The 1954 translation machine that fooled headlines knew just 250 words
In 1954 newspapers hailed a bilingual robot brain turning Russian into the King's English. The Georgetown-IBM demonstration actually translated 49 hand-picked sentences using a 250-word vocabulary. Cycles like this, wild promises followed by bitter disappointment and funding cuts, have hit artificial intelligence so often that researchers named them AI winters.
The term surfaced in 1984 at a public debate during the AAAI's annual meeting, where Roger Schank and Marvin Minsky, veterans of the 1970s slump, warned business leaders that hype had run out of control and predicted a chain reaction like a nuclear winter: gloom among researchers, then in the press, then deep funding cuts, then the end of serious work. By 1987 an AI business worth a billion dollars was starting to fall apart. Historians count two big winters, roughly 1974 to 1980 and 1987 to 2000, plus smaller chills.
Machine translation was an early casualty. Cold War agencies, the CIA among them, wanted instant translation of Russian papers, and researchers expected breakthroughs. They underestimated how hard it is to know which sense of a word is meant; an apocryphal round trip through Russian turned the spirit is willing but the flesh is weak into a line about good vodka and rotten meat. A 1966 advisory report judged machine translation slower, costlier and less accurate than people, and after about 20 million dollars the National Research Council stopped funding it.
Neural networks suffered too. Frank Rosenblatt promised his perceptron might one day learn and translate languages, but Minsky and Seymour Papert's 1969 book Perceptrons highlighted its limits, and nobody yet knew how to train multilayer versions. Rosenblatt died in a boating accident soon after and never saw the field's revival in the mid-1980s. In Britain, Sir James Lighthill's 1973 report attacked AI's grandiose aims and the combinatorial explosion that stalled its algorithms on real problems, leaving research alive at only a few universities, including Edinburgh.
American money dried up for a different reason. The 1969 Mansfield Amendment forced DARPA to back mission-focused work instead of open-ended research, and Hans Moravec blamed colleagues whose promises kept escalating. One DARPA gamble did pay off: a battle-planning tool saved billions in the first Gulf War. Interest recovered from its early-1990s low, and from about 2012 machine learning drove the current boom.
Source: AI winter
