Know a company with years of records? Refer it and earn up to $100K
Translation, localisation and subtitling companies

Licensing Translation Memory and Subtitle Files for AI

Can a translation agency license its translation memories or subtitle files to AI buyers? Who owns the memory, what clients must agree to, what to strip.

A signed licensing agreement with a fountain pen and archive boxes
Short answer

A translation company can license a translation memory only where it owns the memory or its clients agree. Each segment pairs a client's source text with a translation, so client confidentiality and copyright in both halves come first. Memories built from the company's own content, or from clients whose contracts allow reuse, are the candidates. Subtitle files follow the rights in the film or programme, which usually belong to the content owner.

  • A memory is two things at once: a database you built and thousands of sentences that came from clients.
  • A translation is a derivative work. Rights in the source text do not pass to the translator.
  • Public parallel text is abundant. Reviewed, domain-specific memories with metadata are what is scarce.
  • Strip client names, personal data and unreviewed machine output before describing a memory.

Can a translation agency license its translation memories?

Yes where the agency has the right to, which is narrower than where it has the files. Sort memories by client and by contract. The memory for a client whose agreement says the memory is theirs, or that all materials are confidential, stays out unless that client agrees.

The table shows the usual positions. This page is general information, not legal advice. Your client and vendor contracts decide.

Memory or fileUsual positionCan it be licensed?
Memory built for one client under a contract that assigns itThe client'sOnly by the client, or with its written agreement
Memory built for a client, contract silent on memoriesDisputed. Confidentiality usually still appliesAsk the client
Memory of the agency's own content or of public-domain textThe agency'sYes
Master memory merged from many clientsMixedOnly the segments that can be traced to cleared clients
Termbases and glossariesOften the client's terminologySame test as the memory
Subtitle and caption filesFollow the rights in the programmeBy the content owner, or with its agreement
Freelancers' own memoriesSubject to the agency and client contracts they signedRarely

Who owns a translation memory: the agency or the client?

The contract decides, and many contracts say the client. Where the contract is silent, the answer is unsettled: the client owns the source text, someone owns each translation, and the agency assembled the database.

US copyright law treats a translation as a "derivative work" (17 U.S.C. 101), and "the copyright in a compilation or derivative work extends only to the material contributed by the author of such work" (17 U.S.C. 103). So the translator's rights never reach the client's source sentences. A commissioned translation can also be a work made for hire: the statute's list of commissioned works includes one ordered for use "as a translation", "if the parties expressly agree in a written instrument signed by them" (17 U.S.C. 101). The commissioning party then "is considered the author" (17 U.S.C. 201).

The EU adds a separate right in databases for "the maker of a database which shows that there has been qualitatively and/or quantitatively a substantial investment in either the obtaining, verification or presentation of the contents". The maker is "the person who takes the initiative and the risk of investing", which "excludes subcontractors in particular" (Directive 96/9/EC). An agency that paid to build a memory at its own risk has an argument. One that built it as a paid service for a client has a weaker one.

What about client confidentiality inside the segments?

It applies to every segment. A memory is a sentence-by-sentence copy of what clients sent: contracts, manuals, clinical documents, unreleased product text. Owning the database does not release you from a promise to keep the content confidential.

The profession's own code commits translators "to hold in confidence, not divulge, and protect privileged and/or confidential information obtained in the course of our work" (ATA Code of Ethics). Most client agreements and non-disclosure agreements say the same with a remedy attached.

Segments also carry personal data: names in correspondence, patients in medical reports, parties in legal documents. Under European law that is "any information relating to an identified or identifiable natural person" (GDPR, Article 4), and it has to go before a memory leaves. The de-identification guide covers the method.

What are translation memories and subtitle files worth?

No public source gives a price for this kind of record, so this page does not quote one. Plain bilingual text is not scarce, so value sits in what public data lacks.

Large open collections of parallel text already exist. One public index describes itself as "an overview of the OPUS collection" and lists subtitle and web-crawled corpora among its largest sets (OPUS). A general-domain memory in a major language pair competes with that. A reviewed memory in a regulated or technical field, with the revision history kept, does not.

Raises the valueLowers the value
Specialist domains: legal, medical, patents, engineering, financeGeneral marketing and web copy
Language pairs with little public dataMajor pairs in general domains
Segments reviewed by a second linguist, with the status recordedUnreviewed or machine-translated segments mixed in
Both the first translation and the reviewer's correctionFinal text only
Dates, domain tags and a matching termbaseBare segment pairs
Document order kept, so context can be rebuiltShuffled, de-duplicated segments
Written client consent on fileMemories no one can trace to a contract

A quality process leaves the review trail buyers look for. ISO 17100 "provides requirements for the core processes, resources, and other aspects necessary for the delivery of a quality translation service" (ISO).

Are subtitles covered by the rights in the film or programme?

Yes. A subtitle file is a transcription or translation of the dialogue, so it depends on the rights in the programme. The studio, distributor or platform that commissioned the subtitles usually owns them or controls their use.

A subtitling company can license files only for content it owns or where the content owner agrees. The same reasoning about derivative works applies (17 U.S.C. 101). The guide for video production archives covers rights in the pictures and sound.

Which formats can a buyer use?

The standard exchange formats, exported with their metadata. Do not flatten a memory to a two-column spreadsheet: the attributes are part of the value.

FormatWhat it holdsKeep
TMXTranslation units: "an entry consisting of aligned segments of text in two or more languages is called a Translation Unit" (TMX 1.4b)Creation and change dates, user IDs replaced by codes, domain and project properties
XLIFFProject files. Its purpose "is to store localizable data and carry it from one step of the localization process to the other" (XLIFF 2.1)Segment state, match origin, notes
WebVTT and other subtitle formatsTimed text. "WebVTT files provide captions or subtitles for video content" (W3C)Timings and line breaks, with the programme identified by a code
Termbases (TBX or CSV)Approved terms and forbidden terms per client or domainDefinitions and usage notes

Example: inventory of a 40-person translation company

A fictional agency has worked in legal and life-science translation for fifteen years. Its operations lead builds a one-page inventory. The figures are invented to show the method.

MemorySizeContract positionDecision
Life-science master, 11 clients9 million segments, 14 pairsThree contracts assign the memory to the client. Eight are silentThree out. Eight in only if each client agrees in writing
Legal memory, law-firm clients4 million segmentsAll under confidentiality termsOut unless a firm and its client both agree
Own marketing and training content200,000 segmentsAgency contentIn
Public-sector tenders, published text1.5 million segmentsSource documents are publicIn after a check of each contract
Subtitle files for a streaming client6,000 programmesClient owns all deliverablesOut
Post-edited machine translation3 million segmentsMixed clientsTagged separately. Useful only with the edit history

Fictional example. The agency starts by writing to the eight clients whose contracts are silent. What it can describe on day one is small. What it could describe after those answers is most of the life-science memory.

What is removed first?

Anything that identifies a client or a person, and anything that is not human translation you can stand behind.

  • Segments from clients who have not agreed.
  • Client, product and project names in segment text and in properties.
  • Personal data: names, addresses, case and patient details.
  • Translator and reviewer names in the user fields. Replace them with stable codes.
  • Unreviewed machine translation, or tag it so it can be excluded.
  • Segments copied from a client-supplied memory that another vendor built.

Common mistakes translation companies make

Most come from treating the master memory as the company's own asset.

  • Assuming the agency owns a memory because it runs the server.
  • Merging all clients into one memory so segments can no longer be traced.
  • Forgetting that freelancers and partner agencies signed confidentiality terms too.
  • Exporting plain segment pairs and dropping dates, status and domain tags.
  • De-duplicating and shuffling segments, which removes document context.
  • Mixing raw machine translation with human work.
  • Offering subtitle files for content the company does not own.

A translation memory as a data asset

Fifteen years of reviewed specialist translation, with each correction kept, is a record of expert judgement in two languages at once. Buyers who build translation and multilingual models look for that depth. A company with speech as well as text may also hold multilingual call recordings. The part you can license is the part your clients have cleared.

Origin works with buyers and helps a translation company sort its memories by contract, describe the cleared part, find out what it could be worth and decide whether to go further. Nothing is exported at the start. Origin Data Partners is an independent data partner that helps companies get their data bought. A buyer may pay Origin when a deal closes.

Does your company hold records like these? See what your data is worth.

Common questions

Can you sell translation memory to AI companies?

Yes where you own the memory or your clients agree. Each segment contains a client's source text, so contracts and confidentiality terms decide. Memories of your own content, or of clients who consent in writing, are the candidates.

Who owns a translation memory, the agency or the client?

The contract decides, and many assign it to the client. Where the contract is silent the position is unsettled: the client owns the source text, and the agency may hold rights in the database it invested in. Confidentiality still applies.

Can you sell subtitle files to AI companies?

Only for content you own or with the content owner's agreement. Subtitles are a transcription or translation of the dialogue, so they follow the rights in the film or programme.

What are translation memories worth to an AI buyer?

No public source gives a price. Reviewed memories in specialist domains or scarce language pairs, with revision history and metadata, are worth more than general text in major pairs, which public corpora already cover.

Are rare language pairs worth more in a translation memory?

Generally yes, because little public parallel text exists for them. The memory still has to be human-reviewed, in document order and cleared by the clients whose text it contains.

What has to be removed from a translation memory before licensing?

Segments from clients who have not agreed, client and product names, personal data, translator names in user fields and unreviewed machine translation. Keep dates, status and domain tags.

Sources and further reading

What is your company’s data worth?

AI companies pay $20K to $5M for company records. Six questions, no files.