Machine translation and the trade it might open
Machine translation papers rose to 34 in the last year, almost all in AI journals, while economics and trade journals in the archive carry none. Language is one of the oldest barriers in trade models, and translation by language models is now nearly free. Has the barrier shrunk, and for which language pairs?
What prompted this#
Since October 2025, 34 papers carry a machine-translation keyword, up from 19 in the year before: "machine translation" on 20, "neural machine translation" on 10, "multilingual neural machine translation" on 5. Thirty-two appeared in AI journals; over three years, 73 of 80 did, and none appeared in economics or finance journals. They sit next to large language models (17), low-resource languages (16) and post-editing (4). A study of a large online marketplace found that introducing machine translation raised cross-border sales (Brynjolfsson, Hui and Liu, Management Science, 2019), before language models made translation far better. A related note asks how finance reads text beyond English; this one asks what translation does to trade.
Questions#
- In gravity models of trade, has the effect of sharing a common language shrunk since language models arrived?
- Do language pairs that machine translation still handles poorly, often low-resource languages, keep their trade gap?
- Which flows respond most: goods on online marketplaces, freelance services, tourism?
- What happens to translators and interpreters in the exporting firms?
Data that could help#
CEPII gravity data, bilateral services trade statistics from the OECD and WTO, translation-quality benchmarks by language pair such as FLORES-200, and freelancing platform data.
Next step#
Interact the common-language term in a gravity model with machine-translation quality per language pair from FLORES-200, and compare its coefficient before and after 2023.