Last updated: 2026-07-13

Every time you paste a paragraph into Google Translate or ask ChatGPT to translate a document, that text leaves your machine and lands on a server you don't control — often to be logged, and in some cases used to improve the underlying model. For a restaurant menu, that's irrelevant. For an employment contract, a medical record, an immigration filing, or a client deliverable under NDA, it's a data-handling decision you're making without thinking about it.

This guide covers who actually needs to care about this, why the popular translation tools are the wrong default for sensitive material, and how to build a translation workflow that keeps the source document, the translated output, and everything in between off someone else's servers.

Why Document Translation Is a Bigger Privacy Problem Than Chat

Translating a sentence in a chat window and translating a document are not the same risk. When you type a question into an AI chatbot, you're usually sharing a fragment — a thought, a query. When you translate a document, you're uploading the entire thing: names, account numbers, medical diagnoses, salary figures, case details, signatures, letterhead. Translation tools need the full context to produce an accurate result, so there's no way to redact your way around it and still get a usable translation.

A few examples of documents people run through free translation tools without a second thought:

Employment contracts and offer letters for a job in another country
Medical records and lab results for care abroad or for a second opinion
Immigration paperwork — birth certificates, marriage certificates, visa applications, sponsor letters
Client contracts and deliverables for consultants and agencies working with international clients
Financial statements for cross-border business, tax filings, or loan applications
Legal correspondence — demand letters, court filings, discovery documents

Each of these has a real cost if it leaks: identity theft, an immigration case compromised by an intercepted document, a client NDA violated because a "free" translation tool's terms of service permit human review of submitted content. And the tools most people reach for by default were never built with that threat model in mind.

What Google Translate, DeepL, and ChatGPT Actually Do With Your Text

None of these are surveillance tools by design — but their default terms are built around a free consumer product, not confidential document handling.

Google Translate: Text submitted through the free web interface can be logged and used to improve Google's translation and AI systems. Google's privacy policy applies the same broad data-usage terms it applies to Search and Gmail. There's no confidentiality agreement, no data processing addendum, and no way to guarantee deletion.

DeepL: The free tier's terms permit DeepL to store and use submitted text for a period of time, and historically for training on the free consumer product. DeepL Pro (paid) adds a "no data retention" option, which is a meaningfully better posture — but you have to know to enable it, and it still means your document reaches DeepL's servers even if it isn't retained afterward.

ChatGPT / Claude / Gemini (consumer tiers): Pasting a document into a general-purpose chatbot for translation means the full text becomes part of that conversation's history, subject to whatever retention and training settings are active on your account. Most people never check.

The pattern across all three: your document leaves your device, and the guarantee that it isn't stored, read, or used somewhere downstream depends entirely on a terms-of-service page you didn't read. For a menu, fine. For a lease agreement or a lab result, that's a real exposure.

The Local Alternative: Translation That Never Leaves Your Machine

The fix isn't complicated — it's translating documents with a model that runs on your own hardware, so the text never crosses the network at all.

Option 1: Ollama with a Multilingual Model

Ollama runs open-weight language models locally on macOS, Linux, or Windows. Several current open models handle translation well across dozens of languages, including Qwen2.5 and Aya Expanse, which were specifically trained for strong multilingual performance.

```bash

Install Ollama

curl -fsSL https://ollama.com/install.sh | sh

Pull a multilingual-strong model

ollama pull qwen2.5:14b

Run it

ollama run qwen2.5:14b

```

Once it's running, you can paste text directly into the terminal or, more comfortably, pair it with Open WebUI for a browser-based interface at localhost. A translation prompt looks like this:

```

Translate the following document from Portuguese to English.

Preserve the original formatting, paragraph breaks, and any

legal or technical terminology as precisely as possible. Do

not summarize or omit any section.

[paste document text]

```

For longer documents — a multi-page contract or a full medical record — break the document into sections and translate each one, then reassemble. Local models have smaller context windows than the frontier cloud models, so this extra step matters for documents over a few thousand words.

Hardware reality check: A 14B parameter model runs comfortably on an M1 Mac or better, or a Windows/Linux machine with 12GB+ of VRAM. You don't need specialized hardware — a mid-range laptop from the last three years is usually enough.

Option 2: LibreTranslate (Self-Hosted, Dedicated Translation Engine)

If you'd rather use a tool built specifically for translation instead of a general-purpose LLM, LibreTranslate is a free, open-source translation engine you can self-host with Docker:

```bash

docker run -d -p 5000:5000 libretranslate/libretranslate

```

That gives you a private translation API and web interface running entirely on your own machine or home server. It supports dozens of languages out of the box and integrates into scripts if you need to batch-translate a folder of documents. It's less nuanced than a large language model for idiomatic or legal phrasing, but it's fast, lightweight, and genuinely private — nothing leaves your network.

Which One to Use

For contracts, medical records, and anything where phrasing precision matters, use the Ollama + multilingual model route — LLMs handle nuance, tone, and technical terminology noticeably better than dedicated translation engines. For high-volume, lower-stakes batch translation (internal notes, bulk document triage before a human reviews the important ones), LibreTranslate is faster to set up and lighter on hardware.

What Happens Before and After the Translation Matters Too

Running the translation locally solves the biggest leak, but the document has a life before and after that step — and that's where most privacy-conscious workflows quietly fall apart.

Where the source document lives: If your contract or medical record sits in a regular Google Drive or Dropbox folder before you ever translate it, you've already exposed it to a provider with broad data-usage terms. Proton Drive provides end-to-end encrypted cloud storage under Swiss privacy law — even Proton can't read the files you store there. Keeping source documents and translated outputs in an encrypted vault closes the gap that a private translation step alone doesn't cover.

Affiliate Disclosure: This article may contain affiliate links. If you make a purchase through these links, we may earn a small commission at no extra cost to you. We only recommend products we genuinely believe in. This helps support our work and allows us to continue providing free content.

Researching terminology without exposing the document itself: Legal and medical translation often requires looking up a specific term, a certification requirement (is this translation "certified" for USCIS purposes?), or a formatting convention for a target country — without pasting the sensitive document into a search engine. Perplexity Pro is well-suited to this narrower task: you ask a general question ("what format does USCIS require for certified translations of foreign birth certificates?") without ever sharing the document you're actually working on. Keep the research query generic and the source material stays local.

A Complete Private Translation Workflow

Putting it together, a privacy-respecting document translation workflow looks like this:

Store the source document in encrypted storage (Proton Drive or Tresorit) the moment you receive it — not a general-purpose cloud drive.
Translate locally with Ollama and a multilingual model (Qwen2.5 or Aya Expanse) for anything where phrasing and terminology matter, or LibreTranslate for high-volume, lower-stakes batches.
Research unfamiliar terms or formatting requirements with a privacy-conscious research tool like Perplexity Pro, keeping queries generic rather than pasting document excerpts.
Review the output against the source for accuracy — local models are good, not infallible, and legal or medical translations should get a human review pass regardless of which tool produced the draft.
Deliver the translated document through encrypted, access-controlled sharing (Tresorit) rather than a plain email attachment, especially for anything covered by NDA or privilege.
Delete the working copy from any temporary location once the final version is stored and delivered.

None of these steps require specialized skills or expensive tools — the entire local translation setup takes about fifteen minutes, and the encrypted storage and sharing tools have free tiers to start with. The only real change is a habit: routing sensitive documents through a workflow you control, instead of the free tool that happens to be a browser tab away.

When It's Fine to Use the Free Tools

Not every translation needs this level of care. A restaurant menu, a product review, a casual message from a pen pal — there's no meaningful downside to running that through Google Translate. The dividing line is simple: if the document contains information you wouldn't want stored indefinitely on a server you don't control, or that's covered by a confidentiality obligation (NDA, attorney-client privilege, HIPAA-adjacent medical context, immigration proceedings), route it through a local workflow instead. Everything else, use whatever's convenient.

Want the exact Ollama model recommendations for translating into 12 common languages, plus a copy-paste prompt template that preserves legal and medical formatting? Subscribe below and we'll send the full reference guide straight to your inbox — no account required, no tracking pixels.

Subscribe for the Private Translation Reference Guide →