The Second AI Winter

Part 1 — When Commercial Success Turned into Disappointment

Introduction

History often teaches an important lesson:

Success can create expectations that become impossible to satisfy.

The First AI Winter taught researchers that intelligence was far more complex than they had originally believed. In response, the AI community shifted its focus away from ambitious attempts to build machines with general intelligence and toward solving practical, domain-specific problems.

This strategic shift gave rise to Expert Systems, one of the first commercially successful applications of Artificial Intelligence.

Unlike the experimental programs of the 1960s, Expert Systems generated real economic value. They assisted physicians in diagnosing diseases, helped engineers configure complex computer systems, supported financial institutions in decision-making, and enabled organizations to preserve valuable expert knowledge in software.

For the first time, Artificial Intelligence moved beyond research laboratories and entered corporate boardrooms.

By the mid-1980s, AI was no longer viewed as a speculative academic discipline. It had become a commercial technology attracting billions of dollars in investment. Major corporations established AI divisions, governments expanded funding programs, and universities introduced specialized courses to meet growing demand.

Everything appeared to indicate that Artificial Intelligence had finally found its place in industry.

Yet within only a few years, this optimism collapsed.

• Projects were abandoned.
• Dedicated AI companies disappeared.
• Investors withdrew funding.
• Organizations that had enthusiastically adopted Expert Systems began questioning whether the promised benefits justified their enormous costs.

This period of declining confidence became known as the Second AI Winter.

Unlike the First AI Winter, which was primarily driven by unrealistic scientific expectations, the Second AI Winter emerged because commercial reality exposed the practical limitations of deploying AI at scale.

Understanding why this happened is essential because many of the challenges organizations faced during the 1980s continue to influence AI projects today.

From Research Success to Commercial Expectations

The revival of AI during the Expert System era created a fundamental change in how organizations viewed Artificial Intelligence.

During the 1960s, AI was primarily a research discipline. Researchers asked scientific questions such as: Can machines reason? Can computers solve logical problems? Can software imitate human expertise?

By the 1980s, those questions had evolved into business questions. Executives were asking: Can AI reduce operational costs? Can AI replace expensive specialists? Can AI increase productivity? Can AI improve decision-making? Can AI become a competitive advantage?

This shift dramatically changed how AI projects were evaluated.

Academic research often measures success through scientific discovery. Businesses measure success through measurable outcomes.

The Explosion of the Expert System Industry

The commercial enthusiasm surrounding Expert Systems grew rapidly throughout the early and mid-1980s.

• Large organizations invested heavily in AI initiatives.
• Consulting firms established AI divisions.
• Specialized AI software companies appeared around the world.
• Universities expanded Knowledge Engineering programs.
• Professional conferences attracted thousands of participants.

Several governments—including those of the United States, Japan, and Europe—announced ambitious national AI initiatives.

The Japanese Fifth Generation Computer Systems Project, launched in 1982, became one of the most ambitious government-funded AI programs of its time. It aimed to develop computers capable of advanced logical reasoning and knowledge processing, reflecting the widespread confidence that AI would soon transform computing.

At the same time, corporations believed Expert Systems would fundamentally change how businesses operated.

This optimism created an investment cycle remarkably similar to the one experienced during the 1960s. Once again, expectations began growing faster than technological capabilities.

Why Organizations Initially Loved Expert Systems

To understand why disappointment eventually followed, we must first understand why businesses embraced Expert Systems so enthusiastically.

From a management perspective, Expert Systems appeared almost ideal.

Imagine a company with one highly experienced engineer responsible for diagnosing failures in expensive industrial machinery. That engineer possessed twenty years of accumulated knowledge. If the engineer retired, the organization risked losing decades of practical expertise.

Expert Systems promised to solve this problem by converting human expertise into software.

This offered several compelling advantages:

• Expertise became available around the clock.
• Knowledge could be shared across multiple locations.
• Decision-making became more consistent.
• New employees could receive guidance without requiring constant supervision from senior specialists.
• Organizations became less dependent on individual experts.

For executives, these benefits translated into reduced operational risk and potentially significant financial savings. It is therefore understandable why businesses invested so aggressively in AI during this period.

The Hidden Assumption Behind Expert Systems

Despite their success, Expert Systems were built upon an assumption that gradually proved unrealistic.

Researchers believed that expertise could be completely described using explicit rules.

Consider an experienced pilot responding to unexpected weather. A pilot simultaneously evaluates wind speed, aircraft weight, fuel reserves, visibility, air traffic, airport conditions, previous experience, and intuition developed over thousands of flight hours.

Much of this decision-making occurs almost automatically. When asked to explain every step, even the pilot may struggle.

This revealed an important reality: not all human expertise can be converted into written rules.

Some knowledge exists as intuition, experience, pattern recognition, and subconscious judgment.

The Knowledge Acquisition Crisis

As organizations attempted to expand Expert Systems into increasingly complex domains, another challenge emerged.

Building the software was no longer the most difficult task. Collecting expert knowledge became the primary bottleneck.

This process required Knowledge Engineers to spend months—or even years—interviewing specialists, documenting decision processes, validating rules, and resolving contradictions.

The larger the system became, the slower this process grew.

For many projects:

• Development timelines expanded significantly.
• Costs exceeded original budgets.
• Experts became unavailable.
• Business requirements changed before the system was completed.

Organizations discovered that maintaining an Expert System often required continuous involvement from both software engineers and domain specialists.

What initially appeared to be a one-time investment gradually evolved into an ongoing operational expense.

Knowledge Changes Faster Than Software

Another unexpected challenge involved the changing nature of knowledge itself.

Business policies evolve. Medical guidelines change. Tax regulations are updated. Manufacturing processes improve. Scientific discoveries replace older assumptions.

Every change required modifications to the Knowledge Base.

Unlike modern Machine Learning systems, Expert Systems could not automatically learn from new information. Human experts needed to manually update the rules.

As systems expanded to thousands of interconnected rules, maintenance became increasingly difficult. A small policy change could require modifications across dozens—or even hundreds—of related rules.

Organizations soon realized that maintaining these systems could become as expensive as developing them.

The Limits of Rule-Based Intelligence

Perhaps the most important lesson of the Expert System era was that intelligence cannot always be reduced to predefined rules.

Real-world environments contain uncertainty, incomplete information, conflicting evidence, changing conditions, and unexpected situations.

Humans adapt naturally because they continuously learn from experience. Expert Systems did not.

If a situation had not been anticipated by the system’s designers, the software often failed to produce useful recommendations.

This limitation became increasingly apparent as organizations attempted to deploy Expert Systems beyond narrow, well-defined domains.

Recap: Expert Systems won commercially then stalled on tacit knowledge and maintenance—next, see how data, compute, and learning algorithms revived the field in AI Renaissance.