A Brief History of AI

Our relationship with Artificial Intelligence has a long history, much longer than one might think based only upon recent reporting. In the broadest possible terms, AI’s roots stem from myths and legends of non-human intelligence found in cultures all over the world. Many of these cultural memories influenced early scientists who, in parallel traditions, developed formal mathematics and automata, as well as several early analog “thinking machines”. These devices, some developed over a thousand years ago, may have been rudimentary but operated on the same principles of modern computing. Accordingly, computing developed from these same roots, and indeed the distinction between AI and computing is a relatively recent phenomenon. The modern concept of AI, and indeed the English term “Artificial Intelligence” itself, is typically credited to the 1956 Dartmouth College Summer AI conference. This conference sought to formalize a discipline to eventually create an artificial equivalent of a human mind, capable of solving complex problems across many domains. We now call this highly general type of AI Artificial General Intelligence, or AGI.

However, the subsequent trajectory of AI research has been highly non-linear, experiencing several periods of growth followed by so-called “AI Winters” where science, investment, and regulatory confidence in the technology waned for a time. These setbacks are largely attributable to actors’ over-promising the current or expected near-term capabilities of AI followed by the actual capabilities falling short of these expectations. 1950s-1970s AI models were effectively toys — interesting proofs of concept but incapable of solving practical problems, which eventually resulted in a large decrease of funding for their development. In the 1980s, expert systems focused on decision support for highly specific business and logistics problems achieved early success and became widely popularized. This prompted a refocus of most AI research onto tools tailored for specific problems instead of AGI. The early success of these systems led to their over-adoption and an eventual decline in confidence as over-leveraged companies began to fail in the early 1990s, prompting the second AI winter.

It is worth noting that the first two AI winters were caused by different entities. The first was the result of a loss of US government confidence, while the second was the result of industry over-investment creating a bubble. As we shall see, the ebb and flow of AI became much more complex after the turn of the century, and although several individual technologies have risen and fallen, we have not yet seen a third AI winter and the mass loss of confidence that it entails.

By the mid 2000s the large increase in compute power, storage capacity, and the popularization of the internet led to the success of several AI technologies. However, lingering public and business distrust of the AI label caused scientists to use alternative category names for their technologies such as data mining, machine learning, data science and “big data”. Like their expert system forebears, the majority of these technologies were narrowly focused on particular problems. This expansion of compute capacity also led to the prominence of Artificial Neural Networks (ANNs), a branch of AI that had existed for over 50 years but only became practical by the early 2010s. The surprising generalization ability of ANNs and their broad applicability to diverse problems such as natural language, computer vision, and audio processing eventually led to a renewed interest in pursuing Artificial General Intelligence. Perhaps, many thought, it was finally possible to build tools that are broadly applicable rather than purpose-built for a narrow collection of problems.

In the mid-2010s there was a wide belief that AI technologies of the time were ready to solve major problems. Perhaps most famously, business leaders promised that fully autonomous vehicles were on the cusp of becoming commonplace. The intervening years have shown that, while some viable assistive technologies for driving achieved widespread adoption, fully autonomous driving was not yet practical and is still an unsolved problem in 2025. In prior decades, these conditions might have created another AI winter as investors abandoned the technology. However, it seems that the field has penetrated enough of the academic and business world that high-profile shortfalls are no longer sufficient to cause a full-blown AI winter. Instead there appears to be a cycle of popularity churn as new models rise to replace those that fall short, capturing the attention span of the industry before it can form a judgement of the whole field. For autonomous vehicle routing this has meant that, by mid 2025, there has been enough sustained investment that some businesses have deployed limited autonomous vehicle services to moderate success.

As this autonomous vehicle routing drama played out, a different branch of AI technologies was steadily growing to an unprecedented boom. Indeed, our current moment in AI’s history is overwhelmingly focused upon generative models - tools that are able to produce a human-comprehensible response to (usually) natural language prompts provided by a non-expert user. For example, Large Language Models (LLMs) are sophisticated neural networks that can produce readable, if not always accurate, plain language responses to queries on many topics. Similarly, diffusion and related models can produce visual images or even video that attempt to realize a text-based description. LLMs are particularly in vogue due to their perceived ability to augment or even replace a lot of human knowledge work. Although there is a lot of hype that LLMs have or will shortly achieve artificial general intelligence, it is important to learn from the history of AI and not overestimate their abilities.

Unfortunately, in mid-2025 it is clear that many business and government leaders believe that LLMs and related AI systems will reduce or eliminate much of the need for human workforce in several sectors, a phenomenon that is most visible in the arts but is also present in point-of-service (e.g., customer service chatbots), IT infrastructure, point-of-sale, as well as text production sectors such as software development and professional writing. Furthermore, stakeholders believe so strongly in the future dominance of these technologies that layoffs, hiring freezes, and a widespread reluctance to hire early career staff in many fields has become commonplace. 

What can we learn from the history of AI, and how can its lessons help us navigate our present? It is important to keep in mind that LLMs and diffusion models are only a subset of AI and are not the only route towards solving the problems that AI hopes to solve. Furthermore, AI is not, and likely never will be, a drop-in replacement for a human mind. We are beginning to see some recognition that the current state of the technology is not AGI as advertised. Research is beginning to suggest that LLM use not only does not drastically improve productivity but may in fact reduce critical thinking and problem solving abilities. Moreover, pop science publications and commentators have begun to notice that prominent AI companies are consistently promising exponential improvement six months in the future, promises upon which they never seem to deliver. The purpose of this article is not to predict an incoming AI winter — indeed, the global economy has invested so heavily in AI that such a thing would cause a devastating recession that the ultra wealthy will fight to avoid. It is instead a reminder that we’ve been in similar situations before, where a technology with interesting applications has been oversold by a profit-seeking investor class. This status quo is unlikely to persist, and eventually new technologies and ideas will arise to supplant those that are currently in vogue.