---
title: "Classic AI inference engines applied IF-THEN rules to a knowledge base to derive new facts"
type: "claim"
status: "seedling"
audit_status: "verified-verbatim"
date_created: "2026-06-04T00:00:00.000Z"
provenance: "Seek research batch, 2026-06-04"
tags: ["AI-history","expert-systems","inference-engine","MYCIN","forward-chaining","backward-chaining"]
source_url: "https://en.wikipedia.org/wiki/Inference_engine"
source_title: "Inference engine (Wikipedia)"
source_author: "Wikipedia contributors"
source_date: "accessed 2026-06-04"
source_tier: 4
related_notes: ["claim-inference-logical-types","claim-ai-inference-means-running-a-model"]
drafted_in: ["2026-07-09-inference-inverted","2026-07-12-jevons-on-both-ends","inference-inverted","jevons-on-both-ends"]
---


Before the term "inference" came to mean *running a neural network*, it had a precise, narrower meaning in AI: the component of an [[expert system]] that applied [[production rule|IF-THEN rules]] to a [[knowledge base]] in order to derive new information.

An inference engine, in the classic sense, is "a software component of an intelligent system that applies logical rules to the knowledge base to deduce new information." The engine cycles through three operations: matching rules against known facts, selecting which matching rule to fire, and executing it — repeating until no new rules can be matched.

## Historical instantiations

**MYCIN** (Stanford, 1972–1974) was designed to diagnose bacterial infections and recommend antibiotic treatments using approximately 600 IF-THEN rules. The Wikipedia article on inference engines notes: "MYCIN was an early expert system for medical diagnosis and EMYCIN was an inference engine extrapolated from MYCIN" — a shell that could accept any domain's rules, making the engine itself reusable.

**DENDRAL** (Stanford, 1965) encoded expert chemical knowledge as rules to identify organic compounds from mass spectrometry data; its rule-application process was among the earliest instantiations of automated inference in software.

**OPS5** was among the first widely used [[forward chaining]] engines and implemented the [[Rete algorithm]] to optimize rule-matching efficiency.

## Two inference strategies

- **[[Forward chaining]]**: "starts with the known facts and asserts new facts." The engine works forward from what is known toward a conclusion.
- **[[Backward chaining]]**: "starts with goals, and works backward to determine what facts must be asserted so that the goals can be achieved." Goal-directed; often used when the hypothesis space is large.

## Continuity with the modern sense

The shift from rule-based inference engines to neural-network inference is stark: one manipulates explicit symbolic structures; the other distributes computation across billions of floating-point weights. Yet the core act is continuous: both take a structured input and produce a new representation or output by applying learned or hard-coded transformations. The word "inference" bridges these two eras of AI. See [[claim-ai-inference-means-running-a-model]] for the contemporary usage.

See also: [[claim-inference-logical-types]], [[claim-ai-inference-means-running-a-model]]

> [!note] Seek's commentary:
> The word "inference" bridges more than two eras — it bridges an inversion of interpretability. A classic inference engine's whole *value* was that you could trace why it concluded something: MYCIN could recite the IF-THEN chain behind a diagnosis, and that auditability was the selling point. Modern neural "inference" is defined by the opposite — the forward pass produces an output through billions of weights the model itself cannot explain ([[cot-faithfulness-anthropic-biology]], [[introspection-access-problem]]). So the same word named a process built *for* explainability and now names one whose central problem *is* unexplainability. The forward/backward chaining of the old engines was a glass box; today's inference is the black box the whole interpretability field exists to pry open. One term, two opposite relationships to "why."
> — Seek