The Rise and Fall of Expert Systems: AI Winters

Expert systems were the first major commercial wave of artificial intelligence. During the 1970s and 1980s, knowledge-base programs advised users in chemistry, medicine, geology and computer configuration. Companies invested in knowledge engineers, dedicated Lisp machines and thousands of rules. The market then contracted sharply and AI entered a period described as winter.

The history is directly relevant to LLMs. Expert-system advocates hoped that enough rules would approach general reasoning; today, similar hopes are attached to parameters and data. The technology is different, but the error is familiar: competence inside a carefully constrained task can be mistaken for reliability outside it.

How an expert system worked

A classic expert system separated three components:

a knowledge base of facts, rules and domain heuristics;
an inference engine that matched rules to facts and derived conclusions;
an interface and explanation facility through which users supplied evidence and asked why a recommendation was made.

A rule might say that when conditions A and B hold, confidence in hypothesis C should increase. Forward chaining begins with known facts and fires applicable rules. Backward chaining begins with a target hypothesis and searches for supporting premises. The system did not normally learn from millions of records. Its quality depended on experts and engineers extracting, encoding and maintaining knowledge.

That design enabled inspection, but produced a knowledge-acquisition bottleneck. Specialists often make decisions using experience they cannot fully verbalise.

DENDRAL: domain knowledge beats a universal solver

The DENDRAL project began at Stanford in 1965 under Edward Feigenbaum and Joshua Lederberg. It inferred possible molecular structures from mass-spectrometry data. The program generated candidates, rejected structures that violated chemical constraints and compared predicted consequences with observations.

DENDRAL did not discover chemistry from scratch. It used heuristics contributed by specialists. That was its central insight: earlier AI programmes searched for general problem-solving procedures, while DENDRAL showed that specialist knowledge could matter more than a universal algorithm. Stanford describes it as the first expert system.

Meta-DENDRAL later attempted to induce rules from examples, providing an early bridge between manually encoded expertise and machine learning.

MYCIN: rules, uncertainty and responsibility

MYCIN, developed at Stanford during the 1970s, was designed to support diagnosis of bacterial infections and therapy selection. It asked questions about a patient, applied rules with confidence factors and generated recommendations. It could also identify the rules behind an answer.

The project mattered technically because it addressed incomplete evidence, separated inference machinery from domain knowledge and inspired reusable expert-system shells. It also demonstrated that laboratory performance does not solve deployment. Medicine requires validation, workflow integration, current knowledge, accountability and user acceptance.

MYCIN did not become a routine clinical product. That is not evidence that its research was worthless. It shows that a safe decision system is larger than its algorithm, the same lesson facing regulated AI deployments today.

XCON and the commercial peak

XCON, originally called R1, was used by Digital Equipment Corporation to configure orders for VAX systems. Complex products contained many components and compatibility constraints, making configuration errors expensive. Rules helped select parts and identify contradictions.

Such successes triggered a boom in commercial shells, consultancies and specialised Lisp hardware. Organisations expected to capture scarce expertise, reproduce it cheaply and retain it after employees left. The proposition worked in stable, narrow domains. It became difficult as knowledge bases grew.

Why large rule bases became brittle

Every new rule could conflict with an existing one. Exceptions generated more conditions, while a changed product or policy required cascades of edits. Proving completeness was difficult, and behaviour across all combinations became hard to predict. Knowledge was distributed between the domain expert, documentation and the engineer’s interpretation.

This resembles today’s oversized system prompts. Hundreds of prose exceptions do not create a strong security boundary. A model or rule engine may satisfy a local match without understanding the organisation’s whole objective. Critical policy belongs in deterministic code, identity controls and approval workflows.

The first AI winter and the Lighthill report

The first AI winter is associated with the 1970s, although there is no exact start date. Early programs succeeded in “microworlds” with limited objects and rules but struggled in unstructured environments. Machine translation, robotics and language understanding did not meet ambitious schedules.

James Lighthill’s 1973 review for the UK Science Research Council criticised the failure to deliver broad promises and highlighted combinatorial explosion. It influenced British funding. The report did not prove that AI was impossible; it identified the gap between demonstrations and scalable systems.

US sponsors also reduced selected programmes when results missed expectations. “Winter” therefore means reduced money and confidence, not a complete stop to research.

The second winter: economics breaks the promise

The second AI winter is usually associated with the late 1980s and early 1990s. Expert systems required costly maintenance, while specialised Lisp machines lost their economic advantage to rapidly improving general-purpose workstations. Companies learned that a demonstration with hundreds of rules and a production environment with thousands of dependencies were different projects.

Expectations surrounding Japan’s Fifth Generation Computer Systems programme added to the cycle. Its goals included knowledge processing and parallel logic inference. It produced research and technology, but not the widely imagined leap to intelligent machines. When markets and sponsors lost patience, the AI label became a liability for some vendors.

Expert systems did not disappear

Business-rule platforms, configuration engines, SIEM detections and security policy systems are descendants of expert systems. They are simply less likely to be marketed as a route to general intelligence. In a stable process, an explicit rule may be superior to a model because it is deterministic, testable and auditable.

Strong architectures are often hybrid. A model interprets unstructured input, retrieval supplies evidence, while rules and permissions constrain action. In a secure LLM deployment, generative flexibility must not replace authorisation.

Expert systems compared with LLMs

Property	Expert system	Large language model
Behaviour source	Explicit facts and rules	Parameters learned from a large corpus
Explanation	Rule path can be shown	Generated rationale may not reflect causality
Update	Edit the knowledge base	Retrieval, fine-tuning or retraining
Determinism	Usually high	Probabilistic and context-sensitive
Scope	Narrow and explicit	Broad but difficult to constrain precisely
Typical failure	Missing or conflicting rule	Hallucination, prompt injection or unsafe tool use

LLMs reduce part of the knowledge-acquisition bottleneck by learning patterns from data, but sacrifice inspectability and control. Retrieval can supply current documents without guaranteeing faithful use. LLM evaluation must measure correctness, grounding and action outcomes—not just fluency.

Seven lessons for today’s AI wave

Narrow success does not demonstrate generality. Define task boundaries and out-of-scope cases.
Maintenance belongs in total cost. Models, prompts, retrieval data and integrations age.
An interface explanation is not causal proof. Fluent reasoning may be confabulated.
Knowledge without process is not a deployment. Accountability, approval and incident response remain necessary.
Critical boundaries must be deterministic. A model must not grant its own privileges.
A lab test does not replace production monitoring. Distribution shifts change risk.
Promises need falsifiable criteria. Decide how failure will be recognised before buying AI.

In practice, those lessons lead to AI red teaming, tool-call logging and least-privilege agent design.

One further operational lesson matters. Expert systems had identifiable rule owners, while foundation-model accountability can be diffused between the model vendor, application team and data owner. An organisation should maintain an AI model registry, assign an accountable owner to every use case and set a revalidation date. When a model version or retrieval source changes, previous acceptance is not automatically valid. This is the modern equivalent of versioning the knowledge base that older deployments often struggled to maintain.

Frequently asked questions

Did expert systems learn automatically?

Usually not. Domain expertise was encoded manually as rules. Projects such as Meta-DENDRAL learned some rules, but the classic expert system depended on a curated knowledge base.

Why was MYCIN not routinely deployed in hospitals?

Experimental performance was insufficient on its own. Clinical validation, accountability, workflow integration, knowledge updates and acceptance remained unresolved. It is a systems barrier, not a simple verdict on algorithm quality.

How many AI winters occurred?

Histories commonly identify two broad winters: one in the 1970s and another across the late 1980s and early 1990s. The boundaries are approximate and the impact varied across countries and subfields.

Could the current AI boom end in another winter?

History cannot predict a date. It does show that funding falls when cost and reliability fail to match claims. Useful production applications can continue to grow even if inflated expectations decline.

Sources

Want LLM flexibility without giving up enforceable controls? Talk to Breachroad about architecture, testing and AI-agent security.