Can a company license its records in other languages to AI companies?
Yes, and records in languages with less text online can be among the harder ones for labs to find. A company that serves customers in Spanish, Portuguese, Tagalog or Arabic, or that keeps paired translations of its documents, holds language from real work that is thin on the public web. The same ownership and privacy checks apply as for English records.
What a company may hold in other languages
| Support conversations in other languages | Chats and emails with customers who wrote in their own language, with the agent replies. |
| Paired translations | A document and its professional translation, kept side by side. Each pair shows how meaning moves between languages. |
| Mixed-language team chat | Teams that switch between two languages in one message. This kind of writing is common at work and rare online. |
| Local market sales and operations files | Quotes, orders and notes from offices abroad, written in the local language. |
| Interpreted call transcripts | Calls with an interpreter, where the transcript records both languages. |
Why language matters to a buyer
Most training text is English. Models do worse in languages with less text, and real conversation in those languages is harder to find than translated web pages. The mandate of a lab may ask for a named language, a region, or text that mixes languages the way people really write.
Paired translations are their own category. A document and its checked translation teach models to translate, and a translation memory kept by a company over years can hold hundreds of thousands of paired sentences.
Where the records were made changes the rules
Records made outside the US follow the law of the place they were made. Data about people in the EU falls under the GDPR, India has its Digital Personal Data Protection Act, and Brazil has its LGPD. Counsel in each place confirms whether personal data can be removed to their standard and whether the records may leave the country.
Data from outside the US, UK, Canada and Australia is often priced lower today. A rare language can offset that when a buyer has a mandate for it.
Checking and cleaning
Count each language separately: messages or words per language, years and the source system. Mark which records are machine translated, because buyers treat them differently from text people wrote.
Names, phone numbers and addresses take local forms, so cleaning tools need settings for each language, and a speaker of that language checks a sample. The owner approves the sample before it goes anywhere.
Records that stay out
| Machine translations offered as human text | Labeled as machine output or left out, since a buyer needs to know the difference. |
| Records that local law keeps in the country | Where transfer rules forbid moving them, they stay put. |
| Translations owned by an agency | If the translation contract kept rights with the agency, the pair stays out until it agrees. |
| Personal data that cannot be found reliably | Records in a language the cleaning tools handle poorly are reviewed by hand or left out. |
More on this: how personal details are removed, who carries the risk and how you get paid.
An example, invented for illustration
Imagine an online travel agency with a hundred staff that has answered customers in English, Spanish and Portuguese for nine years, and kept a translation memory for its hotel descriptions. The support chats are counted by language, customer names and booking numbers are replaced, and a Spanish and a Portuguese speaker check the cleaned sample. Chats handled by its machine translation tool are marked so a buyer can include or exclude them.
Questions owners ask about this
Which languages are most in demand?
It changes with each buyer. Languages with less public text tend to draw interest, and a buyer with a specific mandate will name the language, so it helps to know your counts per language first.
Is our translation memory licensable?
Often, if the company owns it. Check the contracts with translation agencies, because some keep rights to the memories they build.
Does mixing English with another language in one message cause problems?
No, it is useful. Mixed writing is how many teams talk, and models need to see it. Mark the main languages so the set can be described.
Didn't find your question? Ask us directly.
Keep reading
Related kinds of records
- Support tickets: The whole thread, wrong guesses included, with customers replaced by stand-ins.
- Team chat: Slack and Teams threads, where one person must stay one person and direct messages stay out.
- Speech and audio recordings: Dictation, voicemail, meeting audio and narration, where the consent of every speaker decides what can be used.