Sixty years of false dawns: how AI got here

In July 1958, the New York Times reported on a machine built for the US Navy. The embryo of an electronic computer, the paper said, was expected to walk, talk, see, write, reproduce itself and be conscious of its existence. The machine was the perceptron, built by the psychologist Frank Rosenblatt. It could learn to tell a mark on the left of a card from a mark on the right.

That gap between the headline and the hardware is the whole history of artificial intelligence in miniature. For sixty years, the field promised minds and delivered tools. And then, quite suddenly, the tools started to look like minds.

A question before a field

The story usually starts with Alan Turing. In 1950 he published Computing Machinery and Intelligence, which opens with a blunt question: can machines think? Turing found the question too vague to answer, so he replaced it with a game. If a judge conversing by text cannot tell a machine from a person, he argued, we have no good reason to deny the machine intelligence.

Turing was also the first to say out loud where this might lead. In a talk given around 1951, he suggested that once machine thinking had started, it would not take long to outstrip our own feeble powers, and that at some stage we should expect the machines to take control. He said it calmly, as a scientist stating a consequence. Almost nobody listened.

Five years later, in the summer of 1956, a small group of researchers met at Dartmouth College. The proposal for the workshop, written by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, coined the term “artificial intelligence”. It also contained one of the most optimistic sentences in the history of science: the organisers believed a significant advance could be made on language, abstraction and self-improvement if a carefully selected group worked on it together for a summer.

The first winter

The early results were genuinely exciting. Programs proved theorems in logic, played checkers well enough to beat their authors, and solved algebra problems from textbooks. Researchers made bold forecasts, and funding followed.

The trouble was that everything worked in the lab and nothing survived contact with the real world. Translating Russian into English, a priority for Cold War governments, turned out to be far harder than parsing a sentence’s grammar. Meaning depended on context, and context depended on knowing a great deal about the world.

Neural networks took a separate blow. In 1969, Minsky and Seymour Papert published Perceptrons, a rigorous book showing what a single-layer perceptron could not compute. The book was careful about its scope, but its influence was not. Interest and money drained away from the whole approach.

In 1973, a report written for the British government by the mathematician James Lighthill concluded that AI had failed to deliver on its grand promises. Funding was cut in the UK, and the mood spread. Researchers later gave the period a name: the first AI winter.

Experts, then another winter

The field recovered in the 1980s with a humbler idea. If machines could not learn about the world, perhaps experts could simply write the knowledge down. Expert systems encoded thousands of hand-written rules: if the patient has this symptom and that test result, consider this infection. Some of them worked well enough to be sold, and companies spent heavily.

They were also brittle. Every rule had exceptions, every exception needed more rules, and the systems broke in ways their builders could not predict. By the end of the decade the market had collapsed, and a second winter set in.

Something important happened quietly during those years, though. In 1986, David Rumelhart, Geoffrey Hinton and Ronald Williams published a paper in Nature popularising backpropagation, a method for training neural networks with several layers. It solved much of the problem Minsky and Papert had pointed to. There was simply not enough data or computing power to show what it could really do.

Milestones that proved less than they seemed

Year	Event	What it showed
1950	Turing’s imitation game	Intelligence could be framed as a testable question
1956	Dartmouth workshop	A field was born, with wildly optimistic timelines
1958	Rosenblatt’s perceptron	Machines could learn simple patterns from examples
1973	Lighthill report	Promises had run far ahead of results
1986	Backpropagation paper	Deep networks could, in principle, be trained
1997	Deep Blue beats Kasparov	Brute-force search could master a narrow domain

Deep Blue’s victory over the world chess champion Garry Kasparov in 1997 was the most public moment of all. It was also, in a sense, a dead end. The machine evaluated around 200 million positions per second, using rules tuned by grandmasters. It knew nothing except chess, and it learned nothing by itself.

The pattern behind the failures

Looking back, the false dawns share a single cause. Every approach tried to hand intelligence to the machine: as logic, as rules, as expert knowledge. And every time, the world turned out to be too large and too messy to write down.

The roboticist Hans Moravec noticed a strange corollary in 1988. The things humans find hard, like chess or calculus, were easy for computers. The things a toddler does without thinking, like recognising a face or picking up a cup, were almost impossible. What we call intelligence, it turned out, was mostly the part we never have to think about.

Sixty years of failure taught the field one thing: intelligence cannot be written down. It has to be learned.

The answer had been sitting in Rosenblatt’s perceptron all along. Machines would need to learn from examples, at enormous scale. In 2012, a graduate student’s program entered an image recognition contest, and the winters ended for good. That story, and the lesson it forced on the whole field, comes next in this series.