How translation and localization agencies can sell data to AI companies

Every finished translation project leaves behind aligned sentence pairs, a reviewer's corrections and decisions about terminology. For machine translation and multilingual models, that is the hard part to find: human judgment on real text, across language pairs that public data covers thinly.

Very little of it belongs to the agency alone. The source text belongs to the client, the translation may belong to a freelancer, and the documents often hold personal data.

What translation data do AI companies buy?

AI companies buy translation data that pairs source text with a human-quality target and shows where people corrected machines. Post-edited machine output kept alongside the raw machine output is especially useful, because the difference between the two is a correction a model can learn from.

Record language pairs, domains, segment counts and date ranges in your data manifest. Buyers of multilingual text data check those details first.

Who buys translation data?

Language data buyers are among the few with published supplier routes. Defined.ai lists machine translation among its services and runs a marketplace where vetted partners sell datasets, and it says only 30% of partners are approved after ethical vetting. Submitting its partnership form means accepting its supplier terms, so read them first.

The Linguistic Data Consortium publishes contributed corpora and asks providers for ISO 639-3 language codes, size, format, origin and collection methods. It does not require exclusive rights and does not describe payments to providers. No buyer publishes a price range for translation data, so a free data valuation is the place to start.

Who owns a translation memory?

A translation memory usually carries at least two sets of rights. Under 17 U.S.C. section 101, a translation is a derivative work, and under section 103(b) the copyright in a derivative work extends only to the material its author contributed, independent of the copyright in the preexisting work. The client or the original author keeps the rights in the source text.

The target text belongs to whoever wrote it, unless a contract says otherwise. Section 101 lists a translation among the commissioned works that can be a work made for hire, but only if the parties expressly agree in a written instrument signed by them. Without that agreement or a signed assignment, a freelance translator owns the copyright in their translation. Work by employees within the scope of employment belongs to the agency under section 201(b).

Audit your freelancer agreements before you license anything. Where a signed work made for hire clause or assignment exists, you may hold the target text. Where it does not, leave that translator's work out or obtain a written license.

Client NDAs and services agreements often stop you licensing translation data, because the source text is the client's confidential information. Contracts, product documentation before launch, patent filings and internal communications are typical translation work, and an NDA that limits use of confidential information to performing the services rules out a training license.

Daniel Goldberg of Frankfurt Kurnit Klein and Selz points out that email and chat records often contain other companies' confidential information covered by NDAs and commercial agreements, and holding a record does not give you the right to sell it. Ask clients for written permission, and license only work from clients who agree and material you created yourself.

What about personal data in translated documents?

Personal data in translated documents is common in legal, immigration, medical and HR work. Birth certificates, court papers and medical reports name individuals and often include sensitive details, so exclude those projects or remove identifiers before anything leaves your systems.

Healthcare clients add HIPAA. HHS defines a business associate as a person or entity that performs functions or provides services for a covered entity involving access to protected health information, and says a business associate may use or disclose that information only as permitted or required by its business associate contract. A training license is not normally one of those uses, so leave the material out.

If you hold UK or EU data

UK copyright treats translation the same way. Under section 21(3) of the Copyright, Designs and Patents Act 1988, a translation is an adaptation, and making an adaptation is an act restricted by the copyright in the original work.

Agencies that handle client personal data in the UK or EU are usually processors. UK GDPR Article 28(3)(a) requires a processor to process personal data only on documented instructions from the controller, and Article 28(10) says a processor that determines the purposes and means of processing is considered a controller for that processing.

How do you sell translation data through Sell My Data To AI?

You start with a consultation about your language pairs, domains, client terms and freelancer contracts. Sellers pay nothing to be assessed. We charge a success fee only when a deal completes, agreed in writing before we introduce you to any buyer.

Before any buyer sees a sample, your data passes our eight-point vetting standard, which covers ownership and chain of title, contract restrictions and third-party content. You approve every buyer, every sample and every license term. Book a consultation to start.