WHAT IS AI, REALLY?

Almost everything started with one paper. In 1950, British mathematician Alan Turing — already famous for helping crack the Enigma code during World War II — published a plainly titled paper in the journal Mind: Computing Machinery and Intelligence. He felt his own starting question, "can a machine think?", was itself poorly framed: the words "think" and "machine" were too vague to be answered seriously.

So Turing pulled off a move that's now famous: instead of settling the question philosophically, he replaced it with a concrete test. In what he called the "imitation game" — later renamed the Turing test — a human interrogator holds blind, written conversations with another human and with a machine. If the interrogator can't reliably tell which one is the computer, the machine is considered to have passed the test.

Perhaps the most telling detail is that Turing didn't write this piece as a dry logical treatise — unlike his more technical mathematical papers. According to someone who knew him, he wrote certain passages with visible enjoyment, almost playfully (The Essential Turing). The paper wasn't meant so much as a rigorous philosophical contribution as a piece of persuasion: Turing wanted to convince his contemporaries that they should take seriously the idea that a machine might one day display behavior that, in a human, would be called intelligent. It would take another six years for that idea to get a name.

We often picture a grand, famous conference, but the Dartmouth conference was really more of a workshop among peers.

In the summer of 1956, on the quiet campus of Dartmouth College in New Hampshire, about twenty researchers gathered for a workshop meant to last eight weeks. The event was organized by four men: John McCarthy, a young assistant professor of mathematics at Dartmouth; Marvin Minsky, then a young researcher at Harvard; Nathaniel Rochester, director of research at IBM; and Claude Shannon, the father of information theory, then at Bell Labs. Among the participants were also Allen Newell and Herbert Simon, who came to present their Logic Theorist program — capable of proving formal logic theorems, a first.

It was in the funding proposal written the previous year by the four organizers that the phrase "artificial intelligence" appeared for the first time. The term wasn't chosen lightly: up until then, people talked instead about "cybernetics" or "automata theory." McCarthy deliberately chose a new label to mark a break with the past and unite a still-fragmented research field.

The mood of that famous summer remains poorly documented — two participants, Ray Solomonoff and Trenchard More, had the good idea of keeping a daily journal, which today makes them the most precise source on what actually happened at the gathering. What we do know is that optimism ran high: the researchers sincerely believed that within a few decades, or even less, most of the problems of machine intelligence would be solved. Arthur Samuel, one of the participants, would that same year develop a checkers-playing program capable of improving by playing against itself — and would coin the term "machine learning" in 1959.

This confidence would carry the research forward for nearly twenty years, backed by generous funding, notably from DARPA, the Pentagon's research agency. But the promises made by the Dartmouth pioneers — a truly intelligent machine within twenty years — were about to hit a somewhat unforeseen wall.

"AI winter" refers to a period when funding dries up and enthusiasm collapses abruptly, after years of unfulfilled promises. AI's history has seen two of them, and in both cases the root cause was the same: machines were being asked to reason about the real world without the data or the computing power needed to do so.

A first warning sign came as early as 1966, with a US report (ALPAC) finding that machine translation — a field heavily funded since the start of the Cold War in hopes of instantly translating Russian — simply wasn't working satisfactorily. Some budget cuts followed.

The real chill came in 1973, with a report commissioned by British scientific authorities from mathematician James Lighthill, who himself had never worked in AI. His verdict was harsh: researchers, he argued, had massively oversold their results. Lighthill put his finger on a problem that persists today under a different name: "combinatorial explosion." A program able to solve a toy problem in the lab — a game, a simplified logic puzzle — sees its computing needs grow exponentially the moment it's applied to a real-world-sized problem. And the computers of the time, with their few kilobytes of memory and processors capable of barely a few tens of thousands of operations per second, simply couldn't keep up. Nor was there any large-scale digitized database for a program to learn from — most knowledge had to be hand-coded, line by line, by researchers.

The Lighthill report compounded another blow: as early as 1969, the book Perceptrons, written by Marvin Minsky himself together with Seymour Papert, had demonstrated the mathematical limits of early neural networks — the famous perceptrons couldn't even learn a logical function as simple as exclusive OR. It would take until the mid-1980s, with the rediscovery of the backpropagation algorithm, to get past this theoretical roadblock. But even that new tool would remain impractical for a long time: training a multi-layer neural network required computing resources that simply wouldn't exist before the 2010s. The result: starting in 1974, DARPA cut most of its fundamental research funding, the British government did the same, and AI entered a first winter that would last until the early 1980s.

Between the two AI winters, there was nonetheless a very real golden age, driven by a family of programs known as "expert systems." The idea was simple: instead of aiming for general intelligence, you encode a human specialist's decision rules for one very specific domain into software, in the form of "if this condition, then this conclusion" rules. The program doesn't literally understand anything, but it methodically applies hundreds, sometimes thousands, of rules. Expert systems would be the first AI programs to actually make money (R1 being a prime example), briefly pulling the field out of its winter.

Expert systems briefly kick-started the field again in the 1980s, with a specialized-hardware market — the "Lisp machines" — worth several hundred million dollars a year by the middle of the decade. But the business model collapsed almost overnight in 1987: cheaper, general-purpose workstations from companies like Sun Microsystems caught up with, then surpassed, the performance of the specialized machines. The market for dedicated AI hardware disappeared within months, dragging down many companies in the sector along with it.

The real problem — even more structural than the collapse of one market segment — was a matter of infrastructure: computing power and data, whose limiting effect is hard to fully appreciate today.

On the compute side, first: a personal computer in the early 1980s ran on a few tens or hundreds of kilobytes of RAM. To put that in terms everyone can relate to: the onboard computer that guided Apollo 11 to the Moon in 1969 had only 4 KB of RAM, and the very first IBM PC, released in 1981, had barely eight times that. Storage, meanwhile, cost a fortune: in 1980, a 5-megabyte hard drive cost around $1,500 — that's $300 per megabyte, a price that would fall to less than a thousandth of a cent per megabyte by 2008. In other words, it would have been financially absurd, and technically impossible, to even consider storing and processing the volumes of text or images that training an AI system requires today.

On the data side, next, things were no better: there simply was no public internet to draw on. ARPANET, the military-and-academic ancestor of the web, connected only a few hundred computers in the early 1980s, and would reach barely more than 100,000 — all research labs — by the very end of the decade. The World Wide Web itself wasn't invented until 1989-1991, at CERN, and wouldn't really go mainstream until the mid-1990s, with the arrival of the first consumer web browsers. So throughout the second AI winter, there existed, on no accessible network anywhere, neither a large-scale corpus of digitized text nor a massive database that a program could draw on to teach itself.

This double constraint — no machines powerful enough, and no data to feed them — explains why the systems of the time, starting with the expert systems mentioned above, could only work by hand-encoding a human specialist's knowledge rule by rule: technically, there was simply no other option.

Before the explosion of generative AI, there was a pivotal decade, carried by two twin promises: the self-driving car and computer vision.

It all started with a spectacular failure. In 2004, DARPA organized its first "Grand Challenge": a self-driving vehicle race across 240 kilometers of California desert, between Barstow and Primm, Nevada. The result was brutal. Not a single vehicle finished the course: the best contender — a convoy of buggies and 4x4s bristling with sensors, straight out of a post-apocalyptic movie — covered barely 12 kilometers before getting stuck on an embankment. The press at the time called it a resounding fiasco. But just one year later, in 2005, several vehicles completed the full circuit — including Stanley, the vehicle fielded by the Stanford team led by Sebastian Thrun. That episode would ripple across the entire industry: Thrun went on to found Google's self-driving car project, which became Waymo, while former Carnegie Mellon participants went on to launch Cruise and Zoox.

At the same time, a second breakthrough happened in 2012, in a field that seemed very different at first glance: image recognition. Every year, the ImageNet competition evaluated the best image-classification algorithms on a dataset of more than a million photos spread across a thousand categories. Progress there had, until then, been slow and incremental. In 2012, a team from the University of Toronto — Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton — submitted a neural network called AlexNet, trained on graphics cards (GPUs) rather than standard processors. The result stunned the entire community: a 15.3% error rate, versus 26.2% for the runner-up — a nearly 11-point leap that nobody saw coming. Within months, Google, Facebook, Microsoft, and Baidu launched massive investments in deep learning. Three years later, in 2015, an even deeper network (ResNet) would drop below a 4% error rate — a level now better than a human's on that specific task.

It's this combination — cars driving themselves, a machine "seeing" better than a human on certain tasks — that fueled a considerable wave of hope between 2015 and 2022. People promised widespread robotaxi fleets within a few years, the scheduled end of the driver's license, total automation of transportation. The reality, seven years later, is more nuanced: computer vision has indeed gone mainstream — smartphone facial recognition, AI-assisted medical imaging, industrial quality control — but the fully self-driving car, in 2022 as today, remains confined to a handful of pilot cities rather than rolled out on country roads or in poor weather conditions. Much like the Dartmouth promises in their time, the scale of the change was real, but slower and more partial than the enthusiasm of the era suggested.

What computer vision had failed to do — make AI tangible for the general public — generative AI would accomplish in a matter of weeks. On November 30, 2022, OpenAI launched ChatGPT, almost as a low-key technical demo. The service reached one million users in five days, then 100 million monthly active users by January 2023 — an adoption rate no consumer product had ever seen before, far outpacing services like Instagram or Netflix in reaching that milestone. ChatGPT would go on to cross the billion-monthly-user mark in just over three years, beating the previous record held by Google Maps, which had taken nearly five years to get there.

This success came alongside a parallel wave in image generation: Midjourney, DALL-E, and then Stable Diffusion popularized, also starting in 2022, the ability to produce an image from a simple natural-language description. For the first time, someone with no technical skills at all could literally "chat" with an AI or ask it for an illustration — a far more immediate experience than an invisible recommendation algorithm or a smartphone camera detecting a face.

It was precisely this visibility that, in the years that followed, created a conflation many industry observers have pointed out: in the mind of the general public — and sometimes even of certain professionals — "artificial intelligence" and "generative AI" became synonyms. Yet generative AI, however impressive, is only one branch of a much broader field that also includes recommendation systems, the computer vision mentioned above, predictive algorithms used in finance or logistics, and the expert systems of the past. This confusion isn't harmless: it feeds both excessive optimism — the idea that generative AI is some magic, universal solution — and disproportionate fears, each one feeding the other.

Whatever judgment one makes about this conflation, the transformation that comes with it is very tangible indeed — at the level of individuals as well as businesses.

On the individual side, perhaps the most telling example is Stack Overflow, the world's largest developer help community since 2008. Within the first few months after ChatGPT's launch, its traffic dropped by more than 10% a month; later academic studies would describe a decline that only deepened over time. The old reflex of asking a coding question on a public forum, for the benefit of thousands of other developers who might one day stumble on the same answer, has largely given way to a private, instant conversation with an AI assistant. An entire model of collectively producing knowledge has been upended by this shift in habits, in just a few months.

On the business side, adoption is moving fast but remains uneven: in Europe, about one in five companies with ten or more employees reported using AI technology in 2025, with that rate topping half among the largest organizations. A new phenomenon illustrates just how fast things are moving: "Bring Your Own AI," where employees use their own generative AI tools at work without waiting for an official internal policy — an informal practice that, according to several recent surveys, involves a majority of employees worldwide.

From the philosophical question Turing asked in 1950 to a billion people chatting with an AI every month in 2026, the thread running through this history is, in the end, fairly consistent: waves of hope driven by real technical breakthroughs, followed by backlash once reality — a lack of data, computing power, or simply robustness — catches up with the promises. Nothing guarantees the current wave will escape this cycle. But unlike previous winters, it has already, and for the first time, changed the habits of several hundred million people in just a few years — something that, historically, had never happened at this scale or this speed.