The Georgetown IBM experiment of 1954 was the first public demonstration of machine translation, and it worked exactly as designed: it produced English from Russian, in public, on a computer, and it convinced the people who controlled research budgets that the problem was nearly solved. Both halves of that sentence matter, because the demonstration was technically honest and publicly misleading at the same time.
What the Georgetown IBM experiment actually did in 1954
Leon Dostert of Georgetown University and Cuthbert Hurd of IBM ran the Georgetown IBM experiment on an IBM 701, the company's first commercial scientific computer. Peter Sheridan wrote the program and Paul Garvin handled the linguistics. Around sixty sentences on chemistry and general topics were transliterated onto punched cards by an operator who knew no Russian, and the machine printed English.
The sentences were chosen in advance. The vocabulary was 250 items, selected so that the chosen sentences would parse. The grammar was six rules, each a numbered instruction attached to dictionary entries telling the program how to reorder or select. Nothing about this was hidden in the technical write-ups, where the scale is stated plainly.
"Mi pyeryedayem mislyi posryedstvom ryechi" came out as "We transmit thoughts by means of speech." The machine did that. It could not have done a sentence nobody had prepared for.
The IBM 701 and the constraints it imposed
The Georgetown IBM experiment ran on hardware that shaped every choice in it. The IBM 701 had Williams-tube memory measured in thousands of words and read from magnetic tape and punched cards. A 250-item dictionary was not a modest choice, it was close to what could be held and searched at speed. The six rules exist because rule application cost cycles nobody had. Reading the experiment as a toy misses that it was a serious piece of engineering against real limits.
How the Russian demonstration was sold, and to whom
The demonstration was aimed at a specific customer. In 1954 the United States wanted Russian scientific literature read at volume, and the press releases said so. Dostert predicted useful systems within three to five years. The reporting went further, and the gap between the six-rule program and the coverage is where the later disappointment was manufactured.
That gap is the direct cause of the reckoning twelve years later. The Georgetown IBM experiment of 1954 was reported as a solved problem rather than a first step. The first public demonstration of machine translation set an expectation that the field spent a decade failing to meet, and the ALPAC report of 1966 measured it against that expectation rather than against what the systems had actually promised in their technical papers.
What the experiment proved, and what it only suggested
It proved three things, all of them real. A machine could hold a bilingual dictionary and search it at speed. A machine could apply ordering rules to produce output in the target language's word order rather than the source's. And an operator with no knowledge of the source language could prepare the input, which meant the process did not smuggle a translator in at the front.
What it only suggested was everything the coverage took from it. Six rules chosen to fit the sentences do not show that a few hundred rules would fit a language; a 250-word dictionary says nothing about the ambiguity that appears at ten thousand. The demonstration was a proof of mechanism presented as a proof of feasibility, and the distinction was lost between the technical write-up and the front page.
That gap set the terms of the next twelve years. Funding arrived sized to the promise rather than to the mechanism, and when the results were measured in 1966 they were measured against the promise.
How many sentences were translated?
Accounts differ between roughly forty-nine and sixty, depending on whether the count is of the prepared set or of what was run on the day. Contemporary press reports say sixty.
Was the Georgetown IBM experiment of 1954 rigged?
Not in the sense of faking output. The machine really translated, the operator really did not read Russian. The sentences were selected in advance to fit a 250-word dictionary, and that limitation was published. The exaggeration lives in the press coverage, not the experiment.
What happened to the Georgetown system afterwards?
It grew into a much larger rule-based system installed at Euratom in Ispra and at the Oak Ridge National Laboratory, where it ran into the nineteen-seventies, several years after United States research funding had been withdrawn.
Who actually built the system?
Leon Dostert of Georgetown directed it, Paul Garvin wrote the linguistic rules, and Peter Sheridan of IBM wrote the program. Cuthbert Hurd provided the machine and the venue. The division of labour between linguist and programmer set the pattern for the next decade.
Before this: the 1949 memorandum. After it: the 1966 report. Or see the whole sequence.