The History of Machine Translation, Recreated: One Sentence Through Every Era

In 1954, if you believe the legend, an early computer was asked to translate "The spirit is willing, but the flesh is weak" into Russian and back. It returned: "The vodka is good, but the meat is rotten."

The story is almost certainly fake. Nobody has ever produced the machine, the printout, or a witness, and versions of the anecdote circulated before any system could have attempted it. But it refuses to die, because it captures exactly how early machine translation failed: confidently, word by word, with no idea that words mean things together.

So instead of retelling the history of machine translation one more time, we decided to rerun it. One sentence, the legend's own sentence, translated the way each era would have translated it. The old eras are reconstructed by hand using their actual methods, and clearly labeled as such. The modern era is real, unedited output from Translation Mixer, collected in September 2026 with Google's engine doing the work.

The result surprised us, and it comes at the end. First, the map:

1954: Six grammar rules and a dictionary

Machine translation did not start in linguistics departments. It started with the Cold War. The United States government was drowning in Russian scientific papers it could not read fast enough, and human translators were the bottleneck. On January 7, 1954, IBM and Georgetown University staged a public demonstration at IBM's New York headquarters: more than sixty Russian sentences rendered into English by a machine holding a vocabulary of 250 words and six grammar rules.¹ The press called it a breakthrough. Researchers predicted the whole problem would be solved within three to five years.

They were off by about sixty.

The demo was carefully staged with sentences the system was built to handle. Real text collapsed it immediately, and here is why. A rule-based system of that generation worked roughly the way a tourist works with a pocket dictionary: look up each word, take the first plausible entry, apply a reordering rule if one matches, and print the result. The system does not know what the sentence is about. It cannot know, because nothing in it represents meaning at all.

The recreation (constructed by hand, following the 1954 method). Take the legend's sentence into Russian word by word, first dictionary entry wins:

The → (Russian has no articles, drop it) spirit → спирт (the dictionary's first entry: distilled alcohol) is willing → готов (ready) flesh → мясо (the everyday entry: meat, as at a butcher's counter) is weak → слабо (feeble)

Assemble and translate back with the same dictionary and you get "The alcohol is ready, but the meat is feeble." We did this with a Russian dictionary and fifteen minutes, and it is the vodka legend, near enough. The famous joke is not a report of something a machine once did. It is an accurate parody of the method, which is why it stuck. Spirit really is спирт if you only know one sense of the word, and no rule in a six-rule grammar tells you a proverb is not about groceries.

1966: The winter

By 1966 the U.S. government had spent millions on this approach and had little to show. The ALPAC report, commissioned by the Department of Defense, the CIA, and the NSF, concluded there was "no immediate or predictable prospect of useful machine translation."² Funding stopped, the field emptied out, and historians now mark it as the first AI winter. What ALPAC missed was that the task was not impossible. The approach was wrong. You cannot write down all the rules of a language, because a language is mostly exceptions.

1988 to 2015: Counting instead of understanding

The comeback began in the late 1980s with a different question: what if nobody writes rules at all? Feed the system millions of sentences that humans already translated, the proceedings of the Canadian Parliament in English and French, the EU's legislative archive, the UN's output in six languages, and let it count.³ If "maison blanche" keeps showing up where English has "white house," the correlation is learnable with no grammar in sight. Google Translate launched on exactly this principle in April 2006.⁴

Statistical systems were a huge step up, and they failed in a new and recognizable way. Because they translated phrase-sized chunks and then stitched the chunks together, their signature error was the seam: each piece defensible, the joints slightly wrong, agreement and articles wobbling where the pieces met.

The recreation (a characteristic reconstruction, not a live run; the 2006 engines are gone). A phrase-based system handles our sentence as three independent pieces: "the spirit is" + "willing, but the" + "flesh is weak." Each phrase pulls its most frequent translation from the parallel data, which for religious vocabulary meant Bible-adjacent text, so flesh correctly becomes плоть rather than meat. But the pieces do not negotiate with each other, and the stitched result reads like "The spirit is ready, but flesh is the weak." If you used Google Translate in its first decade, you remember this texture: right words, drunken syntax. It was the era that invented "Google Translate fails" as a genre, because the output was finally good enough for its mistakes to be surprising.

2016: The machine learns the verse

In November 2016, Google switched Translate to a neural engine, GNMT, which encodes a whole sentence into a mathematical representation of its meaning and decodes that into the target language.⁴ Users noticed the jump overnight. The transformer architecture followed in 2017⁵ and now underlies both modern translation and the current generation of AI assistants.⁶

So what does the neural era do with the sentence that launched a seventy-year-old joke?

The recreation (real mixer output, September 2026, Google engine). English → Russian → English, the legend's exact route:

In: The spirit is willing, but the flesh is weak. Russian: Дух бодр, но плоть немощна. Out: The spirit is willing, but the flesh is weak.

It came back character for character identical. And the Russian in the middle is not just correct, it is the canonical Synodal Bible wording of Matthew 26:41, the exact phrasing a Russian priest would use. The machine has read the verse, in both languages, many times. Seventy years after the legend, the vodka sentence is one of the safest things you can possibly feed a translator, because it is famous. The failure the joke describes has been fixed so thoroughly that the sentence is now closer to a memorized fact than a translation problem.

That felt like cheating, so we pushed.

2026: What it takes to break it now

If one well-known language cannot budge the sentence, a chain might. Real runs, same engine, same day:

Three languages (Russian → Japanese → Finnish):

Out: My heart burns, but my body is weak.

Six languages (Russian → Japanese → Turkish → Hindi → Swahili → Welsh):

Out: My heart is on fire, but my body is weak.

The interim steps tell you exactly where the drift happens, and it is not where the 1954 method broke. Russian hands Japanese the biblical wording faithfully. Japanese, the language that does more damage to chained translations than any other we've measured, has no everyday idiom for an abstract willing spirit, so it reaches for its own figure of speech: 心は燃えている, "the heart is burning." From that hop on, every language downstream translates the Japanese metaphor with perfect competence. Turkish, Hindi, Swahili, and Welsh change almost nothing. Three more hops only polish "burns" into "is on fire."

That is the modern failure mode in miniature, and it is the opposite of the old one. The 1954 method broke on every word and knew nothing. The 2026 method translates each hop so well that a single reasonable substitution, one language's honest attempt at another language's metaphor, gets carried faithfully to the end of the line. There are fewer errors now, and they are much harder to spot. We have watched the same mechanism turn a duck into an Australian and "I'll be back" into a clergyman named Father Goshen, and it is why history's famous mistranslations still mostly reproduce today: the machine no longer fumbles vocabulary, but it still commits to one reading of an ambiguous thing and never looks back.

→ Run the vodka sentence through the six-language chain yourself

One sentence, four eras

Each era failed at the level it worked on. The 1954 method broke on individual words and the 2006 one on the joins between phrases. The current one gets words and grammar right and can still swap out a whole idea, which every later hop then translates faithfully.

Where that leaves us

Modern machine translation is genuinely extraordinary. Google Translate covers 249 languages and moves over 100 billion words a day,⁴ and for well-resourced pairs its output can be hard to tell from human work. The remaining limits are not bugs in a queue. Low-resource languages drift more because there is less to learn from, which is measurable: when we round-tripped one paragraph through 20 languages and scored the damage, the ranking mostly tracked training data and structural distance, not the languages people assume are "hard." Register, irony, and ambiguity still slip through even when every word lands. In medicine and law, that gap is why human translators still sign off.

And it is why the telephone game we run in production stays interesting. Every language in a chain is a slightly different model of the world. Chain enough of them and the changes that pile up are less like typos and more like one language handing its best guess at a metaphor to the next. The vodka sentence needed six languages to move at all, and even then it only moved once, at Japanese. Most sentences are not famous Bible verses, and they move a lot sooner.

---

Related: Famous Mistranslations in History, Recreated reruns the disasters that were real, including the word that may have ended the world. The Translation Telephone Game traces the schoolyard original that this whole field accidentally rebuilt. Why Machine Translation Humor Works explains the linguistics of why the failures are funny. Or skip the reading and feed the mixer a sentence of your own.

---

Sources & Further Reading

1. Hutchins, W.J. "The Georgetown-IBM experiment demonstrated in January 1954." Proceedings of the 6th Conference of the Association for Machine Translation in the Americas, Springer, 2004. See also: Georgetown-IBM experiment, Wikipedia.

2. ALPAC. Language and Machines: Computers in Translation and Linguistics. National Academy of Sciences, National Research Council, Washington D.C., 1966. Chaired by John R. Pierce.

3. Wikipedia. "History of machine translation."

4. Wikipedia. "Google Translate." Launched April 28, 2006; switched to neural MT November 2016; supports 249 languages as of 2026.

5. Vaswani, Ashish, et al. "Attention Is All You Need." 31st Conference on Neural Information Processing Systems (NeurIPS), 2017.

6. Wikipedia. "Attention Is All You Need." Cited more than 173,000 times as of 2025.

The chain outputs above are unedited Translation Mixer results collected September 1, 2026, with Google as the active translation provider. The 1954 and 2006 entries are reconstructions of those eras' methods and are labeled as such; the originals no longer exist to run.