiny ontology might handle one expert domain but would not support reasoning across ordinary life.

CycL provided a formal language for concepts, relations, and rules

Cyc needed more than a database of simple facts. Its representation language, commonly called CycL, supported predicates, variables, quantification, collections, relations, and rule-like implications.[3] This let the system represent both concrete assertions and general knowledge. A statement could say that a particular city is in a country, while another rule could express what normally follows from membership in a class. The expressive language was necessary because common sense includes exceptions, abstractions, temporal relations, and context-dependent claims.

Representation choices determined what could be inferred

A knowledge base is not merely a pile of sentences. Its ontology and logical structure determine which relationships the inference engine can exploit.

Microtheories addressed the fact that truths depend on context

One of Cyc’s important architectural ideas was the use of contexts, often organized as microtheories. A statement may be appropriate within one domain, time period, hypothetical situation, or level of abstraction without being universally true. Guha’s work on formalizing contexts grew directly from the Cyc project and explored how knowledge could be localized while still connected to broader reasoning.[4] This helped the system avoid treating every assertion as an unconditional fact about one flat world.

Context is a defense against contradiction

The same object or rule can be described differently in legal, physical, fictional, or hypothetical settings without forcing every statement into one universal interpretation.

The project treated knowledge acquisition as long-term infrastructure work

By 1995, Lenat described Cyc as a large-scale investment in knowledge infrastructure, reporting years of manual work, large numbers of concepts, and vast numbers of hand-authored axioms.[5] The language of infrastructure was deliberate. Cyc was supposed to become a shared semantic foundation that other programs could build on, much as an operating system supplies services to applications. The ambition went beyond one expert system: common sense would be a platform.

Cyc exposed the difficulty of deciding what ordinary knowledge actually means

Encoding common sense is harder than listing trivia. Concepts such as ownership, intention, physical containment, social roles, and causation have edge cases and depend on context. Knowledge engineers must choose distinctions that support useful inference without making the ontology impossible to maintain. Cyc therefore became a laboratory for ontology engineering itself. The challenge was not only entering facts but deciding which conceptual categories and relations were stable enough to support millions of later assertions.

Statistical AI changed the competitive landscape but not the underlying problem

Large language models acquire broad regularities from text rather than relying on teams manually entering commonsense axioms. This made Cyc’s hand-authored strategy look less attractive for many applications. Yet neural models still struggle with consistency, explicit provenance, exceptions, and reliable multi-step reasoning. The contrast is useful: Cyc optimized for inspectable symbolic structure, while statistical models optimize for learned patterns at scale. Hybrid approaches continue to revisit the possibility that learned models may benefit from explicit knowledge structures.

Why Cyc belongs in the history of knowledge systems

Cyc belongs in computing history because it pursued the commonsense knowledge problem more systematically and at greater scale than almost any earlier symbolic AI project. Lenat and Guha treated ontology, contexts, inference, and manual knowledge acquisition as infrastructure rather than application-specific details.[1][2]

The project also clarified a fundamental tension in AI. Human intelligence relies on enormous background knowledge, but explicitly writing that knowledge down is slow, contentious, and difficult to validate. Learning it statistically is faster at scale but can make the resulting knowledge hard to inspect or control. Cyc represents one extreme of that design space: the belief that common sense should be encoded explicitly enough for a machine to reason over it as a structured body of knowledge.

RESEARCH / PROVENANCE

Works Cited

5 SOURCES
  1. 01
  2. 02
  3. 03
  4. 04
  5. 05

CodeHistory is a living archive. Citations document the evidence used for this edition; later evidence may refine the account.

Contribute / Corrections

Improve the record.

Use this moderated submission form to suggest a correction, provide a source, challenge a priority claim or identify a missing contributor. Submissions are treated as research leads, not automatically published comments.

Submit a research lead

Please do not submit confidential material or claims you cannot support.