Artificial intelligence has a remarkable ability to disappoint people without ever having promised them anything. We seem quite capable of taking care of the promises ourselves.
We develop a technology, discover what it can do, and almost immediately start imagining what it ought to be capable of next. At some point, it is no longer enough for a machine to play chess, process language or write a reasonably convincing article. We want it to understand, think, solve problems and preferably improve the productivity of the entire global economy.
And when it fails to deliver, we're disappointed.
With the machine, naturally.
This has happened before. Twice, in fact. The history of artificial intelligence includes two major periods when enormous expectations collided with technical limitations and economic reality. Research funding declined, companies lost interest, and technologies that had recently seemed to represent the future became considerably less fashionable.
These periods became known as AI winters. I find the expression rather wonderful: an entire field of technology gets its own season because humans are unhappy with its progress.
The history behind it says more than the metaphor, though. Especially now, as we make another impressive collection of promises about what artificial intelligence will soon be able to do.
What Is an AI Winter?
An AI winter is a period of declining enthusiasm, investment and research funding in artificial intelligence, usually following a period of high expectations that the technology fails to meet.
The term describes more than a temporary loss of public interest. During historical AI winters, research programmes lost financial support, commercial projects struggled and confidence in the field weakened.
Historians generally distinguish between two major AI winters:
The first AI winter (approximately 1974–1980): Early AI research encountered serious technical limitations, while critical assessments contributed to reductions in research funding.
The second AI winter (from around 1987 into the early 1990s): Commercial enthusiasm for expert systems and specialised AI hardware declined as costs, limitations and unmet expectations became increasingly apparent.
These dates are approximate. AI research did not develop uniformly across countries, institutions or scientific disciplines. Nor did artificial intelligence simply stop working during these periods.
An AI winter doesn't necessarily mean that a technology has lost its capabilities. Sometimes it means that people have changed their minds about how impressive those capabilities are.
To understand why that happens, we need to go back to a time when computers were considerably less powerful, but expectations were already remarkably ambitious.
The Beginning of Artificial Intelligence: Big Ideas, Small Computers
In 1956, a group of researchers gathered at Dartmouth College in the United States to explore an ambitious possibility: could aspects of human intelligence be described so precisely that machines could simulate them? The Dartmouth research project helped establish artificial intelligence as a field of study.
The underlying idea was fascinating. If reasoning, learning and problem-solving could be represented through formal processes, perhaps computers could eventually reproduce some of the abilities we associate with human intelligence.
Early AI programmes demonstrated that machines could solve mathematical problems, perform logical operations and complete certain tasks in controlled environments. For the time, these were remarkable achievements.
The computers themselves were somewhat less remarkable by today's standards. Processing power was expensive. Memory was limited. Tasks that a modern smartphone handles without attracting attention required substantial technical resources.
Researchers began imagining what could become possible anyway. Some predicted that machines would develop increasingly sophisticated intellectual abilities within a relatively short period. Once a computer could solve a logical problem, the distance to a genuinely intelligent machine suddenly appeared rather manageable.
Looking back, it's tempting to smile at some of those predictions. Although I would like to know which of our current AI forecasts will sound equally optimistic in seventy years.
I suspect there will be a few.
The First AI Winter: Why AI Research Lost Momentum in the 1970s
During the 1960s, artificial intelligence research produced several promising results. But as researchers attempted to move beyond carefully controlled problems, the limitations became increasingly difficult to ignore.
Machine translation was a particularly revealing example. The basic idea sounds straightforward: a computer receives a sentence in one language and produces the equivalent sentence in another. Anyone who has ever tried to translate a joke knows that the situation is slightly more complicated.
Words have multiple meanings. Sentences depend on context. Some expressions only make sense because the reader brings background knowledge that never appears in the text. Early translation systems struggled with precisely these difficulties.
In 1966, the US government's Automatic Language Processing Advisory Committee published the ALPAC report, which concluded that progress in machine translation had fallen short of expectations. The report contributed to a substantial reduction in support for machine translation research in the United States.
Other branches of AI faced their own difficulties. In 1969, Marvin Minsky and Seymour Papert published Perceptrons, examining mathematical limitations of certain early neural-network models. Their findings were important, although some of the broader conclusions later drawn about neural networks went beyond what the analysis itself established.
Then, in 1973, the British mathematician James Lighthill published a highly critical assessment of artificial intelligence research. The Lighthill Report questioned whether the field had delivered on its ambitious promises, particularly when it came to handling complex, real-world problems.
Funding priorities changed. Some research programmes lost support. Expectations became considerably more restrained.
The first AI winter had arrived.
What actually failed during the first AI winter?
The computers hadn't suddenly forgotten how to solve the problems they could solve a few years earlier. A programme capable of completing a particular task in 1972 could generally still perform that task in 1975. What had changed was the interpretation of its achievement. Something that had once appeared to be an early sign of machine intelligence now looked like a limited technical application.
The capability remained.
The excitement did not.
The research limitations were real, of course, and so were the funding consequences. But the episode shows how much our assessment of technological progress depends on what we expected to happen next.
The Second AI Winter: When Expert Systems Became Less Impressive
By the 1980s, artificial intelligence was attracting renewed attention, and this time much of the enthusiasm centred on expert systems. The idea was commercially appealing. If human experts possessed valuable knowledge, why not capture that knowledge in a computer? A company could then have access to specialised expertise whenever it was needed. The digital expert wouldn't retire, take holidays or request a salary increase.
From a management perspective, this must have sounded almost suspiciously convenient.
Expert systems used collections of rules and structured knowledge to draw conclusions within particular domains. They could assist with medical diagnoses, technical troubleshooting, industrial processes and the configuration of computer systems. Some were genuinely useful.
But human expertise turned out to be considerably more difficult to capture than expected. An experienced doctor doesn't simply apply a list of rules. They recognise unusual cases, interpret incomplete information and sometimes notice that the original question was wrong. An experienced technician knows what the manual says should happen. They may also know what regularly happens despite the manual insisting that it shouldn't.
Turning all this knowledge into explicit rules required substantial effort, and as expert systems became more complex, maintaining them became increasingly difficult. New situations required new rules. Existing rules needed updating. Different rules could contradict one another. Eventually, companies discovered that a digital expert requiring constant attention from human experts wasn't quite as independent as originally hoped.
There were economic problems too. Many AI applications relied on specialised computers known as Lisp machines. As conventional computers became more powerful and affordable, the economic advantages of this hardware began to disappear. From 1987 onwards, the Lisp-machine market declined sharply, and commercial expectations surrounding expert systems became harder to justify.
Investment weakened. Enthusiasm declined. The second AI winter began.
Once again, this didn't mean that every AI application had failed. Some expert systems continued to provide value, and research carried on. But the broader promise of quickly transforming human expertise into reliable computer systems had proved much more difficult to fulfil than many people expected.
The pattern is familiar by now. We develop a system that can perform a useful task. We imagine all the other tasks it might eventually perform. Then we become disappointed when the technology stubbornly insists on remaining good at the original task.
Why Artificial Intelligence Returned After Two AI Winters
Artificial intelligence didn't disappear after the second winter. Research continued, and approaches that had previously received less public attention became increasingly important.
Machine learning offered a different way of developing intelligent systems. Instead of explicitly programming large collections of rules, researchers could train systems using data, allowing them to identify patterns and make predictions. Many of the underlying ideas had existed for decades. What changed was the environment in which they could be applied: computers became more powerful, larger datasets became available and algorithms improved. Applications that had previously been impractical became technically and economically feasible.
Neural networks, which had gone through periods of reduced attention, returned to prominence. A widely recognised milestone came in 2012, when the deep-learning system AlexNet achieved a breakthrough result in the ImageNet image-recognition competition. In 2017, researchers introduced the Transformer architecture, which became a major foundation for modern large language models.
Then, in November 2022, OpenAI released ChatGPT to the public. Suddenly, people could interact with an advanced AI system simply by typing a question. They didn't need to understand neural networks, machine learning or programming. They could ask for an explanation, a translation, a piece of code or an argument, and the system would produce an answer that often sounded as though someone had understood the question.
That was a remarkable development.
Particularly because the answer could sometimes sound equally convincing when it was wrong.
The AI Hype Cycle: Why Expectations Grow Faster Than Technology
The Gartner Hype Cycle describes a recurring pattern in how expectations around emerging technologies can develop.
A new technology attracts attention. Early demonstrations suggest exciting possibilities. Companies identify potential applications, investors become interested and media coverage increases. Expectations rise, sometimes considerably faster than the technology's actual capabilities.
Eventually, limitations become visible. Applications fail to deliver the expected results, costs become harder to justify, and initial excitement gives way to disappointment. Gartner calls this phase the Trough of Disillusionment, which sounds like somewhere you might end up during a hiking trip after being a little too optimistic about the route.
In the model, more realistic applications may eventually emerge. Organisations learn what the technology can reliably do, and useful implementations become part of everyday life.
The Hype Cycle isn't a scientific law that predicts exactly how a particular technology will develop, and it can't tell us when another AI winter will begin. But it offers a useful distinction: technological progress and public enthusiasm are not the same thing. A technology can continue improving while investment enthusiasm declines. It can be genuinely useful in certain applications while failing to meet broader commercial expectations. And it can become so familiar that we stop considering its achievements particularly impressive.
Think about machine translation. What once represented an enormous research challenge is now available on millions of phones. Most people don't organise a press conference when their device translates a short message. They simply expect it to work.
Usually, it does.
Anyone translating literature may still have a few comments.
What Makes Today's AI Boom Different?
Today's artificial intelligence differs substantially from the systems involved in the first two AI winters. Modern large language models can generate and analyse text, translate languages, write code, summarise documents, work with images and support a wide range of professional and creative tasks. These aren't merely theoretical possibilities. Many are already useful.
I work with AI regularly as a writer and content strategist. It helps me explore research questions, examine arguments, structure information and identify weaknesses in drafts. In AI for Writing: What I Use It For and What I Don't, I explain how I integrate these tools into my writing process.
But a demonstrated capability turns into a much larger expectation remarkably quickly.
An AI system can produce an article, so we begin discussing whether it will replace writers. It can summarise research, so we expect its answers to be reliable. It can formulate a convincing argument, so we sometimes assume it understands the argument in a way that we haven't quite defined.
These are not necessarily unreasonable questions. But they are different questions.
Can AI perform a task without possessing human intelligence?
The distinction between weak AI and strong AI is useful here. Weak AI generally refers to systems that perform particular tasks or groups of tasks without demonstrating unrestricted, human-like general intelligence. Strong AI is often associated with Artificial General Intelligence, or AGI, although the terminology becomes more complicated in philosophical discussions about understanding and consciousness.
Modern language models are considerably more versatile than traditional narrow AI systems. Nevertheless, impressive performance across many tasks does not, by itself, settle the question of whether a system possesses human-like general intelligence or conscious understanding. I explore these distinctions in my German article Schwache und starke KI: Warum ChatGPT intelligenter wirkt, als wir es erklären können.
For me, the practical question is slightly different. A machine can produce a beautifully formulated answer. That doesn't mean the answer is correct, and it doesn't automatically tell us whether the system understands what it has produced.
A confidently expressed mistake remains a mistake.
Even when the grammar is excellent.
Will There Be Another AI Winter?
Given the history, it's tempting to ask whether we're heading towards a third AI winter. The possibility deserves serious consideration. AI development requires substantial investment in computing infrastructure, energy, research and commercial applications. Companies and investors have made ambitious predictions about productivity, automation and future economic value. If those expectations prove difficult to fulfil, investment could weaken and some projects could lose support.
But that wouldn't necessarily mean that all current AI capabilities would disappear, or that the entire field would experience a downturn comparable to the previous winters. Today's AI industry is different from the research environment of the 1970s and the expert-system market of the 1980s. AI is already integrated into a wide range of products, services and professional workflows. Some applications may become commercially successful. Others may fail. Entire business models could change while the underlying technology continues to improve.
That is why I would be cautious about declaring another AI winter simply because a company struggles, a model disappoints or investors become less enthusiastic. A genuine AI winter would involve a broader, sustained loss of confidence and support across significant parts of the field.
As of 2026, whether the current AI boom will eventually lead to another AI winter remains uncertain. The historical parallels are useful.
They just aren't a reliable weather forecast.
AI Can Produce More. But Who Decides What Matters?
The history of artificial intelligence isn't simply a story of excitement followed by failure. Early researchers developed ideas that influenced entire fields of computer science. Expert systems provided useful applications. Machine learning kept advancing even when AI was no longer particularly fashionable. At the same time, the setbacks mattered. Research programmes lost funding, businesses failed, and scientists had to adapt their work to changing priorities. For the people involved, an AI winter was more than an entertaining metaphor.
What the history does show is how readily we turn individual successes into predictions about entire categories of work. A machine solves a difficult problem. We begin imagining all the other problems it might solve. And eventually, someone puts a date on the prediction.
Optimism is very good at turning possibilities into deadlines.
The same thing happens on a much smaller scale in everyday AI work. An AI system can suggest a perfectly reasonable paragraph, improve a sentence or produce an outline. Sometimes the suggestion is genuinely helpful. Sometimes it's technically excellent and completely wrong for what I wanted to say. The problem isn't always the quality of the output. It's whether the output belongs in the work at all.
That is the question behind my Human Decision Map: not how much AI contributed, but which decisions the human actually made.
Are we evaluating what the technology can reliably do? Or are we already judging it against what we've decided it ought to become?
The Forecast Was Always Ours
I don't know whether artificial intelligence will experience another winter. Perhaps some current expectations will prove wildly optimistic. Perhaps other developments will surprise us. Some applications will probably become ordinary parts of everyday life, while others disappear despite having attracted considerable attention.
That seems a fairly reasonable prediction. Although the history of AI suggests I should probably be careful about making predictions.
Meanwhile, we get used to the extraordinary very quickly. A machine can translate languages, generate images, explain complex subjects and write entire articles, and we are already asking why it can't do everything else.
Perhaps that's simply part of being human. We imagine possibilities, develop expectations and become impatient when reality takes longer than planned. Without that impatience, some technological progress might never happen. With too much of it, we risk confusing genuine limitations with disappointment over promises that were never realistic.
Artificial intelligence has experienced two major winters. Both involved real technological and economic difficulties. But both also tell us something about the people evaluating the technology.
The machines didn't write the forecasts.
We did.
This article is also available in German.
Frequently Asked Questions About AI Winters
What caused the first AI winter?
The first AI winter developed during the 1970s as early artificial intelligence systems encountered technical limitations and failed to meet ambitious expectations. Critical assessments, including the 1973 Lighthill Report, contributed to reduced research support. Earlier disappointments in machine translation also influenced funding decisions.
When was the first AI winter?
The first AI winter is commonly dated from approximately 1974 to 1980. These dates are approximate because funding reductions and changes in research activity occurred at different times across countries and institutions.
What caused the second AI winter?
The second AI winter began in the late 1980s when commercial enthusiasm for expert systems and specialised AI hardware declined. Expert systems were expensive to develop and maintain, while cheaper general-purpose computers undermined the market for specialised Lisp machines.
How many AI winters have there been?
Historians generally identify two major AI winters: one during the 1970s and another beginning in the late 1980s. The precise dates and boundaries vary between historical accounts.
Is another AI winter coming?
Another AI winter is possible, but it cannot be reliably predicted from historical patterns alone. A broad decline in AI investment and research support would be more significant evidence than individual product failures or changing market valuations.
What is the difference between an AI winter and the AI hype cycle?
An AI winter describes a historical period of reduced confidence, funding and investment in artificial intelligence. The Gartner Hype Cycle is a conceptual model describing how expectations around emerging technologies may rise, decline and eventually become more realistic. The two ideas are related, but they are not interchangeable.
Sources and Further Reading
Dartmouth College: Artificial Intelligence Coined at Dartmouth
ALPAC (1966). Language and Machines: Computers in Translation and Linguistics. National Academy of Sciences, National Research Council.
Minsky, M., & Papert, S. (1969). Perceptrons: An Introduction to Computational Geometry. MIT Press.
Lighthill, J. (1973). Artificial Intelligence: A General Survey. Science Research Council.
Computer History Museum: Timeline of Computer History: AI & Robotics
Gartner: Hype Cycle Research Methodology
About the Author
Carsten Jan Weichelt is a writer and content strategist based in Barcelona. He works with artificial intelligence in content research, writing and editing, and explores how AI changes creative work, professional identity and human decision-making. His projects include the Human Decision Map and the developing W.E.I.C.H.E.L.T. Method, both examining how people can work with AI while retaining creative direction and editorial responsibility. He also writes fiction and develops the satirical storyworld Alma's World. More about Carsten.