Following the start of research into AI in the 1950s, AI development saw bursts of activity followed by periods, known as AI winters, where general interest (and research funding) declined. The initial optimistic claims made for the technologies proved harder to realise resulting in disillusionment and a decline in interest by commercial and political (read financial) supporters. This, aligned with researchers establishing the natural limits and problematic hurdles in continuing to use the previously promising technologies, prevented further progress at that time.    

Two AI winters are generally recognised to have occurred between 1974 and 1980, and then again between 1987 and 1994. In the main, AI research did continue throughout the AI Winters – just with reduced funding and less commercial and public interest.

Despite these rises and falls many of the pre-winter AI technologies continued to find successful industry applications and have been adapted to fit a wide range of real-world problems.  The technology developed did not go away but tended to lose its association with the label of “Artificial Intelligence” and more realistic claims in application and performance were made.  The rise of ‘Smart’ devices, marketing product ‘recommenders’, optimisation tools and route planners, are all examples of pre-AI winter AI technologies finding commercial application in later years, and there are many more.  

Many of the early attempts to build AI systems focused on the ability to build rule-based systems.  Such systems held rules defining how to recognise symbols and combinations of symbols alongside rules that could manipulate those symbols.

Created using ChatGPT free version 04/03/2025 by Kevin Waugh. Prompt used: Draw a cartoon of a litter picking robot choosing where to place an item it has picked, it can choose between 4 bins -one red, one blue, one green and one orange.

Imagine a litter picking robot sorting recyclable waste – when it has an item in its grip, it needs to decide where to put it. A simple linear rule set might be:

Early successes in the 1950s were in systems that could do mathematical reasoning and automated logic-proof systems in highly constrained systems. These techniques could also be applied to simple rule-based games, using rules to capture common strategies. 

This works well in small, simple games like Draughts (aka checkers) and Noughts and Crosses, but they weren’t able to play games such as chess where rules could not exhaustively cover strategic play.

https://www.shutterstock.com/shutterstock/photos/2126891213/display_1500/stock-vector-tic-tac-toe-game-vector-illustration-hand-drawn-tic-tac-toe-game-2126891213.jpg

It became clear that the limiting factor for rule-based systems was the number and complexity of the rules needed to navigate real-world problems. Think about the complexity of the rules needed to walk down the street, cross the road, enter a coffee shop, order a coffee and sit down to drink it, or to identify a trophy fish in a photograph. Research continued into rule-based systems, but it was gradually recognised that it was not the rules themselves that were important, it was how those rules were combined and chosen to be applied.  The focus turned to capturing rules and the knowledge of how to apply them. The aim was to build systems that would attempt to apply rules according to how an expert would solve such problems.  They became known as expert systems.

There was also a shift from developing techniques intended to make general AIs that try to be applicable to a wide range of tasks, to techniques that would be able to mimic a knowledgeable user working on well-defined limited tasks. The research focus also moved to knowledge management.

The first expert systems were developed just before the first AI winter. MYCIN (1972) was a system that could diagnose diseases of the blood. The system started out with around 600 rules and would ask the operator (a doctor) Yes/No questions to explore the available evidence, before combining these in a simple ‘inference engine’ (effectively assessing the weight of the available evidence) to produce a decision. When asked, it would produce the English paraphrases of the rules used to reach that decision. The rule-based and expert system approaches were technologies that could be easily understood – MYCIN, for example, could report the rules and choices that it used to reach a decision. So, it was possible to put trust in what the system was doing; if you wanted re-assurance, you could always ask it to report its decision making.

Research conducted at the Stanford Medical School discovered that researchers found MYCIN’s recommendations to be acceptable 65% of the time. This compared to between 42.5% and 62.5% of the recommendations of five human experts (Yu, 1979).

MYCIN demonstrates the utility of these rule-based expert systems. However, there are limits to how big a rule set and how complex the interaction of rules (strategies) can become and remain ‘processable’ in a reasonable time.  It’s also very hard for these systems to extend their application to novel problems.  For example, an expert system for solving medical problems in blood disorders could not learn to deal with auto-immune diagnosis.

The second AI winter began with the realisation that expert system knowledge bases and inference engines would not be sufficient for expert systems to tackle open real-world problems.  There are limitations to asking experts how they solve some problems – not least that it turned out that sometimes experts don’t know; they’ve solved some problems so many times they apply intuition when the problem recurs.

However, as with rule-based and script-based processing, the work on expert systems did not go away.  Currently, expert systems are widely used across a range of focused,  real-world problem areas such as medical diagnosis and treatment recommendations, financial risk assessment, customer service chatbots, cybersecurity threat analysis and legal case analysis; essentially, any domain where complex decision-making requires specialized knowledge and can be codified into rules, allowing the system to mimic the reasoning of a human expert in that field.  

The early 1990s saw a move away from capturing and encoding rules and knowledge towards a data-centred approach and a growth in the technologies of machine learning.   These allow data to be automatically classified and rules to be inferred without the need to explicitly capture and represent them manually.   At the same time, natural language processing researchers had found better ways of making representations of language with numbers to encode words, phrases, and sentences in data structures that could be processed more easily within computers. Knowledge engineers had started to build ‘semantic webs’ that could capture relationships between words and phrases to encode knowledge.  This allows language processing systems to understand sentences where simple grammar rules tell you about the structure but not the meaning. The groundwork had been laid for the Generative AIs and Large Language Models. 

Aside note

This is a limited account of the technologies and applications and problems in the periods concerned – a much richer picture is shown by the  graphic “A brief, mostly complete, history of Artificial Intelligence.” by Danielle J Williams (history of AI poster) which highlights important successes across a range of related application areas.