What the engine does with your words.

These terms govern every AI-assisted feature in the Service: machine translation, automated quality review, glossary extraction, in-context suggestions including the ones that read a screenshot, the assistant, and anything else we label as AI. They are part of the Terms of Service and are read together with the Privacy Policy, the Data Processing Addendum and the Acceptable Use Policy. Where they conflict with the Terms of Service on an AI-specific question, these terms win.

"AI Feature" is any feature of the Service that uses a machine learning model. "Input" is the Customer Content, context and instructions sent to an AI Feature: source strings, key names, file paths, glossary entries, style guides, comments, screenshots and the surrounding strings that give a key its meaning. "Output" is what the AI Feature returns: translations, quality findings, severity scores, glossary candidates, suggestions and explanations. "Engine" is the model configuration a translation runs on. "Glossa" is the name of our engine suite; "Flux" is the fast lane, available on every plan including Free, and "Deep" is the higher-effort lane, available from the Team plan.

When a key is pushed or edited, the Service assembles a batch: the source strings, the surrounding context that gives them meaning, your glossary terms, your style guide and the target locale. It resolves what it can from translation memory without leaving our systems. It sends the remainder, as structured batches, to the engine, which is operated by the subprocessors disclosed by category in our Trust Center and named in the list we furnish on request under a non-disclosure agreement. It then runs the quality gates over the result, applies at most one corrective retry to a row that fails, and blocks the row rather than shipping it if the retry also fails. There is no separate "send to translation" step: the engine runs on the push.

You keep every right you have in the Input. As between you and us, you own the Output produced from your Input, and to the extent we hold any right in it we assign that right to you, so a translation the engine produces for you is yours to use, modify, publish, license and sell without any further payment or attribution to us. We take no ownership interest in your translated content, your translation memory or your glossaries. The only licence we take over Input and Output is the operational one in the Terms of Service: hosting, processing and transmitting them to provide the Service to you.

First, copyright law in several jurisdictions does not protect text generated by a machine without sufficient human authorship, and no contract can change that. We make no representation that Output is protectable by copyright, and if that matters to you, a human review pass is what creates the authorship. Second, Output is produced statistically from Input: the same or similar Input can produce the same or similar Output for a different customer, and a common product string ("Save", "Your cart is empty") will translate the same way for everybody. Output is therefore not exclusive to you, and that non-exclusivity is not a breach of confidentiality, of this agreement, or of any duty we owe you.

We do not use Customer Content, Input or Output to train, fine-tune, evaluate or otherwise develop any machine learning model, on any plan, and we do not permit our engine vendors to do so: they process your text for inference only, under commercial terms that prohibit training on it. We do not offer an opt-in that would change this and we are not holding it back as a paid tier. We may use aggregated statistics that identify neither you nor any individual (run counts, latency, gate failure rates, reuse ratios) to operate and improve the Service, and those statistics contain no strings.

The subprocessors that operate our engines are disclosed in the Trust Center by category, with the purpose, the region and the condition under which each is engaged, and they are restated in the Subprocessors document; the vendors themselves are named in the list we furnish on request under a non-disclosure agreement. That disclosure is enforced rather than promised: our build fails if a translation provider becomes reachable in a deployment without a matching disclosure row, so routing your text to an undisclosed vendor is not a thing an operator can quietly do. Some engines are conditional and off by default. We give notice before a new subprocessor starts processing your content, and the Data Processing Addendum gives you a right to object.

We select the engine that a tier runs on, and we may change it to improve quality, reliability, cost or latency, subject to the disclosure obligation in clause 07. Flux and Deep describe the lane rather than a named third party model, deliberately, because the underlying model is an implementation detail that will change and the guarantee we make about the lane will not. We do not today offer a per-project choice of model vendor; if you need one contractually, raise it before you sign an Order Form rather than assuming it.

The Service maintains two kinds of memory and they behave differently. A human-reviewed entry is scoped to your organization, is used for near-match lookup within your organization only, and is never returned to another workspace. An exact-match entry is contributed to a shared pool that stores the source and its translation keyed by a hash of the source, together with plural forms and reuse counters, and stores no user column and no team column at all, so nothing in it can be traced to a person. The key name, the file path, screenshots, comments and every other piece of surrounding context are stripped. Any project can leave the pool with a per-project switch, on every plan including Free, which stops it both contributing and drawing. We disclose the pool rather than burying it, because a shared memory is a real trade and you should be able to decline it.

Glossary terms and style guides you create belong to your organization and are used only for your organization’s translations. The engine is instructed with them and the gates enforce them, so a term you mark as required is checked in the output rather than merely suggested to the model. Glossary extraction proposes candidate terms from your own content; a proposal is a suggestion, and nothing enters your glossary until you accept it.

Every translated row passes structural checks before it can ship: placeholders and interpolation tokens must survive the translation intact, plural categories must match what the target language requires, required glossary terms must appear, and length limits you set must be respected. A row that fails gets one automatic corrective retry, and if it fails again it is blocked rather than shipped. That is a warranty about mechanics, and it is the only warranty we make about a translation. It is not a warranty that a translation is accurate, idiomatic, culturally appropriate, legally sufficient in any market, or fit for a regulated purpose.

Do not ship unreviewed machine output as medical instructions, drug labelling, safety warnings, legal notices, contractual terms, regulatory filings, financial disclosures, or anything else where an error can hurt a person or breach a rule. Route those strings through human review, keep the reviewer in the workflow rather than in the plan, and treat the gates as a floor rather than as a sign-off. The Service does not produce a certified, sworn or notarised translation, and no output from it should be presented as one.

Usage is metered in words and charged against your plan allowance. The unit is a source word multiplied by each target locale it is translated into, so one hundred source words into five languages is five hundred words of usage. Counting is script-aware: a language written without spaces is counted on a basis appropriate to its script rather than by splitting on spaces, so that a Japanese string is not counted as one word. A row resolved from translation memory without calling the engine does not consume allowance. A re-translation consumes allowance again, because it is another call. Quality review, glossary extraction and in-context suggestions consume allowance on the same basis as the translation they belong to. The allowance resets each billing cycle and does not roll over, and the console shows consumption against it in real time.

Machine translation models can produce text that is fluent and wrong. They can invent a term, mistranslate a negation, resolve an ambiguous pronoun the wrong way, choose the wrong register for a formal market, or reproduce a bias present in the language they learned from. The gates catch structural failures, not semantic ones, and a wrong translation that keeps its placeholders will pass every gate we run. You are responsible for reviewing Output before you rely on it, and for the consequences of publishing it. We do not warrant that Output is free of errors, and clause 32 of the Terms of Service says so in the formal register.

Where the EU Artificial Intelligence Act applies, we consider the Service to be a general purpose AI-based tool used for language translation, and not a high-risk AI system under Annex III as we supply it, because it does not itself determine access to education, employment, essential services, credit, law enforcement outcomes, migration decisions or the administration of justice. That classification depends on how you deploy it: if you build the Service into a process that does make such a determination, you may become the provider or the deployer of a high-risk system and you take on the obligations that follow. Tell us if you intend to, because we may not be able to support that use.

Content produced by an AI Feature is artificially generated, and where the law requires it you must disclose that to the people who read it. We help by keeping the machine origin of a string visible inside the product: a translation carries the state it was produced in, review status is recorded per string, and the version history shows what the machine wrote and what a human changed. You decide what your own product tells your users, and you must not represent Output as having been produced or verified by a qualified human translator when it has not been.

You are responsible for making sure the people in your organization who use AI Features have sufficient understanding of what the engine does, what it cannot do, and how to check it, in proportion to how you use it. We support that with published documentation, in-product explanations of the gates and their findings, per-string version history, and the review workflow. Every AI Feature in the Service is a suggestion to a human rather than an autonomous action: a translation can be edited, a suggestion can be rejected, a gate finding can be inspected, and a version can be restored.

Translation memory contribution and lookup can be switched off per project on every plan. Semantic (near-match) memory is off by default and is only engaged when a project turns it on. The in-context editor, including anything that reads a screenshot, is opt in and needs a key you create. The assistant is optional. Machine translation itself is the product, so switching it off means not translating; if you want the workflow without the engine, you can import and manage translations produced elsewhere.

The Service can be configured to fall back to a third party machine translation vendor using an API key you supply. That lane is off by default and is not used unless you enable it. Because the key and the relationship are yours, you are the controller of that transfer and the vendor is your processor rather than our subprocessor, and their terms and privacy policy govern what they do with the text. If we ever ship that lane with a key of our own or turn it on by default, it becomes our subprocessor and appears on the published list before it processes anything.

The in-context editor can capture a screenshot of the screen a string appears on, so that the engine and the reviewer can see the context the string lives in. A screenshot is Customer Content: it may contain personal data or confidential information from whatever was on the screen, and you are responsible for what you capture. Screenshots are held on a private store with no public path, served only through short-lived signed links, and deleted after 90 days by a scheduled job that removes the stored file as well as the record. Capture is opt in, and a project that never enables the in-context editor never stores one.

An AI Feature labelled beta, preview or experimental is provided as is, may change or be withdrawn without notice, is excluded from every service level commitment and from our intellectual property indemnity, and may run on a different engine or a different configuration from the general one. We tell you it is a beta at the point you turn it on. Do not put a beta AI feature on a critical path.

The Acceptable Use Policy governs, and these are the AI-specific applications of it: do not use an AI Feature to generate content that is unlawful, infringing, deceptive or abusive; do not use it to produce impersonations, fake reviews, coordinated inauthentic messaging or political manipulation at scale; do not use Output to train or evaluate a competing translation or language model; do not attempt to extract, reconstruct or reverse engineer our prompts, model configuration, quality checks or system instructions; and do not present Output as human-authored or human-verified when it is not.

We change engines, prompts, quality checks and routing as the field moves, and we do not promise that a translation produced today will be byte-identical to one produced tomorrow from the same source. Version history is what protects you from that: a string you have reviewed and approved stays as you approved it, and a re-translation is an action you take rather than one that happens to you. Material changes to how a plan-level lane behaves are announced under the notice provisions of the Terms of Service, and a new subprocessor is announced under clause 07.

Questions about these terms, about how a run is routed, or about an engine commitment you need in writing go to hello@transglot.ai. Data protection questions about AI processing go to privacy@transglot.ai. The engine page explains the mechanics in product language, and the Trust Center carries the disclosure list this document relies on.