Other attributes
An AI winter refers to a period of reduced interest in artificial intelligence (AI). The reduced interest can be a reduced public, academic, or enterprise interest in artificial intelligence and results in a reduction of funding for artificial intelligence projects and companies and a commensurate reduction in research and development into AI and underlying technologies. "Winters" are not exclusive to AI, as other technology-heavy and funding-based industries have experienced these periods. Cryptocurrency and blockchain have experienced their own winters and "summers"—periods of intense public interest and funding.
After a summer period comes a winter, which results in a cooling period of public interest and funding and a cooling of research and development into AI. A winter typically occurs due to the overhyping or overpromising of AI technology that occurs during the summer period, when companies may make AI an important part of their growth strategy and the technology does not live up to the expectations of the companies, investors, and general public.
This period of disillusionment tends to see an increase in the realistic scrutiny of the industry. Some researchers have suggested this scrutiny is not a bad thing, as it allows the headier and more aggrandized ideas to be fleshed out and brought to fruition. This also allows regulators and the general public time to catch up with the developments in the space and proper legislation and regulation to be adopted to deal with the possible disruption caused by AI developments.
The AI summer offers some insight into why AI winters happen. In an AI summer, the development of and interest in AI is booming, and expectations for the technologies' capabilities and the enabling technological breakthroughs are high while funding is free-flowing. This is a metaphorical peak of interest in the technology and its promised applications. Many of the significant advancements in AI happen during an AI summer, with the advancements typically met with breathless anticipation that continues to build the hype of the summer.
The natural ebb and flow of the hype cycle is a contributor to AI winters—some suggesting the biggest contributor, as it sets the expectations for the technology. Often those expectations are based on promises of AI researchers and organizations and the technology they are developing. But in the case where those promises are not met by the technology or not met in the way the public imagined they would be, interest is lost. A winter then sets in where the loss of interest leads to a loss of funding and a slowing—if not halting—of development.
Gartner's hype cycle
AI winters are strongly correlated with Gartner's hype cycle, which attempts to explain that new technology triggers unrealistic expectations at first, leading to an increase in research and investigation into the technology. However, this leads to increasing disillusionment, and the research and general interest see a drastic decrease. However, during the winter, often slow and steady technology improvements can lead to a new cycle of hype, which, in turn, can lead to another winter.
Often the development of AI comes through universities or similar research-focused institutions. This significantly impacts AI developers and development, as the researchers, and at times their institutions, are not able to control funding changes. If budget cuts are made, or funding to AI is cut off, the researchers have no choice but to stop the development of and research into AI.
The use of AI has raised ethical concerns, including the potential for bias, discrimination, privacy, security, transparency, and the accountability (or lack thereof) of AI systems. Many countries and organizations may develop guidelines and frameworks for the use of AI, which can hamper the technology's development and reduce the interest in the technology, helping to contribute to an AI winter. It can also be important for ensuring that AI can be trusted by users.
The economic status of a country will also have a major influence on the development of AI. For example, during times of war, recessions, and pandemics, resources will be reallocated to other projects and research areas, often directed away from AI, unless AI developers have a proven use case for the given economic factor.
AI is often marketed toward and developed for enterprises, and AI use in these sectors has been impacted by AI winters. And the challenges enterprise use cases include and can face can trigger an AI winter. Internal concerns, such as operational costs, or external concerns, such as a lack of adoption, can drive enterprises to pivot to other technology offerings or internally sour on AI and therefore not market products as AI, leading to a lack of interest in AI and a resulting AI winter.
Although a technical challenge, a lack of data will impact an enterprise before it will impact public interest. AI systems and models often need large amounts of data to train and function effectively, with data being a critical component of AI and machine learning. Therefore, if organizations do not have access to high-quality, relevant data or cannot manage and process it effectively, or if the cost of storing and processing data is prohibitive, the enterprise may not be able to realize the full potential of AI and will sour on the technology and its possibilities.
As with any technology, to achieve success, an organization needs to ensure it can access people with the necessary skills and expertise. This includes data scientists and data engineers who can help an organization build and deploy machine learning models, domain experts who have deep understanding of AI applications, and the necessary infrastructure and tools to develop and deploy those tools. An organization with these in place can fully realize the potential of AI, but those without may struggle to get a return on their investment in AI.
AI systems often need to be integrated with existing enterprise systems, or the existing systems have to be overhauled and updated, which can come with challenges and resource costs. These challenges and costs can come from the AI system not being compatible with current infrastructures and the complexity of data integration as the AI system requires data from multiple sources within an organization, which may never have connected their various data pipelines and sources before.
This cost can cause an organization to question the investment into AI, and in the case where they may lack the necessary systems and skilled workers, the return on that cost may take longer to be realized, creating a greater challenge for the adoption of AI systems in an organization.
The artificial intelligence industry dates back to the early 1950s. During this period, Alan Turing first broached the concept of whether machines could be capable of human-like thought in his paper "Computing Machinery and Intelligence." Around the same period of time, the field of computer science was also beginning. During this period, Turing created the imitation game known as the Turing Test. In 1954, one of the first experiments in machine translation occurred, which included a 250-word dictionary for translation combined with syntactical analysis with demonstrations in translations between English and Russian. This garnered the machine a lot of hype and funding.
Arthur Samuel playing checkers against a computer.
In 1955, Arthur Samuel used a learning algorithm he called "temporal-difference learning," based on a combination of tree search with heuristics and learned weights to develop a program that could play checkers well. The learning algorithm included the learned weights, which were adjusted using the "error" between the score initially calculated and the score after the search was completed. During this period, AI research began to gain increasing amounts of funding from U.S. Defense Establishments, originally ONR and ARPA, later called DARPA. The hope was that the technologies would be useful for the military and garner further enthusiasm and optimism for AI, especially in the case of machine translation.
An early implementation of the perceptron machine.
In 1956, the Dartmouth Summer Project was created, and the term artificial intelligence was coined. Researchers were from various fields and developed many ideas, papers, and concepts, but also saw an initial wave of disappointment as the workshop showed the experts and researchers that AI may be harder than initially thought. But by 1957, Frank Rosenblatt invented perceptrons, a type of neural network where binary neural units are connected via adjustable weights and are inspired by the work of the 1940s. This led Rosenblatt to a crude replication of the neurons in the brain, including various layouts and learning algorithms, to try different layouts of the input and output and data flows. Computers of the time were too slow to run the perceptron, and Rosenblatt built a special-purpose machine with adjustable resistors controlled by motors and proved capable of learning to classify different images of shapes or letters.
After increases in funding and enthusiasm for the potential for AI in the 1950s and early 1960s, progress stalled. A lot of the interest was in the possibility for machine translation. However, during the decade, some experts suggested machine translation was not feasible and the computers needed too much information for a correct translation, which led the Automatic Language Processing Advisory Committee to conclude in 1964 that there was no immediate prospect of useful machine translation. Further, corresponding reports led some to believe AI to be similar to alchemy, which led to diminishing returns, disenchantment, and pessimism. This has largely been considered the beginning of the first AI winter, although some consider it an AI winter in itself that was not followed by any real AI summer before the events of what is largely considered the first AI winter.
Minsky and Papert's Perceptrons.
In 1969, Perceptrons was published. Written by Marvin Minsky and Seymour Papert, the book was a critique of Rosenblatt's perceptrons, and they proved that perceptrons could only be trained to solve linear separable problems. This proved a problem for a group of researchers known as connectionists, who believed AI could be best achieved by mimicking the brain. Minsky and Papert suggested multiple layers could solve the problem, but the algorithm to train such a network would not be devised for another seventeen years and would be known as backpropagation (which would later be discovered to have been previously invented before the publication of Perceptrons).
James Lighthill, author of the influential Lighthill Report.
In 1973, came another major roadblock to the continuing development of AI research and technology. A scholarly report was published, now referred to as the Lighthill report. It was an evaluation of the current state of AI written for the British Science Research Council. James Lighthill, author of the report, came to the conclusion that discoveries in the field had so far failed to make the impact that was promised, with the most disappointing area being machine translation, where few useful results had been achieved despite the sums of money and time spent on the field.
Lighthill believed the failure came in the problem known as combinatorial explosion. This referred to a well-known problem in search spaces, like trees, where the number of nodes increases exponentially the further down the tree. For example, in a game like chess, the possible number of moves increases from 20 at the beginning of the game, to 300 by the second move, to 4,865,609 possibilities by the fifth move.
The report drew a lot of criticism, and a debate on it was even filmed for the BBC but would never air. Much of the debate saw comparisons to other fields of science where results were not fast, nor expected to be fast. But the report had been received, and the United Kingdom's government would cut funding for all but two universities engaged in research in the field in what was the beginning of a wave that would see its impact globally, including a reduction of funding from the agency that would become DARPA.
Beginning in the 1980s, the previous AI winter began to thaw, and a new focus on AI began. This time, the focus was less on academic and military research as the previous summer had been, and instead was focused on creating commercial products. At the same time, large conferences, like the Association for the Advancement of Artificial Intelligence (AAAI), began and experienced a rapid increase in tickets sold. And a renewed interest in the technology was found in general industry and government officials.
But at the heart of the summer is the commercialization of expert systems with AI. These systems were handcrafted by surveying experts and creating "if-then" rule sets. This approach to AI is known as a top-down approach, with many believing that expert knowledge was the best way to create AI. These systems were implemented in fields like financial planning, medical diagnosis, geological exploration, and microelectronic circuit design. The adoption and the increasing interest led to the hype, as publications began to declare that AI had arrived, had built a better brain, and had begun programming human knowledge and experience into a computer.
As the hype built, researchers began to worry about the hype and excitement and the potential for another AI winter. By 1984, at the AAAI conference, scientists began discussing if an upcoming AI winter could be prevented, fearing a loss of funding when unrealistic expectations could not be fulfilled. They would be proven correct, as in the following years, the claims of what AI systems were capable of slowly had to face reality.
John McCarthy, AI researcher and critic.
In 1984, John McCarthy criticized expert systems at the center of the AI hype because they lacked common sense and knowledge about their own limitations. One of these systems criticized by McCarthy was the MYCIN physician-assistance system. He laid out a scenario where a patient has cholera vibrio in his intestines, for which the system prescribed two weeks of tetracycline, which would kill off the bacteria, but the patient would be dead by the time it took effect. Many of the systems developed at the time were too simplistic for the complicated tasks they were attempting to tackle, with engineers not capable of designing rules for each task manually, especially as more and more edge cases occurred.
Jacob Schwarz, Director of Darpa's Information Science and Technology Office from 1986 to 1989, would go on to conclude that AI research had limited success in particular areas, followed by a failure to reach the broader goal at which the initial success seemed to hint. AI funding began to decrease, in line with declining general interest, as expectations continued to fail to be met. Companies that developed AI systems began to close their doors. And AAAI conferences that attracted over 6000 companies in 1986 would decrease to only 2000 by 1991. AI-related articles also saw a decrease beginning in 1987 to reach a low point by 1995. It would not be until the end of the 1990s and the beginning of the 2000s that AI research would begin to generate interest again.
