The cost of translation approaching zero

In May 2014 eBay switched on machine translation for buyers in Latin America. A shopper typing a search in Spanish had the query translated into English, and the titles of English-language listings came back translated into Spanish.¹ Erik Brynjolfsson, Xiang Hui and Meng Liu compared US exports on eBay to the Spanish-speaking countries that got the feature with exports to countries that did not. Exports rose by 17.5 to 20.9 percent, depending on the time window.¹ By their estimate, that is the trade effect you would expect from cutting the distance between the countries by 37.3 percent.¹

The translation touched only search terms and listing titles. That it moved trade so much says the barrier had been large. It also says where the gains go first.

Who gains first

The effect was largest for differentiated products, for cheaper items, for listings with longer titles, and for buyers with less experience on the site.¹ The authors read this as a fall in search costs: the people and goods for whom understanding a listing was hardest gained the most.¹

That pattern generalizes by simple reasoning. Translation that costs nothing helps most where language was the main barrier and the others were already low: an online market with payment, shipping and ratings in place. Where trust, law or logistics are the real barriers, free translation changes less.

The cost does not fall evenly

"Approaching zero" also hides a very uneven fall. Translation systems learn from text that exists in two languages at once, and that text is plentiful for a few languages and scarce for most. The team behind Meta's No Language Left Behind model notes that the data needed are not equally available for the world's more than 7,000 languages, and that most of the gains in machine translation have gone to high-resource languages, which they define as those with at least a million sentences aligned with another language.² Their own model covers 200.²

So the cost of translation is falling toward zero fastest between languages that were already well connected, and slowest for the languages whose speakers were already the most cut off. A market or a library that opens through free translation opens first to people who already had the most access.

Reading is not being read

Science is where the deeper asymmetry shows. Tatsuya Amano and colleagues surveyed 908 environmental scientists from countries with different languages and incomes to measure what it costs to do science in English when English is not your first language.³

Reading is one cost. Among researchers who had published one English-language paper, those from countries of moderate English proficiency took a median 46.6 percent longer to read a paper in English than native speakers did, and those from countries of low proficiency took 90.8 percent longer.³ In their first language, they read faster than native English speakers read English.³ That cost is exactly the kind free translation removes.

But the larger costs come from being judged. Among those who had published one English-language paper, about 38 and 36 percent of non-native speakers in the two groups reported having had a paper rejected because of its English, against 14.4 percent of native speakers, a rate 2.5 to 2.6 times as high.³ They were asked to improve their English during revision about 12.5 times as often.³ At the time of the survey, use of machine translation was relatively low in every country, and the authors urged journals and universities to allow AI tools for English proofreading.³

A translator in the reader's hand fixes the first problem. It does not fix the second, because the paper is still judged in English by people reading English. Free translation equalizes reading. It does not equalize authority, the question of which version counts.

Where the burden goes

That distinction runs through every domain the question touches. A contract, a regulation or a published paper usually has one version that counts. When everyone can read a translation of it, the people reading the translation carry the risk that it is wrong, and the people who wrote the original carry none. The barrier moves from understanding to being understood: from reading in a second language to being judged in one.

Education feels this before anything is measured. If any text can be read in your own language, the reason to learn a second one for reading weakens. The reason to write in the language of authority does not, because that is where work is judged. The likely result is people who read in many languages through a tool and write, when it matters, in English.

There is a quieter effect that no one is positioned to measure. If reading English through a translator is easy, some of what would have been written in smaller languages may not be written at all. No market prices a language's production of new text, and no institution's ledger records the paper or the story that was written in English instead.

Where this breaks

The trade evidence is one platform in 2014, from a working paper; the published version of the study was not read for this piece and may report different estimates.¹ The science evidence is a survey of self-reported effort in one field.³

Authority can be shared, and some institutions do it. Where several language versions are made equally binding, readers of each carry the same risk, and the asymmetry described here disappears, at the cost of producing every version carefully. Machine translation could make that cheaper, which would be the strongest version of the change the original question imagines.

And quality keeps improving. Some of what is called uneven here may close quickly. The data gap for small languages is the hardest part to close, because more text cannot be produced on demand.²

The cost that remains

Translation approaching zero is real, measurable and already changing who trades with whom. What it does not change on its own is where judgment happens. The cost of language used to be paid by the reader. Increasingly it will be paid by whoever is judged in a language that is not theirs, and by the languages that stop being written because reading in someone else's has become easy.