This book reveals how an AI winter, caused by poor machine translation, nearly killed the technology

We are currently in the throes of an AI renaissance – but it hasn’t always been that way. AI hasn’t advanced in a single, upward exponential curve that those tech companies currently at the forefront of the generative AI revolution love to suggest. Instead, it’s a stuttering story of two steps forward, one step back.

To understand why, we have to go back to the Cold War.

For the United States, artificial intelligence held huge promise because of its ability to help the country maintain a scientific and military edge over its big ideological and geopolitical rival, the Soviet Union. The US government saw the potential of AI to provide near-simultaneous translation of diplomatic and spying messages sent in Russian that they were tapping. Similarly, the US government, in a race for scientific supremacy, would also want to know what Soviet scientists were publishing in academic papers. Speedier translations of technical reports would help them achieve that. In short, AI could aid the US in peeking behind the inscrutable Iron Curtain.

Machine translation, then (remember AI as a term would not be developed until the Dartmouth summer conference of 1956), was put at the forefront of development. And Georgetown University and tech titan IBM were in the vanguard. The two entities worked together to try and develop a machine translation tool throughout the early 1950s, with some success.

A New York Times article, published on January 8, 1954, indicated just how far the Georgetown–IBM partnership had come in a short space of time. A day earlier, the partners in the project – Professor Leon Dostert and Dr Paul Garvin of Georgetown, and Dr Cuthbert C Hurd, of IBM’s division of applied science – had demonstrated the fruits of their partnership in the skyscraper IBM owned at 590 Madison Avenue in New York.

In front of the world’s media, Dostert, Garvin and Hurd explained how a program they had developed on the IBM Type 701 Electronic Data Processing Machine – only 12 of which had been sold, all to military, commercial and university laboratories since its release the previous April – could translate 250 words in Russian into English almost instantaneously. But they didn’t just explain; they demonstrated it.

A female typist (typically of the time, she goes unnamed in the contemporary reports, despite her importance in the history of the technology) typed a sentence in Russian into the Type 701 computer: “Mi pyeryedayem mislyi posryedstvom ryechi.”

Within a flash, the computer spits back a sentence in English via a printer – the translation of the phrase: “We transmit thoughts by means of speech.” Those 39 characters took around one second to print.

Emboldened, the typist input another sentence in Russian: “Vyelyichyina ugla opryedyelyayatsya otnoshyenyiyem dlyini dugi k radyiusu.” Again, the computer returned with a near-simultaneous translation: “Magnitude of angle is determined by the relation of length of arc to radius.”

More sentences and phrases were tested out, including ones covering politics, communications and military affairs. “The sentences were turned into good English without human intervention,” the New York Times reported. The way the IBM Type 701 machine – technically a calculator, rather than a computer – did this was by following six broad rules for Russian syntax and grammar it had been “taught” through a series of punch card commands. It then interrogated its memory of 250 Russian words to try and discern meaning from the sentence in English. Sometimes it would have to add in words to make sense in English; other times, it had to remove extraneous ones. And when a Russian word had multiple possible English translations, the computer was commanded to pick the one that best fit the context.

IBM’s Hurd was exuberant about the experiment and its potential. While this trial had been tested on just 250 words, the Type 701 calculator had enough memory to store one million five-letter words. The Georgetown academics were equally bullish: Russian had been chosen as the language first tackled, naturally because of the Cold War, but also because of its grammatical complexity. A system able to understand Russian could be put to work on any language.

Taking the successful trial in their stride, the team confidently predicted they would then move on to broadening out the Russian language corpus, then on to French and German. “Then other Slavic, Germanic and Romance languages can be set up at will,” the New York Times wrote. Georgetown’s Leon Dostert, who had worked for US President Dwight D Eisenhower during the Second World War, and was a translator by trade, was defiant about its future. “Five, perhaps three years hence, interlingual meaning conversion by electronic process in important functional areas of several languages may well be an accomplished fact,” he told reporters.

That confidence would turn out to be misplaced.

Like many, the US government read the New York Times report of the demonstration. And like many, they were wowed by what it had managed to do. US government agencies began supporting machine translation research in June 1956, around the same time the world’s pre-eminent thinkers in the field met at Dartmouth College. In part, the government’s willingness to splash the cash was down to the successful trial at IBM headquarters in January 1954.

Three US government departments – the Department of Defence, the National Science Foundation, and the Central Intelligence Agency (CIA) – clubbed together to create the Joint Automatic Language Processing Group (JALPG). The JALPG’s mission was to support the further development of machine translation technology such as that produced by the Georgetown-IBM researchers.

An estimated $20 million was spent on machine translation and other, closely related subjects in the decade after that first New York test – the equivalent of $186 million today. And despite being good at that small vocabulary, the Georgetown IBM team struggled to get similar results when the dictionary expanded.

To find out why, a US government inquiry was opened by JALPG in 1964, led by the Automatic Language Processing Advisory Committee (ALPAC). Overseeing the ALPAC report was John R Pierce, a poindexterish employee of Bell Labs, supported by a phalanx of researchers, including those at the Massachusetts Institute of Technology.

The 124-page report’s findings, when they arrived in November 1966, were damning.

The last decade or more of development, bankrolled by the government, had been for nought. According to Pierce and his team, the experiment in January 1954 was the high watermark of the Georgetown–IBM project, rather than the dawn of a new age. Sometimes, the translations were wrong. Sometimes, they needed significant editing post-translation. Even when they were decent, they hadn’t evolved nearly as quickly as those behind them claimed they would.

As for trying to automatically translate Russian scientific publications as and when they appeared using technology, there was no point. The number of papers was actually low and there were plenty of human translators who could render them into English.

The report’s authors concluded that funding of machine translation should be halted until it could demonstrate that it could provide measurable returns. At that moment, it couldn’t provide those assurances, the ALPAC advised. Nor could four other machine translation systems that had followed in the Georgetown-IBM team’s footsteps: in fact, they were worse than the decade-old technology.

“Early machine translations of simple or selected text,” the report claimed, “were as deceptively encouraging as ‘machine translations’ of general scientific text have been uniformly discouraging. […] We do not have useful machine translation [and] there is no immediate or predictable prospect of useful machine translation.”

The government-backed investigation into the promise of machine translation – an early form of AI – could not have been more transparent. And while the funding didn’t disappear overnight for AI projects, it started to dwindle as confidence in the feasibility of the technology waned.

Little did those working on the early sparks of AI development, reading that ALPAC report in 1966, know that worse was to come, because the US was far from the only country investigating the promise of AI.

The UK, home to the inventor of the Turing test, was, too. And it would imminently come to similar conclusions, which signalled the death knell of the first AI revolution and the chill of the first AI winter.

Prior to 1971, Sir James Lighthill was best known not for his expertise in AI but instead for his knowledge of fluid dynamics. A prolific publisher of academic research, he was the Lucasian Chair of Applied Mathematics at Cambridge University – a role that Stephen Hawking would later fill. Lighthill’s connection to Cambridge went back years: he joined Trinity College at the university in the late 1930s, at the tender age of 15, a true child prodigy. He was involved in designing the wing of Concorde, the supersonic aircraft, and was a known entity to the government. Which is why, in September 1971, Brian Flowers, then the chair of the UK’s Science Research Council (SRC), which oversaw the funding of scientific research in the country, asked Lighthill to look into whether the SRC was spending its money wisely .

“There are few subjects which at any particular time strike one as having a very special potential for being pervasively important,” Flowers wrote, but: “Artificial Intelligence seems to me to be such a field, overlapping as it does with neurobiology, psychology, linguistics, and computer-aided learning, not to mention mathematics and computer science proper.” Flowers added a note of caution: subjects like AI “are highly complex, very much the preserve of experts and perhaps of plausible charlatans.”

Lighthill was flattered to be asked to investigate the tech, but was initially reluctant. However, the researcher was won over and began scoping out the current state of AI in the UK. He began sending letters to pre-eminent researchers asking them how they felt about the present scale and future potential of AI. The letters triggered a series of conversations with academics and external stakeholders who shared their thoughts with Lighthill.

By March 1972, the academic had completed his audit. And he was wary about what was to come. “Quite frankly,” he wrote, “I am fully aware that when my report becomes widely available, I shall be involved in a great deal of controversy.” Why? “In no part of the field have discoveries made so far produced the major impact that was then promised.” An entire section of the investigation was titled “Past disappointments”. The tone was like a chiding, downhearted parent disdainful of their children.

The conclusion of the UK government, which ultimately had to decide where to spend taxpayer money wisely on the development of new technologies, was that there were too many charlatans in the AI field and too few experts. The government pulled large amounts of funding out of the space, only maintaining its support for AI research in four universities nationwide.

By the mid-1970s, the AI winter had become a global ice age.

Excerpted with permission from How AI Ate the World: A Brief History of Artificial Intelligence and Its Long Future, Chris Stokel-Walker, The Bombay Circle Press.