From Turing to Today: A Short History of Artificial Intelligence

Every institution now fielding questions about AI adoption is really asking a version of the same question: is this the real thing this time, or another cycle we'll quietly forget in five years? The honest answer is that both have been true before. Understanding how AI got here is the fastest way to make a clear-headed decision about where it's going next.

1950: a question, not an answer

The starting point most historians agree on is a 1950 paper by the British mathematician Alan Turing, "Computing Machinery and Intelligence." Turing didn't ask whether machines could think — he thought the question was badly formed — so he replaced it with a practical test: could a machine hold a conversation well enough that a human questioner couldn't tell they were talking to a machine? Seventy-five years later, that test, now known as the Turing Test, is still the reference point every new chatbot gets measured against, informally if not formally.

1956: the field gets a name

The term "artificial intelligence" was coined six years later, at a summer workshop at Dartmouth College in the United States, organised by a young researcher named John McCarthy. The pitch to funders was strikingly confident: the organisers believed that "every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it," and that meaningful progress could be made by a small group working over a single summer. It could not. But the label stuck, and so did the ambition.

The first boom and the first winter

The 1960s and 70s produced real, if narrow, progress: programs that could solve algebra problems, prove logical theorems, and hold limited conversations. Government funding, particularly from the US defence research agency DARPA, flowed generously on the promise that general machine intelligence was close. It wasn't. By the mid-1970s, machines that worked well on toy problems collapsed when faced with the ambiguity of the real world, and funders noticed. The resulting funding collapse became known as the first "AI winter" — a term deliberately borrowed from nuclear winter, to signal how sudden and severe it felt to researchers living through it.

Expert systems, and a second winter

AI revived in the 1980s through "expert systems" — programs that encoded the specific rules a human expert would follow, used in fields like medical diagnosis and mineral prospecting. Businesses invested heavily, including a wave of specialised "AI hardware" companies. But expert systems were brittle: they only knew what had been explicitly programmed into them, they were expensive to maintain, and they couldn't learn from new situations. By the late 1980s, a second, longer winter set in, and "artificial intelligence" became close to a dirty word in serious computer science funding circles for most of the 1990s.

The lesson institutions should take from this isn't that AI fails — it's that AI succeeds unevenly, and the gap between demonstration and dependable deployment is where most of the money gets lost.

The quiet decades: machine learning takes over

What actually kept the field alive through the 1990s and 2000s was less glamorous: statistical machine learning. Instead of hand-coding rules, researchers built systems that learned patterns directly from data — spam filters, fraud-detection models, recommendation engines, and speech recognition all matured during this period, largely without being marketed as "AI" at all. This is the quiet inheritance that today's systems sit on top of, and it's also the part of the story most useful to institutions: the boring, well-tested statistical methods from this era still do a large share of the genuinely reliable work in production systems today, including in tax administration and banking, which we cover in separate pieces in this series.

2012: the deep learning breakthrough

The modern era usually gets dated to 2012, when a neural network called AlexNet dramatically outperformed every competing approach in a major image-recognition competition. What changed wasn't a new idea — neural networks had existed since the 1950s — but the arrival of enough data and, critically, enough cheap computing power (in this case, graphics chips originally built for video games) to train much larger networks than had ever been practical before. This combination of old ideas and new hardware is what actually restarted the field, and it's why "AI" and "GPU" have been discussed in the same breath ever since.

2017: the architecture behind today's tools

A 2017 research paper from Google, titled simply "Attention Is All You Need," introduced an architecture called the transformer. It turned out to be extraordinarily good at handling language, and it is the technical foundation underneath essentially every large language model in use today, including the tools now reaching Nigerian ministries, banks, and universities. The five years between that paper and the public release of consumer chatbots in 2022 were spent scaling that architecture up, an unglamorous but decisive engineering effort.

What seventy years actually teaches an institution

The pattern across every wave — the 1960s, the 1980s, and now — has been the same: genuine capability arrives, is oversold beyond what it can actually do, and the correction that follows punishes institutions that committed hardest during the hype rather than the ones that adopted carefully. Three practical conclusions follow directly from that history:

• Distinguish the demonstration from the deployment. A convincing demo is the easiest part of any AI project. The 1980s expert systems boom collapsed almost entirely on this gap.
• Favour the boring, well-tested layer first. Statistical methods from the "quiet decades" — pattern detection, anomaly scoring, classification — are lower-risk and already proven in exactly the use cases most relevant to public institutions.
• Build governance alongside adoption, not after it. Every previous AI winter was, in part, a trust collapse. Institutions that can show how a system reached a decision will survive scrutiny that ones relying on an unexplainable "black box" will not.

This is not the first time the world has been told AI changes everything. It's the fourth. The institutions that come out ahead this time will be the ones that treat that history as useful data, not as a reason for either blind enthusiasm or blanket refusal.