The First AI Winter, 1974–1980: The Great Disillusionment

The funding collapse, the broken promises, the researchers who lost their jobs — and the underground survivors who kept the flame alive.

Edinburgh, Scotland. Autumn 1974. The machine intelligence unit at Edinburgh University — one of the most productive AI research groups in the world, the place where Donald Michie had been building robots and training programs and asking fundamental questions about machine learning for the better part of a decade — is being dismantled.

Not all at once. Not with a single announcement or a clean break. The way research groups die is slower and more demoralising than that. First the grants don’t get renewed. Then the postdoctoral researchers who were counting on those grants take positions elsewhere. Then the graduate students finish their theses and leave — some for other universities, some for industry, some for fields where the funding is more reliable and the future less uncertain. The equipment sits idle. The corridors get quieter.

In an office that is getting emptier around him, a researcher named Rod Burstall is trying to finish work that he has been pursuing for years. He will finish it. Others will not.

This is what an AI winter looks like from the inside: not a sudden freeze but a gradual cooling, a slow withdrawal of warmth and resources, a landscape that becomes incrementally harder to survive in until some things cannot survive at all.

Edinburgh was not the only place this was happening. It was happening in the United States too — in the laboratories that had grown fat on DARPA contracts, in the universities that had built AI departments on the confidence of the 1960s, in the research programmes that had attracted the best students and the most ambitious researchers on the strength of predictions that were not being fulfilled. The winter was everywhere. And it lasted longer than anyone hoped.

What Caused the Winter: A Summary of Forces

The first AI winter was not caused by any single event. It was the product of several forces converging — the Lighthill Report and similar reviews, the ALPAC machine translation assessment, the general reassessment of AI funding by DARPA and other government agencies, the accumulated evidence that the predictions of the 1960s were not going to be met on the promised timelines, and the growing scepticism of funding bodies that had expected more and received less.

We have discussed the Lighthill Report in detail in the previous article in this series, and the ALPAC Report before it. Here, the focus is different: not on the causes of the winter but on its experience — what it was like to be a researcher in AI during the years when the funding dried up and the credibility collapsed, what happened to the people and the programmes, and what, if anything, survived.

Understanding the winter experientially matters because the experience shaped the field in ways that the policy analysis of causes and consequences does not fully capture. The researchers who lived through the first winter carried its lessons for decades. Those lessons — about the danger of overpromising, about the importance of demonstrable results, about the need to manage expectations and maintain credibility — shaped the culture of AI research for a generation. They shaped how research was communicated, how results were qualified, how the relationship between basic research and practical application was managed.

The winter was not just a funding event. It was a formative experience for the people who lived through it. And those people built the field that emerged on the other side.

The Funding Landscape Before the Winter

To understand what was lost in the winter, it helps to understand what the funding landscape looked like in the years before it — the golden age of the early 1960s, when AI research was generously supported and expectations were high.

DARPA — the Defence Advanced Research Projects Agency, the US military’s research and development arm — was the primary funder of American AI research in the late 1950s and throughout the 1960s. DARPA’s model of research funding was unusual and, for a time, extraordinarily productive: it funded basic research with long time horizons, trusted researchers to follow their own intellectual instincts, and did not demand immediate practical results. The contract that DARPA signed with MIT’s Project MAC in the mid-1960s gave the lab millions of dollars per year with minimal strings attached — an arrangement that allowed the lab to attract the best researchers and to work on problems at the frontier of what was possible.

Similar arrangements funded the Stanford AI Laboratory, Carnegie Mellon’s computer science department, and a handful of other institutions that became the centres of American AI research. The funding was not just generous — it was culturally enabling. It created environments where researchers could take intellectual risks, pursue long-term projects that would not produce publishable results for years, and focus on fundamental problems rather than on the incremental, applied work that more tightly constrained funding tended to produce.

The British funding situation was somewhat different — more distributed across universities, more channelled through the Science Research Council and similar bodies, less concentrated in a small number of flagship institutions. Edinburgh’s machine intelligence unit, the groups at Sussex and Essex, the early AI work at Cambridge — these were funded at lower levels than the American equivalents and were therefore more vulnerable to the kind of sweeping review that the Lighthill Report represented.

The funding of the 1960s was justified, to the funders, by the predictions that AI researchers were making. Machine translation would be practical within a few years. Robots would be capable of general manual labour within a decade. Programs would achieve human-level performance across intellectual domains by the end of the century. These predictions attracted money. When the predictions were not fulfilled, the money was reconsidered.

The Withdrawal: How Funding Dried Up

The withdrawal of funding in the early 1970s was not a clean event with a precise date. It was a process that played out over several years, affecting different institutions and different research programmes at different times and to different degrees.

The earliest and most immediate impact was on machine translation. The ALPAC Report of 1966 had specifically targeted machine translation research, and the response from US funding agencies was swift: funding for machine translation was dramatically reduced almost immediately after the report’s publication. Research groups that had been built around the machine translation problem — at Georgetown, at MIT, at the RAND Corporation — lost their primary funding source and had to either redirect their work or disband.

The effect was sharp and lasting. Machine translation research in the United States effectively ceased as a funded activity for more than a decade. The researchers who had been working on it — some of the most linguistically sophisticated people in the AI field — moved to other problems or left academia. The knowledge they had accumulated, the approaches they had developed, the understanding of language that they had built up over years of work — this was partially lost, scattered to other fields or simply abandoned.

The broader impact on AI funding came a few years later, as the effects of the Lighthill Report and the Mansfield Amendment combined with the accumulated disappointment of unfulfilled predictions to change the funding environment.

The Mansfield Amendment, passed by Congress in 1969 and taking effect in fiscal year 1970, restricted DARPA funding to research with direct military relevance. This was a significant constraint on an agency that had been funding basic research with minimal attention to immediate practical applications. The amendment forced DARPA to review its AI portfolio and redirect funding toward work with more obvious military utility.

The review produced significant cuts. The most generous, most unconstrained funding — the kind that had made MIT’s AI Lab and Stanford’s SAIL into world-class research environments — was reduced or restructured. Projects that could not demonstrate a plausible military application lost their funding. Projects that could — computer vision for reconnaissance, natural language processing for intelligence analysis, robotics for military logistics — survived or even grew, but under constraints that shaped what research could be done.

The British cuts were more concentrated and more dramatic, because the Lighthill Report gave the Science Research Council a specific, high-profile justification for reductions that might otherwise have been harder to defend. Edinburgh’s machine intelligence unit was the most visible victim, but it was not the only one. Research groups across Britain found their funding reduced, their positions not renewed, their graduate students unable to secure the postdoctoral positions that would normally follow a successful dissertation.

Edinburgh: The Most Visible Casualty

Edinburgh’s machine intelligence unit deserves particular attention, both because it was the most prominent victim of the British winter and because its story illustrates the human dimensions of what a funding cut means in practice.

Donald Michie had built the unit at Edinburgh from the late 1960s as one of the most ambitious and most productive AI research groups in the world. Michie was a singular figure — a geneticist by training who had become one of the most important figures in machine learning research, a man of enormous intellectual energy and infectious enthusiasm who attracted talented people and inspired them to work on problems that were both fundamental and practically important.

The unit’s research covered a wide range: machine learning, robotics, game-playing programs, pattern recognition, the formal foundations of AI. Michie himself had done important work on the MENACE system — a matchbox-based machine that could learn to play noughts and crosses through reinforcement, a beautiful demonstration of learning principles using the most modest possible computational substrate. He had been involved in work on the Freddy robot — one of the first robots capable of assembling simple objects from their components, a remarkable demonstration of what computer vision and robotic manipulation could achieve even with the limited technology of the early 1970s.

The Freddy project was, in many ways, exactly the kind of work that the Lighthill Report criticised: robotics research that achieved impressive results in carefully controlled laboratory conditions but whose broader applicability was uncertain. Lighthill’s critique of robotics was that the gap between laboratory performance and real-world performance was so large that the research was not justified. Michie disagreed, vigorously and publicly — he was one of the participants in the BBC debate — but he could not prevent the consequences.

The funding cuts that followed the Lighthill Report hit Edinburgh disproportionately hard. Michie’s unit lost a significant fraction of its funding. Researchers left. Projects were abandoned or severely scaled back. The Freddy robot, which had been one of the unit’s most impressive achievements, sat idle as the team that had built it dispersed.

Michie himself continued to work on machine learning — he was not the kind of person who gave up — but the institution he had built was permanently diminished by the winter. Edinburgh never fully recovered its pre-winter status as a global centre of AI research. It remained active and productive, but the concentration of talent and resources that had made it exceptional in the late 1960s was gone.

The American Winter: Less Visible, Still Painful

The American AI winter was less dramatic than the British one — there was no single report with the impact of Lighthill, no single moment of public reckoning — but it was real, and its effects on the research community were significant.

The DARPA funding cuts of the early 1970s did not destroy the major American AI labs — MIT’s AI Lab and Stanford’s SAIL survived, though with reduced resources. But the funding environment changed in ways that mattered. The unconstrained basic research funding of the 1960s was replaced by more targeted, more applied, more militarily justified contracts. The freedom to follow intellectual curiosity without attention to practical application — the freedom that had made the 1960s so productive — was substantially reduced.

The change in funding culture produced a change in research culture. Researchers who wanted funding had to frame their work in terms of practical applications. Projects that were purely basic, purely exploratory, with no identifiable near-term application, became harder to justify. The field began to move, subtly but perceptibly, from the ambitious basic research of the 1960s toward more applied, more incremental work that could be more convincingly justified to funders who had grown sceptical of grand promises.

The graduate students who would have entered AI research in the mid-1970s made different choices. Some chose other areas of computer science — database systems, operating systems, programming languages — where the funding was more reliable and the career prospects more certain. Some chose other fields entirely. The pipeline of new talent into AI research was reduced, and the effects on the field’s development were felt for years.

The industrial interest in AI that had been building through the 1960s — IBM, the RAND Corporation, various defence contractors had been hiring AI researchers and funding AI projects — also cooled. The disappointment of machine translation had been particularly damaging to corporate confidence: companies that had invested in machine translation systems and received systems that required extensive human post-editing had learned a lesson about the gap between AI promises and AI performance. The lesson generalised: AI research in the early 1970s was not something that corporations were eager to fund.

What Was Lost: Research Programmes That Did Not Continue

One of the most important and least discussed consequences of the first AI winter was the research programmes that did not continue — the lines of inquiry that were interrupted before they reached their conclusions, the problems that were set aside before they were solved, the approaches that were abandoned before they had been fully explored.

Machine translation is the most obvious example. The work that was happening before the ALPAC Report was cut off — approaches to syntax analysis, semantic representation, and cross-linguistic transfer that had been accumulating over a decade — was largely abandoned. When machine translation research resumed in the 1980s, it had to start substantially from scratch, rediscovering insights that had already been developed and then lost.

The robotic manipulation work that Michie’s group and others had been doing — the work on Freddy and similar systems that was trying to combine computer vision with dexterous robotic manipulation to produce machines capable of handling real objects in partially structured environments — was similarly cut off. The knowledge accumulated in those projects, the understanding of what worked and what did not, the appreciation of the specific difficulties that made robotic manipulation hard — this was partially lost as the teams dispersed. When robotics research resumed in the 1980s, some of the lessons had to be relearned.

Neural network research suffered a different kind of loss. The Minsky-Papert critique of perceptrons in 1969 had already damaged the neural network research programme before the funding cuts of the early 1970s made things worse. By 1974, neural network research was deeply unfashionable — a territory associated with failed promises, with the discredited pattern recognition approaches that had been tried and found wanting, with the kind of undisciplined speculation that the Lighthill Report had criticised. The few researchers who continued to work on neural networks did so against the prevailing current, with limited funding and limited recognition.

The loss here was not a set of completed results that were then abandoned. It was a set of approaches that were not pursued, questions that were not asked, developments that would have happened but did not because the people who would have made them happen had left the field or been directed elsewhere. The counterfactual is always uncertain, but it seems plausible that the deep learning revolution of the 2010s might have happened decades earlier if the neural network research of the late 1960s and early 1970s had not been disrupted.

The Survivors: Who Kept Working and How

Not everyone left. Some researchers continued to work on AI through the winter, sustained by personal conviction, by modest funding from sources that remained interested, or by the kind of institutional momentum that carries research groups through periods of external adversity.

The most important group of survivors, in retrospect, were the neural network researchers who kept working despite the general scepticism of the field.

Geoffrey Hinton is the most important of these. Hinton had begun his PhD at Edinburgh in 1972 — at exactly the moment the Lighthill Report was being finalised — working on computational models of the brain. He was, from the beginning, committed to the connectionist approach: the belief that intelligence was best understood as the property of networks of interconnected processing units, and that the right way to build AI was to build and train such networks rather than to program explicit rules.

This was not a popular position in the mid-1970s. The symbolic AI approach was the mainstream, and neural networks were the fringe. Hinton’s doctoral research was technically solid but not leading to the positions in major AI labs that would have been the natural next step for a talented graduate student committed to a different approach. He moved between positions in Britain and the United States through the late 1970s, continuing to work on connectionist models and developing the theoretical understanding of how such models could learn that would eventually produce the backpropagation algorithm.

Yann LeCun, who would become one of the key figures in the neural network revival, was an undergraduate in France in the late 1970s and early 1980s, beginning to think about learning systems and neural networks before the field had any serious momentum behind it. Yoshua Bengio was younger still — a teenager in Quebec as the winter was at its coldest. Both of them, along with Hinton, would eventually form the core of the team that revived neural networks and built the foundations of deep learning. But in the mid-1970s, that future was not visible.

In the symbolic AI tradition, research continued — reduced but not eliminated. McCarthy at Stanford, Minsky at MIT, and their colleagues continued to work on knowledge representation, planning, natural language understanding, and the other problems that had been central to the field. The funding was tighter, the atmosphere was more defensive, but the research did not stop. Expert systems — the approach that would eventually revive the field’s commercial fortunes — were being developed during the winter years, quietly and without the public attention that would come later.

DENDRAL and MYCIN: The Seeds of the Recovery

While the broader AI field was contracting, a specific and more focused approach to AI was being developed that would eventually pull the field out of the winter and into a new era of commercial activity. The approach was expert systems — AI programs that captured and applied the specialised knowledge of human experts in specific domains.

DENDRAL, developed at Stanford in the late 1960s by Edward Feigenbaum, Bruce Buchanan, and their colleagues, was an early and influential example. The program was designed to identify the chemical structure of organic compounds from mass spectrometer data — a task that required specialised chemical knowledge and pattern recognition abilities that human chemists had developed through years of training and experience.

What made DENDRAL different from earlier AI programs was its approach to knowledge. Rather than trying to implement general problem-solving strategies — the approach of Newell and Simon’s General Problem Solver — DENDRAL encoded specific, domain-relevant chemical knowledge directly into the program. The knowledge was not derived from first principles or from general reasoning. It was the knowledge of expert chemists, captured and formalised and made available to the program in a form it could use.

This approach — encoding expert knowledge explicitly, rather than deriving it through general reasoning — was less intellectually ambitious than the General Problem Solver approach. It made no claims about general intelligence. But it worked. DENDRAL performed well on its specific task, matching or exceeding the performance of expert chemists in the limited domain it was designed to handle.

MYCIN, developed at Stanford in the early 1970s by Edward Shortliffe and colleagues, applied the same expert-knowledge approach to medical diagnosis — specifically to the diagnosis of bacterial infections and the recommendation of antibiotic treatments. MYCIN used a knowledge base of several hundred rules, derived from interviews with expert physicians, to reason about a patient’s symptoms and test results and produce diagnosis and treatment recommendations.

MYCIN was evaluated rigorously — comparisons of its recommendations against those of expert physicians showed that it performed at or near the level of human experts on the specific class of problems it was designed to address. This was a genuine achievement, and one that had obvious practical significance: if a program could provide expert-level medical advice in a specific domain, then doctors and hospitals in settings where specialist physicians were unavailable might use it to improve the quality of care.

DENDRAL and MYCIN were being developed during the winter years — MYCIN’s development ran from 1972 to 1980, entirely within the period we are discussing. They did not attract the public attention that the grand AI projects of the 1960s had attracted. They were not claiming to have solved the problem of machine intelligence. They were claiming something more modest: that in a specific, well-defined domain, a program that encoded expert knowledge carefully could perform at expert level.

This modesty was precisely what made them credible, and eventually what made them commercially attractive. The expert systems era of the 1980s — which would pull AI out of the winter and into a new period of investment and activity — grew directly from the work on DENDRAL and MYCIN. The seeds of the recovery were planted in the coldest years of the winter.

The Natural Language Processing Thread: What Survived

Natural language processing — the attempt to build systems that could understand and generate human language — was one of the areas most severely damaged by the ALPAC Report and the subsequent funding cuts. But it was not entirely abandoned, and the work that continued through the winter years laid important groundwork for later developments.

Terry Winograd’s SHRDLU, developed at MIT between 1968 and 1970, was perhaps the most impressive natural language understanding system of the early AI era. SHRDLU could understand and respond to natural language commands about a simplified “blocks world” — a virtual environment containing coloured blocks, pyramids, and other simple shapes that a robot arm could manipulate. Within this limited domain, SHRDLU’s language understanding was genuinely impressive: it could handle complex, nested sentences, pronoun resolution, clarification requests, and other aspects of natural language that had previously seemed out of reach.

The problem with SHRDLU — which Winograd himself articulated clearly in his subsequent work — was that its impressive performance in the blocks world did not generalise. The natural language understanding it achieved was dependent on the extremely restricted and carefully designed domain it operated in. Move SHRDLU outside the blocks world, ask it about anything other than block stacking, and its understanding collapsed immediately. The system had not learned anything general about language. It had been engineered to handle the specific linguistic patterns that arose in conversations about blocks.

Winograd went on to draw important conclusions from this experience, published in his influential 1972 doctoral dissertation and in subsequent work with Fernando Flores. His conclusion was not that natural language processing was impossible but that it required a fundamentally different approach — one that was grounded in human practices and social contexts rather than in formal symbolic representations of meaning. This conclusion pointed away from the mainstream of symbolic AI and toward the more phenomenological and pragmatic approaches that he would develop in subsequent decades.

Meanwhile, researchers in computational linguistics — a field that was emerging at the intersection of linguistics, psychology, and computer science — continued to work on the formal analysis of language structure. The Government-Binding theory being developed by Noam Chomsky and his colleagues in the early 1970s was generating new insights into the structure of human language that had implications for natural language processing, even if those implications were not immediately computational. The work on formal grammars, semantic representations, and the syntax-semantics interface that was happening in linguistics departments throughout the 1970s would eventually feed into the computational approaches of the 1980s and beyond.

The thread was thin during the winter years, but it was not broken.

The Robotics Thread: Shaky But Surviving

Robotics research — one of the areas most directly targeted by the Lighthill Report — also continued through the winter, though at reduced scale and with reduced ambition.

The Shakey robot, developed at SRI International (formerly the Stanford Research Institute) in the late 1960s and early 1970s, had been one of the most ambitious robotics projects of the pre-winter era. Shakey could perceive its environment through a television camera, reason about what it saw using a planning system, and execute simple navigation tasks in a carefully controlled environment. It was impressive and genuinely pioneering — but it was also slow, fragile, and very much a laboratory demonstration rather than a practical system.

The STRIPS planning system that was developed for Shakey — a formal language for representing the state of the world and the effects of actions, and an algorithm for generating sequences of actions that would achieve a goal — was one of the most important technical contributions of the era. STRIPS influenced the development of planning theory in AI for decades, and its underlying concepts remain relevant in current AI planning research. The Shakey project produced real and lasting intellectual contributions even as it demonstrated, in its brittleness and its dependence on artificial conditions, the difficulties that Lighthill had identified.

After the funding cuts, robotics research contracted but did not disappear. At MIT’s AI Lab, researchers continued to work on manipulation and robot perception, developing new approaches to the sensorimotor problems that the earlier generation of robots had struggled with. At SRI and at Carnegie Mellon’s Robotics Institute (founded in 1979, near the end of the winter), work continued on mobile robot navigation and on the formal planning methods that were needed to direct robot behaviour.

The Carnegie Mellon Robotics Institute, founded by Raj Reddy and his colleagues, was in some ways a product of the winter — an attempt to create an institutional home for robotics research at a time when its position within computer science departments was insecure. By focusing on robotics as an applied engineering discipline rather than as a branch of abstract AI research, the Institute could justify itself on practical grounds that were more credible to funders who had grown sceptical of AI’s grand claims. The approach worked: the Institute attracted funding and talent and became one of the world’s leading robotics research centres.

The Psychology Connection: Cognitive Science Emerges

One of the unexpected consequences of the AI winter was the accelerated development of cognitive science as a discipline distinct from AI — a discipline that studied intelligence scientifically, without necessarily being committed to building it.

The researchers who had been interested in the relationship between AI and psychology — who had seen AI programs as models of human cognition, as computational hypotheses about how human minds worked — did not lose their intellectual questions when the AI funding dried up. They found different institutional homes and different funding sources in psychology and neuroscience departments, and they brought with them the computational frameworks and the precise, mechanistic style of theorising that AI had developed.

The result was cognitive science: an interdisciplinary field that combined cognitive psychology, computational modelling, linguistics, neuroscience, and philosophy to study the mind scientifically. The first cognitive science conference was held in 1979. The Cognitive Science Society was founded in 1979. The new field attracted researchers from multiple disciplines who found that the questions they were working on could not be adequately addressed within any single established discipline.

For AI, the emergence of cognitive science was both a loss and a gain. It was a loss because some of the most talented researchers who might have built AI systems were instead building theoretical models of human cognition — doing science rather than engineering. It was a gain because cognitive science provided a rigorous scientific foundation for the AI project — a body of empirical results about how human cognition worked that could inform the design of AI systems and serve as a check on unrealistic assumptions.

The relationship between AI and cognitive science has been productive and sometimes tension-filled ever since. The question of whether AI should be trying to replicate human cognition or to achieve intelligent behaviour by whatever means works is a question that the two fields continue to debate. The winter, by temporarily separating the two programmes, gave each space to develop more distinctively — and produced, eventually, two mature fields that have a more nuanced and productive relationship with each other than the seamless unity of the early 1960s had permitted.

The Personal Stories: What Happened to People

The AI winter was not just a policy and funding story. It was a human story — a story of people whose careers were disrupted, whose research was abandoned, whose plans for the future had to be revised.

Some researchers left AI entirely. They moved to other areas of computer science — database systems, programming languages, software engineering — where the funding was more stable and the career prospects more certain. Some of these moves were temporary: researchers who had left during the winter returned when the field recovered in the 1980s, bringing with them skills and perspectives they had developed elsewhere. Others were permanent: people who had been building careers in AI found that other fields were more rewarding and stayed where they had moved.

Graduate students were particularly vulnerable. A PhD student who began their doctoral research in AI in 1970 or 1971, expecting to complete their degree and enter a robust academic or industrial research market, found themselves graduating into a market that had contracted dramatically. The AI faculty positions that had been abundant in the late 1960s were no longer available. The industry jobs at IBM and the defence contractors that had been hiring AI researchers were no longer there. The students who had chosen AI because it seemed like the most exciting frontier in computer science found that frontier suddenly much colder and much less welcoming.

Some of these students redirected their research toward the applied, expert-systems approaches that were beginning to show commercial promise. Some left academia for industry, taking their AI knowledge into companies that could use it in ways that did not depend on the grand claims that had become impossible to sustain. Some left computing entirely, finding that the skills they had developed in AI — rigorous thinking, formal modelling, problem decomposition — were valuable in other fields.

The human cost of the winter is hard to quantify. Some of the people who left the field would have made important contributions if they had stayed. Some of the research programmes that were abandoned would have produced valuable results if they had continued. The winter was not catastrophic — it did not end AI as a discipline or permanently destroy the intellectual project. But it caused real harm to real people, and it delayed progress on real problems, and those facts deserve acknowledgment.

The International Dimension: How Different Countries Fared

The first AI winter was primarily an Anglo-American phenomenon — the countries most affected were the United States and Britain, where the promises had been loudest and the subsequent disappointment most acute. Other countries experienced the winter differently.

Japan was less affected, in part because Japanese AI research had a different character — more focused on specific, applied problems in robotics and natural language processing for Japanese text, less committed to the grand general claims that had been the target of the Lighthill Report. Japanese robotics research, funded by the Ministry of International Trade and Industry and the industrial companies that were developing manufacturing automation, continued to make progress through the 1970s. The Japanese industrial robot industry, which would eventually become the most advanced in the world, had its roots in research that continued through the winter years.

France maintained a distinctive tradition of AI research, rooted in the country’s strong mathematical culture and its interest in formal methods. The PROLOG programming language — developed at the University of Marseille in 1972 — was a product of the French AI tradition. PROLOG, based on the logical programming paradigm that McCarthy had been developing with LISP, provided a different approach to AI programming that was particularly suited to the kind of rule-based reasoning that expert systems required. PROLOG would play an important role in the Japanese Fifth Generation Computer Project of the 1980s — the ambitious national AI initiative that we will discuss in a later article.

The Soviet Union had its own AI research tradition, largely isolated from Western developments but working on similar problems from different starting points. Soviet AI research survived the winter years partly because it was insulated from the Western funding landscape and partly because the Cold War imperative to maintain scientific and technological competition with the West provided its own funding rationale. What Soviet AI research produced in the 1970s remains less well documented in the Western literature than it deserves, in part because of the language barrier and in part because the isolation of Soviet science meant that results were not always published in internationally accessible venues.

The Winter’s End: How It Thawed

The first AI winter did not end with a single event, any more than it had begun with one. It thawed gradually, through a combination of forces that accumulated through the late 1970s and early 1980s.

The most important force was commercial interest in expert systems. DENDRAL and MYCIN had demonstrated that expert knowledge could be captured in programs that performed at expert level in specific domains. By the late 1970s, entrepreneurs and corporate researchers were beginning to recognise that this demonstration had commercial implications. If a program could perform expert medical diagnosis, it could also perform expert financial analysis, expert equipment maintenance diagnosis, expert configuration of complex technical systems. The market for such programs was large.

The first commercial expert system, R1 (later renamed XCON), was developed at Carnegie Mellon in collaboration with the Digital Equipment Corporation in the late 1970s. R1 configured VAX computer systems — determining which components should be included in a system based on a customer’s specifications. This was a complex task that required specialised knowledge and was, at DEC, performed by human experts who were expensive and in short supply. R1 performed the task automatically, accurately, and at a fraction of the cost of human experts. DEC estimated that R1 was saving the company millions of dollars per year by the early 1980s.

R1’s success demonstrated, conclusively, that AI could produce commercially valuable results. It attracted the attention of investors, corporate research departments, and venture capitalists who had not previously taken AI seriously. The commercial expert systems industry that grew from this demonstration was worth hundreds of millions of dollars by the mid-1980s — a scale of commercial activity that had not existed before and that brought new funding, new talent, and new credibility to the field.

The second force was the maturation of computing hardware. The computers of the mid-1970s were substantially more powerful than those of the mid-1960s, and the computing hardware of the late 1970s was more powerful still. The price-performance ratio of computing was improving rapidly — following the trajectory that Gordon Moore had described in 1965, with transistor density on chips doubling approximately every two years. The problems that had been intractable in the 1970s because they required more computing power than was available were becoming tractable as that power became cheaper and more abundant.

The third force was intellectual: the emergence of new ideas and new approaches that offered credible paths forward on problems that the field had struggled with. The development of constraint satisfaction as a general approach to combinatorial problems, the formalisation of non-monotonic reasoning and default logic, the application of probabilistic methods to planning and perception — these theoretical developments of the late 1970s were not dramatic breakthroughs, but they represented real progress on real problems and provided a basis for cautious optimism.

By 1980, the first AI winter was effectively over. Funding was returning. Commercial interest was growing. New approaches were producing real results. The field that emerged was different from the field that had entered the winter — more modest in its claims, more focused on specific applications, more rigorous in its evaluation of results. But it was alive, and it was beginning to grow again.

What the Winter Taught the Field

The first AI winter was one of the most important events in the history of AI — not despite being a period of contraction and difficulty but because of it. The lessons it taught the field were painful but necessary.

The lesson of honest evaluation. Before the winter, AI researchers had a tendency to describe their results in the most optimistic possible terms — to present laboratory demonstrations as evidence of near-term practical capability, to extrapolate from success in toy domains to success in real-world applications, to predict timelines that the evidence did not support. The winter, by making those optimistic predictions directly responsible for funding cuts and public scepticism, taught the field to evaluate its results more honestly.

This lesson has not been perfectly learned — AI has gone through cycles of hype and disappointment since the first winter, as we will see in subsequent articles. But the field that emerged from the first winter was more careful about the gap between demonstration and application than the field that went into it. Researchers who had experienced the cost of overpromising were more disciplined in their subsequent public communications.

The lesson of domain focus. The expert systems approach that rescued AI from the winter demonstrated that impressive results were achievable in specific, well-defined domains, even if general intelligence remained elusive. This was a lesson about how to make progress: not by trying to solve everything at once, but by identifying specific domains where the knowledge could be captured and the performance validated, and building from those specific successes.

The lesson was partly correct and partly misleading. Expert systems were genuinely useful in their domains. But the lesson that domain specificity was the right approach to AI had a dark side: it discouraged the work on general reasoning, common sense knowledge, and broad intelligence that might eventually be more important. The field became, for a time, a collection of domain-specific tools rather than a unified programme of research toward general intelligence. The move toward domain specificity was strategically wise given the field’s damaged credibility, but it came at the cost of intellectual ambition.

The lesson of institutional vulnerability. The winter demonstrated that AI research was vulnerable to external criticism in ways that more established fields were not. Because AI had made specific, testable predictions about near-term capabilities, and because those predictions could be evaluated against evidence, the field was exposed to the kind of review that Lighthill conducted in a way that a field with more modest or more vague claims would not have been.

This vulnerability was partly a consequence of the boldness that had made AI exciting — the willingness to stake out strong positions, to make specific predictions, to claim that intelligence was achievable by machine. The boldness that attracted funding and talent also attracted scrutiny, and the scrutiny, when the predictions were not met, was devastating. The winter taught the field to protect itself institutionally by being more careful about what claims it made and what evidence it offered for those claims.

The Enduring Significance of the First Winter

The first AI winter matters not just as a historical episode but as a recurring pattern — a pattern that AI has repeated, in modified form, in subsequent years, and that remains relevant to understanding the field’s current situation.

The basic dynamic is this: a genuine technical advance produces impressive results in limited domains. Those results attract attention, investment, and predictions about near-term generalisation. The predictions are not met on the predicted timelines. Attention and investment withdraw. The field contracts. A period of more modest, more focused work eventually produces a new advance that begins the cycle again.

AI has gone through this cycle at least twice more since the first winter: the expert systems boom of the 1980s and its subsequent collapse in the second AI winter of the late 1980s and early 1990s, and the neural network revival of the 2000s that became the deep learning revolution of the 2010s. In each cycle, the technical advances were real, the predictions were too optimistic, and the subsequent correction was painful for those who had built their expectations on the optimistic predictions.

The field is in such a cycle now, with large language models playing the role of the impressive advance and the predictions about artificial general intelligence playing the role of the too-optimistic extrapolations. Whether the current cycle will end in another winter is genuinely uncertain — the technical advances are more substantial than those of any previous cycle, and the computing resources available are vastly greater. But the pattern is recognisable, and understanding its previous iterations helps calibrate the current situation.

The researchers who lived through the first winter and continued working — who kept the neural network research alive, who developed the expert systems approach, who maintained the intellectual infrastructure of the field through its most difficult years — made everything that followed possible. They deserve recognition not just for what they built but for what they preserved: the belief that the project was worth pursuing, that intelligence was understandable and buildable, that the difficulties were not permanent barriers but challenges to be overcome by different approaches.

The winter was cold. The spring eventually came. And the people who kept working through the cold made sure that there was something to thaw.