Google to translate poetry
Google announced that it was developing a programme to translate poetry, preserving metre and rhyme in the target language. The Guardian covered it on 24 October 2010, and much of the coverage arrived at the same conclusion: it would fail. That verdict was too neat, but the reasoning behind it is worth taking seriously, because it maps the exact frontier where translation software stops translating and starts guessing.
What Google was actually attempting
The system did not try to understand poems. It worked as a constrained search problem. Given a source line, the engine would generate candidate translations, then filter them against formal requirements: a target number of syllables, a stress pattern, a rhyme with an earlier line. Where an ordinary engine picks the highest-probability rendering, the poetry system deliberately accepted a lower-probability one if it fit the form. The researchers were explicit that quality would suffer in exchange for shape.
That trade-off is the heart of the objection. A poem is not a message with decorative packaging. In a good poem, the packaging is part of the message. Force the machine to hit a rhyme and it will reach for a word that means something slightly different, and the slight difference is often the whole poem.
Why poetry breaks statistical models
Statistical and neural translation engines are built on regularity. It learns from vast parallel corpora that a phrase in one language usually corresponds to a phrase in another, and it gets extremely good at the usual case. Poetry is a systematic attack on the usual case. Consider what a translator has to carry across:
- Sound. Assonance, consonance and rhythm that exist only in the phonology of the source language.
- Ambiguity. Lines built so that two readings stay alive at once, where a machine must pick one and discard the other.
- Register. A deliberately archaic or vulgar word chosen precisely because it clashes with its neighbours.
- Allusion. A phrase that echoes scripture, a folk song or another poem, invisible to a system trained on news text and web pages.
Human literary translators do not solve these problems. They negotiate them, losing something on purpose in order to keep something else. That negotiation requires a view about what the poem is for, and no probability distribution has one.
The experiment was still useful
Dismissing the project as a stunt misses what it demonstrated. Adding hard formal constraints to a decoder is a genuinely difficult engineering problem, and solving it has applications well beyond verse: song lyrics that must be singable, dubbing scripts that must match lip movements, advertising slogans that must fit a fixed line length, subtitles that must be readable in two seconds. All of these are translation under constraint, and all of them were improved by the underlying research.
The commercial world has since absorbed the lesson. Modern engines rarely run unsupervised on anything that matters. They run inside workflows where a professional sets the terminology, checks the output and rewrites what the model got confidently wrong. The practical question is no longer whether to use the tool but how to fold machine translation into a process with a human at the end of it.
What changed after 2010, and what did not
The statistical systems of that era were replaced by neural ones, and the improvement was dramatic. Fluency went from stilted to plausible almost overnight, which created a new problem: output that reads beautifully and says the wrong thing. A clumsy translation announces its own unreliability. A fluent one does not.
Large language models pushed this further. Ask a current model to translate a sonnet with the rhyme scheme intact and it will produce something metrically competent and emotionally flat, or metrically loose and occasionally striking. What it will not do is make the deliberate, defensible sacrifice that a translator makes when she decides that this particular poem lives in its imagery and can afford to lose its rhyme. The machine has no stake in the outcome. It cannot care which loss hurts less.
The honest scoreboard
Where does that leave Google Translate and its descendants? Extremely good at gist. Very good at technical prose with controlled vocabulary. Adequate at commercial copy, provided someone reviews it. Unreliable at anything where tone carries the meaning. Actively dangerous in law and medicine without review. And in poetry, capable of producing a passable imitation of a poem, which is not the same thing as a translation of one.
Translators themselves have long since stopped arguing about whether the machines are coming, a debate that plays out endlessly in forums like r/TranslationStudies. The interesting arguments now are about rates, about post-editing, and about which texts should never be fed to a model in the first place.
The Guardian's sceptics were right that the poetry project would not produce great poetry. They were wrong about why it mattered. Google was not trying to replace poets. It was testing how far a system can be pushed before the thing it produces stops being a translation and starts being a new poem written by an author who has read nothing and felt nothing. That boundary is worth knowing, and the only way to find it is to walk up to it.
Where the money went instead
Follow the investment and you can see what the industry learned. Almost nothing went into automated literary translation, a market that barely exists in commercial terms. The money went into the boring, enormous categories: product manuals, support articles, e-commerce listings, legal boilerplate, user-generated reviews. These are texts nobody wanted to pay a human to translate at volume, and the machines unlocked them. Publishers now translate a support knowledge base into thirty languages as a matter of routine, something no budget on earth would have covered in 2005.
The knock-on effect for professionals was not the extinction that was predicted, but a change of job description. The work drifted from producing sentences to auditing them. A post-editor reads output that is 90 percent acceptable and hunts the 10 percent that is confidently, fluently wrong: the negation that flipped, the dosage that changed unit, the brand name the model helpfully translated. It is a different skill, closer to proofreading crossed with detective work, and it is measured by throughput in a way that drafting never was.
Poetry sits at the far end of that spectrum, which is why it made such a good test case. It is the text type with the lowest tolerance for the machine's core habit of choosing the likeliest word. Everything valuable in a poem lives in the unlikely one.