AI Translation Tools Data Leakage: Governance Guide

Machine translation is one of the most common ways confidential documents leave an organisation, because pasting a contract or uploading a board paper into a free translation website feels like using a dictionary rather than sharing data with a third party. Neural translation services are AI services: text and files are sent to the provider, processed, and under some consumer terms kept to improve the models. Translation is also built into browsers and office suites, so data can leave through a right-click menu nobody thinks of as an AI tool. This guide covers how the main translation services handle submitted content on their free and business tiers, the built-in browser translation features and the policies that control them, and a practical programme to give staff a safe route. It is for security, privacy and IT teams, and for legal and procurement teams that handle multilingual documents. Providers are listed alphabetically, and order is never a ranking.

Key takeaways

The tier decides the data handling: DeepL's free service may use submitted text to train its models for a limited period, and Microsoft's free consumer translation apps record text for service improvement, while the business services of DeepL, Google Cloud and Microsoft state that content is not used for training.
Documents are the high-value leak: whole-file translation sends complete contracts, HR files and technical documents in one action, so file upload to translation services deserves stricter control than short text.
Browsers translate too: Chrome and Edge offer page, selection and PDF translation, governed by the TranslateEnabled policy in each browser; decide deliberately whether to leave it on.
Saved content persists: glossaries, translation memories and custom model training documents are stored in the account until deleted, even on business tiers that do not retain translations.
Offer an approved service: blocking free tools without a sanctioned business alternative drives staff to phones and personal devices; an SSO-backed business account with a data processing agreement is the control that sticks.

Why translation is a distinct leakage route

Translation requests are unusual among AI uses: users nearly always submit complete source material rather than a question, the material is often among the most sensitive the organisation handles (contracts with foreign counterparties, regulatory correspondence, merger documents, patient letters), and the task is so routine that it rarely appears in AI risk assessments. Much of it also involves personal data, which raises a processor question under data protection law as well as a confidentiality one.

Route	Typical user action	What leaves the organisation	Primary control
Free translation website	Paste text or drag in a document	Full text or file, under consumer terms	Approved business service; web filtering; DLP on upload
Mobile translation app	Photograph a letter, speak a conversation	Images, audio and text, often from personal devices	Mobile policy and staff guidance; approved app with work account
Browser built-in translation	Accept the "translate this page" prompt or right-click a selection	Page or selection content, depending on how the browser performs the translation	Browser policy `TranslateEnabled`
Office suite translation	Translate a document or selection inside the editor	Document content to the suite's translation service	Tenant admin settings and the provider's enterprise terms
Translation API in an application	Developers wire a translation call into a product	Customer or internal data at scale	Approved API account, regional endpoint choice, key management
Writing assistants and AI chat	Ask a general assistant to translate	Text under that assistant's terms	See the Grammarly and AI Writing Assistant Security guide and the AI Training Opt Out Settings Guide

Personal data and confidentiality duties

When a translation provider processes personal data on the organisation's behalf, it acts as a processor, and data protection law expects a contract that limits what it does with that data. A free consumer service accepted by an individual employee gives the organisation no such contract, and a service that uses submitted text to improve its models is processing it for its own purpose. DeepL's own terms exclude personal data from its free service for this reason. Contracts with clients, non-disclosure agreements and professional secrecy rules (legal privilege, medical confidentiality) often prohibit disclosure to third parties without consent, and an uploaded file counts. Assess translation providers as you would any other processor: purpose, retention, location, subprocessors and deletion. See the AI and GDPR Guide for the wider legal frame.

How the main services handle submitted content

The statements below are paraphrased from each provider's own documentation as of November 2026. Terms change, so confirm the current wording and your contract before relying on them.

DeepL

Free Translator and Write: DeepL's privacy policy states that content submitted to the free service, its translations and user corrections are processed to train and improve its neural networks for a limited period. Document metadata such as file type, language pair and character count is deleted after 14 days. DeepL's terms do not allow the free service to be used for texts containing personal data.
DeepL Pro (Translator, API, Write): texts are not used to improve DeepL's services and are kept only as long as needed to produce the translation, then deleted with it. Document translation is processed on DeepL's infrastructure in the EEA. Saved translations, glossaries and translation memories stay in the account until deleted. A data processing agreement is available on request.
Admin controls: DeepL's Pro security page lists SSO with SAML and OIDC, multi-factor authentication for non-SSO users, role-based permissions, domain-based management, audit logs, bring your own key encryption and network access restrictions, and SOC 2 Type II.

Google

Consumer Google Translate: use of the free website and apps falls under Google's general Terms of Service and Privacy Policy rather than a business data processing agreement, so treat it as a consumer service for policy purposes.
Cloud Translation API: Google states that submitted text is held briefly in memory to perform the translation, is not used to train or improve Google's translation features, and is not shared with third parties. Glossaries are stored encrypted on Google Cloud and controlled by IAM.
Data location: the Cloud Translation Advanced API offers regional endpoints that can be configured for data location; the Basic and Advanced APIs on the global endpoint, and AutoML Translation, cannot keep data in a specific region. Choose the regional endpoint where residency matters.
Browser: Chrome's TranslateEnabled policy set to false shuts off all built-in translate features, true always offers translation, and unset lets users decide. Chrome 138 and later also expose a client-side Translator API to websites, with the language model downloaded to the device the first time a site uses it.

Microsoft

Consumer apps: Microsoft's No-Trace page states that its free consumer products (Microsoft Translator apps, Translator for Bing, Microsoft Edge translation and Web Translator) record voice clips and text translations for service improvement.
Azure Translator and Office: under No-Trace, data submitted to Azure Translator (standard and custom models), the Speech service and text translation features in Microsoft Office products is not written to persistent storage and is not used for training. Azure document translation stores files temporarily during processing and then hard-deletes them.
Custom Translator: documents uploaded to train custom models are stored encrypted in the workspace until deleted or the account expires; translation requests to the custom model remain No-Trace.
Browser: Edge's TranslateEnabled policy (registry value under SOFTWARE\Policies\Microsoft\Edge, REG_DWORD) turns off all built-in translate features when disabled, including page, selection and PDF translation. It is enabled when not configured, and Microsoft notes it does not apply to profiles signed in with a Microsoft account.

Availability notes: DeepL, Google and Microsoft each separate free consumer offerings from paid business or cloud services, and each documents its business terms on the pages listed in the references. Region, encryption and retention options depend on the plan and endpoint chosen.

Choosing an approved route

Need	Suitable approach	Conditions to set
Everyday text and document translation for staff	Business subscription to a translation service, tied to SSO	Data processing agreement; no training on content; retention and saved-content rules; audit logs on
Translation inside documents already in the office suite	The suite's own translation under its enterprise terms	Confirm the feature is covered by the enterprise data terms, as Microsoft states for Office under No-Trace
Translation in products or pipelines	Cloud translation API on an organisation-owned account	Regional endpoint where needed; API keys or workload identity managed like any other credential
Highly sensitive material (legal privilege, M&A, health records)	Professional human translation under contract, or a self-hosted model	Same handling as the source document's classification
Reading foreign-language websites	Browser translation, if the organisation accepts it	Decide the `TranslateEnabled` setting by user group

Small organisations: one business translation subscription with SSO and a one-line rule ("never use free translation sites for work documents") covers most of the risk at low cost.
Large or regulated organisations: add web filtering of free translation domains, DLP rules for file uploads, browser policy by group, and procurement review of each provider's processing locations and subprocessors.
Developers: translation API keys are credentials. Store them in a secrets manager, scope them to the translation service, and rotate them; see the API key leak response playbook if one is exposed.

Implementation in five steps

Find current use. Query proxy, secure web gateway or CASB logs for translation domains and file uploads to them, and survey legal, HR, sales and procurement teams who handle foreign-language material. The Shadow AI Discovery Guide covers discovery methods.
Procure the approved service. Choose a business tier whose documentation states no training on content and deletion after processing, sign the data processing agreement, connect SSO, and set retention for saved translations, glossaries and translation memories.
Set the rule. Add translation explicitly to the AI Acceptable Use Policy Template: approved services by data class, a ban on free tiers for work content, and what to do for highly sensitive documents.
Enforce gradually. Start with warn pages on free translation domains, pointing to the approved service, then move to block for file uploads once the approved route is in use. Add DLP rules for classified labels on uploads, as described in the Generative AI Data Loss Prevention Guide. Pilot with one department; roll back from block to warn if a business process breaks.
Configure browsers. Set TranslateEnabled in Chrome and Edge as a mandatory policy per user group, test on a pilot ring, and document the decision. See the Browser Security Hardening Guide for policy deployment.

Detection and measurement

Web and CASB logs: visits and uploads to free translation domains, by department and by user, with file names or labels where your tooling captures them.
DLP events: sensitivity-labelled files or matched content sent to translation services, including from the browser's upload dialog.
Provider audit logs: sign-ins and usage on the approved service, which show whether staff are moving to it.
Browser policy reports: devices where TranslateEnabled is not set as intended.

Metric	Target	Why it matters
File uploads to free translation services per month	Falling to near zero	Whole-document leaks are the main exposure
Active users on the approved translation service	Rising until it matches discovered demand	Shows the safe route is used, not avoided
DLP matches on translation destinations	Falling month on month	Labelled or regulated data still reaching translation tools
Managed browsers with the intended translate policy	100%	Policy drift reopens a route
Translation providers with DPA and terms review on file	100%	Contractual protection for content processed

Common failure patterns

Translation left out of AI policy: the AI use rule names chat assistants but not translation, so staff do not see free translation sites as covered.
Business tier bought, free tier still used: staff stay on the free website out of habit or because their bookmark logs them out, and the business account sits idle.
Glossaries and memories forgotten: client names, product codenames and contract terms accumulate in saved glossaries and translation memories with no retention rule or owner.
Phones as the workaround: blocking the desktop route without an approved alternative pushes staff to photograph documents with a personal translation app.
Residency assumed, not configured: a cloud translation API is used on its global endpoint although the data must stay in a region the regional endpoint would support.
Developer keys in code: translation API keys committed to repositories or shared between applications, with no owner or rotation.

Framework mapping

Framework	Reference	Relevance
NIST CSF 2.0	ID.AM-03; GV.SC-05; PR.DS-02	Map translation data flows; put training and retention terms in provider agreements; protect data sent to external services
NIST SP 800-53 Rev 5	SA-9 External System Services; AC-20 Use of External Systems; AC-4 Information Flow Enforcement; SC-7 Boundary Protection	Approve translation providers; restrict free services; enforce DLP and web filtering on uploads
ISO/IEC 27001:2022 Annex A	A.5.14 Information transfer; A.5.23 Information security for use of cloud services	Rules for sending documents to external services; cloud service terms and selection
MITRE ATT&CK	T1567 Exfiltration Over Web Service	Monitoring for uploads to web services also catches deliberate misuse of translation sites

Practitioner checklist

Inventory translation use from web logs, including file uploads, and from the teams that handle foreign-language material.
Procure a business translation service with SSO, a data processing agreement and documented no-training terms.
Set retention and ownership for saved translations, glossaries and translation memories.
Name translation explicitly in the AI acceptable use rule, with a route for highly sensitive documents.
Warn, then block, file uploads to free translation sites, with DLP on labelled content.
Set TranslateEnabled in Chrome and Edge as a deliberate, documented decision per user group.
Use regional endpoints for translation APIs where residency applies, and manage API keys as secrets.
Review provider terms and subprocessors annually and when plans change.

AI Translation Tools Data Leakage Guide: Governing DeepL, Google and Microsoft Translation at Work