A machine translated Russian in 1954. By 1966 the arithmetic had turned against it.

The cryptographer's wager

In a 1947 letter to Norbert Wiener, Warren Weaver set down a hunch that would organize a field for two decades. An article in Russian, he wrote, "is really written in English, but it has been coded in some strange symbols. I will now proceed to decode."¹ Two years later he expanded the idea into a memorandum, "Translation," and mailed it to dozens of colleagues; accounts of exactly how many differ, but it became the single most influential document of early machine translation.²

The analogy was not a whim. Weaver had run the wartime Applied Mathematics Panel and had watched statistical cryptanalysis break enemy ciphers with machinery and probability.² His memorandum opens with a mathematician who deciphered a hundred-word message "done by someone who did not know Turkish, and did not know that the message was in Turkish."¹ If a coded message could be read without knowing its language, translation looked like the same problem: recover the meaning under the symbols. Weaver proposed reading a word through the "N words on either side" of it to fix its sense.¹

He did not promise a perfect machine. "'Perfect' translation is almost surely unattainable," the memorandum says; what he expected was statistical, error-bounded translation of technical prose, useful "even if not at all elegantly."¹ The overreach in the popular retelling was never quite his. His real mistake was subtler, and his recruits named it within a few years: decipherment recovers a known language hidden in a code, while translation crosses between two, and the two are not the same task.²

Sixty sentences at IBM

On 7 January 1954, IBM and Georgetown put the wager on public display. An IBM 701, one of roughly a hundred general-purpose computers then in existence, took Russian sentences on punched cards and printed English. "Russian was translated into English by an electronic 'brain' today for the first time," IBM's release announced; "more than sixty Russian sentences" ran, "all translated smoothly," at about one sentence every six or seven seconds.³ The system held 250 Russian words on a magnetic drum and applied six rules of reordering. Leon Dostert, who led the Georgetown side, called it "a Kitty Hawk of electronic translation" and predicted that within "five, perhaps three years" meaning conversion by machine would be "an accomplished fact."³

The front page of the New York Times the next morning agreed: this was "the first successful use of a machine to translate meaningful texts."⁴ The demonstration was real and the program ran, but its scope was narrower than the headline. A later count put it at 49 sentences and 137 dictionary entries, most drawn from organic chemistry.⁴ The chemistry sentences ran on genuine, extendable rules; the handful of general sentences the reporters quoted "were clearly produced by dictionary entries and codes specifically designed for this particular demonstration," with no negatives, questions, or compound clauses.⁴ It was a showcase built to attract funding, not a prototype, and Dostert's own group abandoned that line of coding within months.⁴ Calling it rigged goes too far. Calling it stage-managed does not.

Lookup and reordering

Strip away the spectacle and a translator of the 1954-to-1966 era did two things. It looked each word up in a stored dictionary, and it applied a short list of rules to reorder the results and pick among a word's stored senses.³ There was no model of what a sentence meant. When a word had several English equivalents, the machine chose by local rule: the surrounding words, the part of speech, a code attached to the entry. On the restricted, third-person, affirmative Russian of the demonstration, that was enough.⁴

The economics of the era made the limit concrete. At Harvard, word-by-word output on a UNIVAC II ran about 25 words a minute, several times faster than a human translator, but the machine's time cost roughly a hundred times more, so "commercial MT is out of the question at present."⁵ Speed was never the obstacle. Reordering a dictionary is not understanding a sentence.

The box was in the pen

In 1959 the United States Office of Naval Research asked Yehoshua Bar-Hillel to survey the field. He was the natural choice: in 1951 he had been the first person appointed full time to machine translation anywhere.⁶ His report, revised in 1960, turned on a single sentence. Put "The box was in the pen" inside a short story: "Little John was looking for his toy box. Finally he found it. The box was in the pen. John was very happy." Every English reader knows that pen here means an enclosure, not a writing implement. Bar-Hillel claimed "no existing or imaginable program" could reach that reading, because the fact that settles it, that a box does not fit inside a fountain pen but does fit inside a playpen, is knowledge about the sizes of objects in the world, "not at the disposal of the electronic computer."⁵

To supply that knowledge, he argued, a translator would need "not only a dictionary but also a universal encyclopedia," and knowledge of the world does not terminate: knowing a room holds eight chairs, one also knows it holds more than five, fewer than nine, and so on without end. Storing all of it was "utterly chimerical."⁵ He named Weaver's "N words on either side" a "fateful fallacy": widening the window works for the human reader the text was written for, never for a machine with no world behind the words.⁵

Here the record is routinely misread. Bar-Hillel was not machine translation's undertaker. He argued only that fully automatic, high-quality translation, which he abbreviated FAHQT, was unreachable; in the same breath he endorsed the alternatives. Post-edited machine output and machine aids to human translators were, he wrote, the "reasonable goals," and he expected working commercial centers "within a period of three to five years."⁵ His target was a specific over-claim, a Georgetown subgroup's promise of fully automatic translation of chemical Russian, which he called "definitely false."⁵

The arithmetic of readers

By 1964 the money behind machine translation was large enough to audit. A committee assembled by the Defense Department, the National Science Foundation, and the Central Intelligence Agency, chaired by John Pierce of Bell Labs, spent two years doing it. Its 1966 report, "Language and Machines," reached a verdict most of its readers had not expected. "There is no emergency in the field of translation," it stated.⁷ English already carried 76 percent of the world's scientific articles, a scientist could learn to read Russian in about 200 hours, and the government's translators were not overrun: of some 4,000 contract translators available to one agency, about 300 were used in an average month, and Washington had no vacant translator posts to fill.⁷ The supply exceeded the demand.⁷

On machine translation proper the report was blunt: there had been "no machine translation of general scientific text, and none is in immediate prospect," and raw output "makes slow and painful reading."⁷ Then it did the arithmetic. Raw machine output was not simply "twice as expensive." It was cheaper to produce: $7.63 per thousand words against $21.60 for a human translation. But readers took twice as long to get through it, about 100 words a minute against 200 for ordinary English.⁷

So the cost depended on how many people would read it. For one reader the machine saved money; past roughly seventeen or eighteen readers the extra reading time overwhelmed the cheaper production, and the human won. Because most translations went to many readers, applying machine translation across one agency's output would have lost about $1.26 million, against a $14,000 loss for the human service it would replace. Post-edited machine translation, the report added, ran slower and dearer than human translation outright.⁷

This was a claim about a ledger, and it is worth keeping apart from Bar-Hillel's. Bar-Hillel said the machine could not, in principle, resolve an ordinary sentence. ALPAC said something narrower and, for a funder, sharper: even where the machine worked, it did not pay.⁷⁸

Two bets, one deferred

The report did not end the field so much as split it. It cut the money for fully automatic translation and, more lastingly, the subject's public standing; for years afterward an interest in it was something to keep quiet about.⁸ But it did not forbid research. It urged continued support for computational linguistics "as science," not to be "judged by any immediate or foreseeable contribution to practical translation," and Pierce added a personal coda defending that funding and asking for a few million dollars a year across four or five centers.⁷ Several projects had already lost their funding before 1966; a few, at Wayne State and Texas, ran on into the 1970s. Even the sums were smaller than the legend: the report's own "about $20 million" over the decade was mostly hardware and unrelated work, and spending on machine translation itself was nearer $12 to $13 million.⁸

What the episode set apart is the distinction the report stated in its own words: "we do not have useful machine translation," but "machine-aided translation may be an important avenue," and "what machine-aided translation needs most is good engineering."⁷ Fully automatic translation and machine-aided translation were now two different bets, and 1966 had backed the second.

The first bet was not wrong, only early. Bar-Hillel's box in the pen needed knowledge no dictionary of the 1960s could hold. What he could not foresee was that the knowledge could be collected rather than written down. Statistical systems trained on parallel texts arrived at IBM around 1990, and neural translation followed at Google in 2016 across nine languages, learning the sizes of the world's objects from data the period had no way to gather.⁹ The diagnosis held: the world knowledge Bar-Hillel demanded was real. It was the prognosis, that no machine could ever acquire it, that the next fifty years overturned. On the evidence in front of it, the committee of 1966 judged the machine it actually had, and judged it right.