The First AI Winter and the Expert Systems Era: 1970s–1990s
AI research went through two major funding collapses (1974-1980 and 1987-1993) when overpromised capabilities failed to materialize and government and corporate sponsors pulled back.
- ·Understand what caused the AI winters and why funding dried up
- ·Know what expert systems were and why they failed to scale
- ·Explain the pattern of AI hype cycles
The AI Winters: When Progress Stalled (1970–2000)
The history of AI is not a straight line upward. It includes two major periods of disillusionment and funding collapse — known as "AI winters" — that deeply shaped how the field developed.
The First AI Winter (1974–1980)
By the early 1970s, the gap between AI's promises and its reality was becoming obvious:
- ›The Lighthill Report (1973): UK government commissioned a review that concluded AI had failed to meet its goals
- ›Funding cuts: DARPA slashed AI research budgets after years of disappointing results
- ›Key failures: machine translation (thought to be almost solved) was exposed as extremely hard
- ›The core problem: researchers hadn't accounted for the "combinatorial explosion" — the number of possible states in any real problem was astronomical
The Expert Systems Era (1980–1987): Brief Revival
A new approach emerged: instead of general intelligence, build systems encoding expert knowledge as rules.
Expert systems were rule-based programs that could reason within a narrow domain:
- ›MYCIN (1972-76): diagnosed bacterial infections better than medical interns
- ›XCON (1980): configured VAX computers for Digital Equipment Corporation — saved $40M/year
- ›DENDRAL: identified chemical compounds from mass spectrometry data
By 1985, expert systems were a $2 billion industry. Fortune 500 companies had AI departments. It seemed AI had found its footing.
The Second AI Winter (1987–1993)
Expert systems turned out to be extremely brittle:
- ›They couldn't learn from new examples — every new rule had to be manually coded
- ›They couldn't handle situations outside their pre-programmed rules
- ›Maintenance was a nightmare: "knowledge acquisition bottleneck"
- ›The hardware they ran on (Lisp machines) was outcompeted by cheaper general-purpose PCs
DARPA again cut funding. AI labs shrank. The term "AI" became toxic — researchers started calling their work "machine learning" or "knowledge engineering" to avoid the stigma.
Key Insights
- Two AI winters (1974-80, 1987-93) caused by gap between hype and reality — a pattern that has repeated
- Expert systems were the dominant AI approach of the 1980s: rule-based, narrow, and brittle
- XCON saved DEC $40M/year — one of the first clear AI ROI case studies
- Expert systems failed because they couldn't learn, couldn't handle exceptions, and were expensive to maintain
- The term 'AI' became so stigmatized that researchers rebranded their work as 'machine learning'
Why It Matters
AI winters happen when expectations exceed delivery for too long. We are not in one now, but the pattern is instructive: hype phases produce overcommitment, and when capabilities plateau, capital flees. Understanding the dynamics helps you discount excessive hype today and recognize when a particular sub-field (e.g., autonomous vehicles, certain agent claims) is heading toward its own mini-winter.