No taker publishes a price per segment or memory (checked 2026-11-09), and the one broker on record, Sell Translation Data To AI, states that the buyer of the segments pays its fee at a rate fixed in an agreement, without publishing the rate (checked 2026-11-16).
This page covers AI training data licensing for translation data: who distributes it, which terms bind a provider, who owns a memory and when not licensing is the better answer. The takers license a copy onward, they do not buy the memory outright. General text corpora and multilingual or low-resource collections have their own pages.
The DataPacts Editorial Team explains how figures are checked.
What a Translation Memory Holds When It Is Licensed
A translation memory is a database of aligned source and target segments built up as translators work, and in a license it is one of several kinds of parallel text a language business holds.
Each segment carries three interests: the client's source text, the translator's target text and the compiler's work in the database. Section 4 sets out how they divide.
The table lists the kinds of parallel text a language business may hold, as of 2026-11-16. General text corpora are covered under every data type, where the neighbors of this page sit.
| Kind of parallel text | What it contains | Typical holder | Where it appears on record | Source |
|---|---|---|---|---|
| Translation memory | aligned source and target segments from finished jobs | translation agencies, in-house language teams | a broker's subject; a TAUS preparation service | selltranslationdatatoai.com; taus.net |
| Termbase or glossary | approved terms per language pair and domain | agencies, client language teams | not named by any taker | n/a |
| Aligned parallel documents | whole documents with their translations | publishers, public bodies, agencies | language resources in distributor catalogs | elra.info; ldc.upenn.edu; gsk.or.jp |
| Post-edited machine translation | machine output with the translator's corrections | agencies, tool operators | not named by any taker | n/a |
Who Takes Translation Memories and Parallel Text
Translation data reaches AI developers mainly through intermediaries: language-resource distributors that license providers' corpora onward, dataset marketplaces that list them, and a broker that looks for a licensee on a firm's behalf.
None of them acquires a memory outright. The table lists the organizations that name translation or language data on their own pages, as of 2026-11-16, every value as the organization states it (checked 2026-11-09).
| Organization (role) | What it handles | Stated condition | Source |
|---|---|---|---|
| Defined.ai (licensing marketplace) | text, audio, image, video and multimodal datasets from data partners | ethically sourced with consent, documentation and right to commercialize | defined.ai/partnerships |
| ELRA / ELDA (language-resource distributor) | language resources under a distribution license from the provider | provider fills description forms | elra.info |
| Gengo-Shigen-Kyokai (GSK) (language-resource distributor) | speech data, lexicons, text corpora | users may not sell, lend or distribute to third parties | gsk.or.jp |
| Linguistic Data Consortium (corpus distributor) | speech, text and lexical corpora; a submissions portal for providers | for-profit users need a For-Profit Membership Agreement | ldc.upenn.edu |
| Mozilla Data Collective (dataset marketplace) | datasets from organizations, speech, language and other; launch providers include TAUS and Pangeanic | provider sets price and license | mozilladatacollective.com |
| Sell Translation Data To AI (broker) | translation memories of translation firms | the buyer of the segments pays the broker's fee; the firm is not charged by the broker (checked 2026-11-16) | selltranslationdatatoai.com |
| TAUS (language-data services) | cleaning and model-training preparation of translation memories | a service, not a license | taus.net |
The distributor model is described under dataset vendors and data labs. No AI developer on record states a program for translation memories, so a direct license to a lab is not described here. TAUS offers preparation, not a license.
Distribution Terms a Language-Data Provider Agrees To
The terms that bind a provider differ by taker: ELRA asks the provider for a distribution license and lets it choose research-only or product-development use, the Linguistic Data Consortium (LDC) grants for-profit members lasting commercial rights, and Gengo-Shigen-Kyokai (GSK) forbids users to pass resources on.
The table compares the terms, as of 2026-11-16, checked 2026-11-09.
| Taker | Grant from the provider | What the end user receives | Provider compensation, as stated | Source |
|---|---|---|---|---|
| Defined.ai Supplier Program | no exclusivity, no minimums (s.3.5); 12-month non-circumvention; 60 days' notice to terminate | customer license "perpetual" (FAQ) and "until terminated" (Data License Agreement s.8(a)), both stated | no revenue share published | defined.ai/supplier-program; defined.ai/faq; defined.ai/data-license-agreement |
| ELRA / ELDA | distribution license from the provider after description forms | non-exclusive, non-transferable right to use, rework and build on; no redistribution; product development or research only, as the provider allows | by contract, not quantified; fees per a yearly schedule | elra.info |
| Gengo-Shigen-Kyokai (GSK) | not stated | no selling, lending or distributing to third parties; credit to resources and providers | not stated | gsk.or.jp |
| Linguistic Data Consortium | submissions portal | for-profit members retain commercial rights to data acquired during membership; perpetual rights to data licensed in membership years | not stated | ldc.upenn.edu |
| Mozilla Data Collective | provider uploads and sets price and license | set by the provider; not standardized | "Own your data and keep 100% of revenue" (home page claim) | mozilladatacollective.com |
The provider's choice of field of use is the main lever: ELRA states that some providers allow product development and others research only (checked 2026-11-09), so a provider that wants no commercial reuse can say so.
Defined.ai words its customer license duration two ways on its own pages, "perpetual" in the FAQ and "until terminated" in the Data License Agreement (checked 2026-11-09).
The "keep 100% of revenue" line is Mozilla Data Collective's own claim, not a market rate. Terms for listing a finished dataset through marketplaces and direct buyers are covered on their own page.
Ownership of a Translation Memory: Client, Agency and Translator
This section is information, not legal advice, and it sets out the interests that overlap in a memory.
The Client's Source Text Inside Every Segment
Every segment in an agency's memory repeats a client's source text, so an agency that licenses the memory discloses its clients' documents, and acquirers screen such client-held material out.
Mercor's listings for software, design and SaaS firms show the practice (checked 2026-11-09). Mercor states "Client work an agency doesn't own" as not a fit on its Software Company listing, requires "The right to license the work, either your own or cleared with clients" and lists "Client work you can't license" as not a fit on its Design & Web Agency listing, and lists "Agencies holding their clients' databases" as not a fit on its SaaS & Subscription listing.
These are Mercor's practice for those company types, not a rule for translation. Its publisher listing excludes "Content distributed under someone else's license" and welcomes multilingual content. Read client confidentiality before licensing alongside any memory.
Under 18 U.S.C. 1839(5), disclosure by one who acquired a trade secret under a duty to maintain secrecy is misappropriation (Pub. L. 114-153, 2016-05-11). Whether a client contract gives the agency or the client the memory differs by contract and is not generalized here.
Translations Made by Staff and by Freelance Translators
An agency owns translations its employees make in the course of their work, but work commissioned from a freelance translator belongs to the agency only under the conditions the statute sets or a written agreement.
Under 17 U.S.C. 201(b), for a work made for hire the employer or other person for whom the work was prepared is the author and owns all rights unless a written instrument signed by both parties says otherwise.
Under 17 U.S.C. 101 a work made for hire is either a work an employee prepares within the scope of employment or a specially ordered or commissioned work in one of nine listed categories with a signed written agreement.
In the UK, CDPA 1988 s.11(2) makes the employer the first owner of a work an employee makes in the course of employment, subject to any agreement to the contrary. For commissioned translators the contract decides, which is why who owns the translations is a question for the rights audit.
Database Right in a Compiled Memory
In the EU, the maker of a database that shows substantial investment in obtaining, verifying or presenting its contents can stop extraction of substantial parts, and that right can be licensed.
Directive 96/9/EC (OJ L 77, 27.3.1996) Art. 7(1) gives the maker that right, Art. 7(3) allows it to be transferred, assigned or granted under a contractual license, and Art. 10(1) sets the term at 15 years from 1 January following completion.
Whether a given memory qualifies, and who its maker is, is not settled here. The database right in detail is covered on its own page.
Translation Tool and Platform Terms Over Stored Memories
Memories kept in a cloud translation tool sit under that tool's terms as well as the client contract, so a seller reads both before a sample leaves.
The broker Sell Translation Data To AI says it compares translation-platform terms. A seller asks four questions: who holds the account, what export rights exist, whether a clause covers AI training, and what happens to stored memories after termination.
Personal Data and Confidential Content in Bilingual Segments
Translated contracts, medical letters, HR files and correspondence carry names and confidential facts in two languages, so personal data must be removed from both sides of each segment before a memory is licensed.
Defined.ai requires data partners' datasets to be "ethically sourced with consent, documentation and right to commercialize" (checked 2026-11-09), and TAUS names a text audit and data cleaning service for translation memories (checked 2026-11-09). ELRA's research-only option is a narrower use that some providers choose.
Legal and medical translations carry their sector's own rules, which this page does not state. The method is set out under removing names from both languages, and it applies to source and target alike.
Preparing a Translation Memory for Licensing: Cleaning, Alignment and Description
A memory is prepared for licensing by removing client and personal content, checking that segments are aligned and in the stated language pair, and describing it so a taker can judge it without a copy.
- Client exclusions: segments from clients that have not cleared the use are removed.
- Personal data: names and identifiers are removed from source and target.
- Alignment and quality review: Defined.ai states a data expertise review of quality, volume and metadata (checked 2026-11-09).
- Description: ELRA has the provider fill description forms (checked 2026-11-09).
- Fields: language pairs, domains and segment counts, with a note on whether client content is cleared.
TAUS names a model training preparation service on its own page (checked 2026-11-09). The usual exchange format is TMX, named here only as a term. Describing a memory without a copy is the job of the data manifest.
Prices and Fees on Record for Language Data
No taker publishes a price per segment, word or memory; the figures on record are fees charged to buyers and two marketplace revenue claims, and the one broker on record publishes no rate.
The table lists each figure as stated by the named party, checked 2026-11-09 (broker row 2026-11-16).
| Figure | Stating party | Value as stated | Who pays whom | Source |
|---|---|---|---|---|
| Membership fee | Linguistic Data Consortium | for-profit Standard USD 34,000 (year not labeled), Subscription USD 40,000; not-for-profit Standard USD 2,400, Subscription USD 3,850 | the corpus user pays LDC | ldc.upenn.edu |
| Distribution fee | Gengo-Shigen-Kyokai (GSK) | non-members JPY 33,000 per incident for any resource; members JPY 5,500 per incident for using the free member resource service more than once in a fiscal year; taxes included | the resource user pays GSK | gsk.or.jp |
| Partner revenue claim | Defined.ai | "$600K+ average partner revenue", no period stated | licensees pay through the marketplace | defined.ai/partnerships |
| Revenue share claim | Mozilla Data Collective | "keep 100% of revenue" | licensees pay the provider | mozilladatacollective.com |
| Broker fee | Sell Translation Data To AI | rate fixed in an agreement, not published; page updated 2026-10-08 | the buyer of the segments pays the broker; the firm is not charged by the broker | selltranslationdatatoai.com |
The LDC and GSK fees are paid by users to the distributor, not income of the provider. No figure is averaged, converted or used to value a memory, and no deal for translation memories is disclosed. Prices across data types are compared under AI training data prices per work, hour and record.
Licensing Translation Memories: Benefits, Risks, Downsides and Alternatives, Including Not Licensing
Licensing a translation memory can earn from years of finished work, but the memory repeats clients' documents, and parallel text that reaches a trained model cannot be called back.
Benefits
- A distributor or marketplace gives catalog reach to many licensees (ELRA, LDC, GSK, Defined.ai).
- At Mozilla Data Collective the provider sets price and license terms.
- Defined.ai's Supplier Program states no exclusivity and no minimums.
Risks
- Client confidentiality: acquirers screen out client-held work and trade secret law applies to disclosure.
- Personal data sits in both languages of a segment.
- Ownership of commissioned translations may be unclear.
- Tool and platform terms may restrict export or AI use.
Downsides
- No public price per segment or memory exists.
- Buyer-side fees are not the provider's income.
- LDC for-profit members keep perpetual rights to data licensed in membership years.
Alternatives
- License only memories built from own or cleared content.
- Choose research-only distribution where a taker offers it.
- License termbases or glossaries without client text.
- Choose not licensing at all and keep the memory for translation work.
The choice is the holder's, and a memory that cannot be cleared is better left unlicensed.
Neighboring Language Data, Translator Work and the Seller Steps for Translation Memories
This page covers translation memories and aligned parallel text; general text corpora, multilingual and low-resource collections, speech and each sector's documents have their own pages.
- Text corpora in general
- Low-resource language collections
- Spoken language data
- Source documents before translation
- Legal translations and privilege
- Other agencies holding client work
- Every broker's published fee
- Documenting a parallel corpus
The seller steps for translation memories follow the general process:
- Read the complete seller guide.
- Sign NDAs before a sample memory leaves.
- Review every license clause.
- Check upfront, milestones, revenue share.
Frequently Asked Questions
Can a translation agency license its translation memories for AI training?
Only the parts it owns or has cleared with clients. Mercor lists client work an agency does not own or cannot license as not a fit (checked 2026-11-09).
How much are translation memories worth to AI companies?
No price per segment or memory is on record. The broker's unpublished fee and the marketplace claims are not prices (checked 2026-11-09). The wider picture is set out in what company data is worth.
Can freelance translators license their own translation memories?
Only for work they own; client text stays the client's. The ownership rules are set out in section 4 above. Read a translator's own data for the individual route; paid language work has its own page.
Which file format do AI buyers want for translation memories?
No taker on record names a format (checked 2026-11-09). Format and transfer are set in the license. The delivery side is covered under formats and transfer.