AI Translation — How Machines Translate Between Languages

Status: 🟩 COMPLETE 🟦 LIVING Section: 10 — AI and LLMs Tags: translation, ai-translation, deepl, google-translate, machine-translation, NMT, multilingual

What it is

AI translation is the use of machine learning to automatically translate text (or speech) from one human language to another. “Bonjour, comment ça va?” → “Hello, how are you?”

Modern AI translation — since roughly 2017 — is dramatically better than older rule-based or phrase-matching systems. Today’s tools can:

Produce translations that read like a human wrote them
Handle idiomatic expressions and cultural references
Maintain document formatting (tables, headers, bullet points)
Translate entire websites, documents, emails, or conversations in seconds
Work in 100+ language pairs

It is now accurate enough to be used for most everyday and business purposes without human review — though high-stakes publishing (legal contracts, medical instructions, literature) still benefits from human post-editing.

How it works (plain English)

Modern AI translation uses the same transformer architecture that powers GPT and other language models (see how-llms-work).

In simple terms:

The model is trained on enormous datasets of parallel text — the same document in two (or more) languages, side by side. Think EU Parliament records (which are translated into 24 languages), Wikipedia, news articles, books, and websites.
The model learns the statistical relationships between language patterns. “When I see this French sentence structure, the equivalent English structure tends to be…”
When you give it a French sentence, it doesn’t look up words in a dictionary. Instead, it generates the most likely English translation given everything it knows about how French and English relate to each other.
Context matters: “Bank” in English can mean a riverbank or a financial bank. AI translation uses the surrounding sentence to pick the right equivalent in the target language.

Neural Machine Translation (NMT): The technical term for this deep-learning approach to translation, which replaced earlier Statistical Machine Translation (SMT) and rule-based systems. NMT produces dramatically more fluent output.

The major AI translation tools (mid-2026)

Dedicated translation tools

Tool	Country	Best for	Free tier?
DeepL	🇩🇪	Highest quality for European languages; document translation; natural-sounding output	Yes (limited)
Google Translate	🇺🇸	243 languages; widest coverage; instant website translation; free	Yes (full)
Microsoft Translator	🇺🇸	Office/Teams integration; Azure API	Yes (free up to 2M chars/month)
Amazon Translate	🇺🇸	AWS-integrated; scalable; developer API	Pay-per-character
Systran	🇫🇷	Enterprise; domain-specific models (legal, medical)	Enterprise pricing
ModernMT	🇮🇹	Adaptive; learns from translator corrections	Enterprise

Built into everyday tools

Google Chrome — auto-translate any webpage (right-click → Translate)
Microsoft Edge — auto-translate on load
Word / PowerPoint — Translate tab in Review ribbon
Gmail — Translate Message button in emails
iOS / Android — translate system-level (Translate app on iPhone)

Inside LLMs (translation via conversation)

ChatGPT, Claude, and Gemini can translate text directly. For individual sentences or paragraphs, this is often excellent. Benefits:

Can explain why a translation was chosen
Can adjust tone (formal/informal)
Can localise culturally (not just translate, but adapt)
Can handle unusual or domain-specific text

Limitations compared to dedicated tools:

Can’t process whole documents with formatting preserved
Slower for batch translation
Less consistent across long documents

Australian translation needs

Australia is highly multilingual. The most-needed languages beyond English include Mandarin, Arabic, Cantonese, Vietnamese, Italian, Greek, Tagalog, Hindi, Punjabi, and Spanish. All are well-supported by DeepL and Google Translate.

For Aboriginal and Torres Strait Islander languages: AI translation is extremely limited. These languages have very little digital text available for training, and most have unique grammatical structures not well-represented in standard NMT systems. Human interpreters remain essential.

Key concepts

Source language: The language you’re translating FROM.

Target language: The language you’re translating TO.

Fluency vs accuracy: Two different quality dimensions:

Fluency: Does the translation read naturally in the target language? Does it sound like a native wrote it?
Accuracy / fidelity: Does it actually convey the same meaning as the source? A fluent translation can still get the meaning wrong.

Back-translation: Translating the translated text back into the original language to check if the meaning is preserved. A rough quality-check technique.

Post-editing (MTPE): Human translators reviewing and correcting machine translations. Typically 40–60% faster than translating from scratch. The industry standard workflow for high-volume professional translation.

Glossary / terminology management: In business and technical translation, specific terms must always be translated consistently (e.g., a product name, a legal term). Professional translation tools let you define a glossary of required translations.

Domain-specific models: General models struggle with highly specialised vocabulary — medical, legal, engineering, pharmaceutical. Specialised models (or fine-tuned LLMs) trained on domain text perform significantly better.

BLEU score: The traditional metric for measuring translation quality — compares AI output to human reference translations. Increasingly replaced by neural quality metrics that better capture fluency.

Where AI translation excels

Information access: Read foreign-language news, websites, social media, research papers, instructions.
Email and communication: Send and receive business emails across languages without a translator.
Customer support: Route and respond to support requests in any language.
E-commerce: Translate product listings and descriptions for global markets.
Travel: Real-time translation in conversation (Google Translate conversation mode, Apple Translate app).
Content localisation (with human review): First pass of website, app, or marketing material translation.
Document translation: Contracts, reports, manuals — DeepL and Google support full document upload with formatting preserved.
Subtitles and captions: Auto-translate video captions into many languages.

Where AI translation still struggles

Literary and creative text: Wordplay, poetry, humour, cultural allusions, and stylistic voice are difficult. A pun in French may have no equivalent in Vietnamese.
Legal precision: Legal language is highly technical and jurisdiction-specific. Mistranslation of a single term can have serious legal consequences. Human review is essential.
Medical instructions: Dosage instructions, drug interactions, informed consent — too high-stakes for unsupervised AI translation.
Low-resource languages: Languages with little digital text available for training (including many African, Pacific, and Indigenous languages) have much lower quality.
Dialects and regional variation: Brazilian Portuguese vs European Portuguese; Mandarin vs Cantonese; Argentine Spanish vs Mexican Spanish — quality varies.
Code-switching: Text that mixes two languages (common in informal writing in multilingual communities) is still challenging.
Context across very long documents: For a 100-page document, early context may be lost in later pages, causing inconsistency in terminology.

Gotchas

DeepL outperforms Google Translate for European languages in quality studies — but Google has far wider language coverage (243 vs 33 languages for DeepL’s free tier).
Don’t use AI translation for legal or medical documents without human review. The consequences of errors are too serious.
“Free” Google Translate sends your text to Google’s servers. For confidential business documents, use DeepL Business (with data processing agreement) or on-premise solutions.
Tone and formality: AI often defaults to a formal register. If you need informal/conversational text, specify in your prompt (or use an LLM that lets you specify).
Translating AI outputs: If you’re using an LLM (like Claude or ChatGPT) to create content and then translating it, tell the LLM your target language upfront — it may do better translating as it writes than having DeepL translate the English output afterwards.
Right-to-left (RTL) languages: Arabic, Hebrew, and Persian are written right-to-left. Translation software handles this; but if you paste translated text into a design tool or webpage, you may need to set the language direction manually.
Australian context: If translating for Australian multicultural communities (health information, government services), use a professional translator who understands the Australian context. Cultural nuance matters as much as linguistic accuracy.

Real-world workflow example

A Melbourne restaurant wants a Vietnamese menu:

Write the menu in English.
Run through DeepL (best quality for Vietnamese → reasonable).
Have a native Vietnamese speaker do a 30-minute review for food names, local expressions, and natural phrasing.
Use the reviewed version.

This is MTPE (machine translation post-editing) — the industry standard. Faster and cheaper than full human translation, still high quality.

Sources

DeepL product documentation and language support (2024–2026)
Google Translate supported languages and API documentation
Microsoft Translator Azure Cognitive Services documentation
TAUS industry translation quality reports (2024)
Koehn, P. — “Statistical Machine Translation” (Cambridge, 2010; still foundational)
Vaswani et al., “Attention Is All You Need” — the transformer paper that transformed NMT (2017)
NAIDOC — guidance on Aboriginal language preservation (2024)
Australian Institute of Interpreters and Translators (AUSIT) — AI translation policy statement

Tech & AI, Explained

Explorer

ai-translation