Black October, explainedAppsGuidesAI DictionarySmall FixesOn RecordAboutFollow the lab

Machine translation

Also called AI translation, Google Translate

Machine translation is a computer turning text or speech from one language into another, like Google Translate turning Hindi into English.

Inside language modelsEveryday

In plain words

Translating is not swapping words one by one. Every language has its own word order, its own idioms and its own ways of being polite. English says "I drink chai"; Hindi says "main chai peeta hoon", literally "I chai drink am".

Early translation programs did swap words, using dictionaries and grammar rules written by hand. Later systems learned from millions of sentences that people had already translated, matching phrases with phrases.

Today's systems are neural networks. They read the whole sentence first, build a sense of what it means, then write it fresh in the other language. That is why modern translations sound far more natural, even if they still slip on idioms, names and rare languages.

Everyday example

A friend who knows both languages

A friend from Chennai helps us read a Tamil wedding card. She does not translate it word for word. She reads the whole line, gets the meaning, then tells us in English: "They are inviting you to the reception on Sunday evening." Modern machine translation works the same way: understand the full sentence, then say it naturally in the other language.

Where the example breaks: Our friend knows the families, the customs and why the card is worded that way. The machine only knows patterns from text it was trained on, so it can miss cultural meaning or get a name wrong without noticing.

Go deeper

How it actually works

Machine translation is one of the oldest goals of natural language processing, and it has gone through three big eras.

  1. Rules (1950s to 1980s)

    On 7 January 1954, Georgetown University and IBM showed a computer translating about 60 Russian sentences into English, with a vocabulary of 250 items and just 6 grammar rules. Organisers predicted the problem would be solved in three to five years. It took decades.

  2. Statistics (1990s to 2016)

    Statistical systems learned, from huge collections of existing human translations, which phrase in one language most often lines up with which phrase in another. Google Translate launched on this approach in 2006.

  3. Neural networks (2016 onwards)

    An encoder-decoder network reads the whole source sentence, and attention lets it look back at the right words while writing each output word. In 2016 Google reported its neural system cut translation errors by an average of 60% on simple sentences compared with its older phrase-based one. The transformer (2017) was first built for translation.

  4. Large language models

    General large language models also translate well, and can adjust tone or explain a phrase. Google says Translate now uses its Gemini models. Projects such as Meta's No Language Left Behind (2022) push quality up for 200 languages, many with little digital text.

  5. How quality is measured

    A common score called BLEU compares the machine's output with human translations by counting overlapping word sequences. It is quick but rough, so serious evaluations also use human judges.

First public demo
Georgetown and IBM, January 1954: Russian to English on an IBM 701 computer
Google Translate
Launched 2006; switched to neural networks in 2016
Scale, as of April 2026
Google says Translate covers almost 250 languages and has more than 1 billion users a month
Indian languages
Google added 110 languages in June 2024, including Awadhi and Marwadi

Questions people ask

How does Google Translate work?

Since 2016 it has used neural networks trained on large amounts of translated text. The model reads your whole sentence, turns it into numbers that capture meaning, and writes a new sentence in the target language. Google says it now uses its Gemini models too.

Is machine translation accurate?

For common language pairs and everyday text, it is usually good enough to understand and often to use. It is weaker for idioms, slang, names, legal or medical wording, and languages with little online text. For anything important, have a fluent person check it.

Is ChatGPT better than Google Translate?

They are different kinds of tools. Google Translate is built for translation. Chatbots such as ChatGPT are general language models that can also translate, and can explain or change tone on request. Quality varies by language and text, so compare on your own examples.

Why does machine translation struggle with Indian languages?

Many Indian languages have far less digitised, translated text than English or French, so models see fewer examples. Mixed-language writing like Hinglish, different scripts, and politeness levels (aap and tum) add difficulty. Coverage has improved quickly in recent years.

What is BLEU score in translation?

BLEU is an automatic score from 0 to 1 (often shown as 0 to 100), created at IBM in 2002, that measures how many word sequences in the machine's translation also appear in reference translations by people. Higher is better, but it misses meaning and style, so it is used alongside human checks.

Further reading

Prepared by Mundhra Labs. Reviewed by Aditya Mundhra. Checked October 2026.