Machine translation means a computer translating text or speech from one language into another, as apps like Google Translate do. Today it is done with AI that learned from huge numbers of existing translations. It is fast and usually good enough to get the gist, but it can still drop words, garble names and miss the tone.
Early machine translation worked like a pocket phrasebook: look up each word, swap it, hope the grammar survives. Modern machine translation is more like a friend who has read millions of pages in both languages. They usually find a natural way to say it, but now and then they confidently say something that sounds fine and means the wrong thing.
In a sentence: “The menu was in Japanese, so I used machine translation on my phone to get the gist before ordering.”
What does machine translation mean?
The plain point: software, not a person, does the translating.
Translation is a classic job for natural language processing, the part of AI that deals with human language, and computers were being tried on it as early as the 1950s (see below). It shows up in translation apps, in “translate this page” buttons in web browsers, and in chatbots you can ask to translate a message.
How did machine translation change?
The plain point: it moved from hand-written rules, to word-and-phrase statistics, to AI that reads the whole sentence.
| Approach | How it worked | Everyday comparison |
|---|---|---|
| Rules | Worked by: grammar rules and dictionaries written by people. | Like: a phrasebook and a grammar book. |
| Phrase-based | Worked by: translating chunks of a sentence mostly one at a time. | Like: translating a sentence in pieces and gluing them together. |
| Neural | Works by: an AI model that takes in the whole sentence at once. | Like: reading the full sentence before saying it in your own words. |
In 1954, IBM had its first commercial scientific computer, the 701, translate Russian into English automatically. IBM’s own history calls the results “thrilling but far from perfect”.1
Later systems learned from existing translations instead. Google explained in 2016 that its older system broke a sentence “into words and phrases to be translated largely independently”, while its new neural system “considers the entire input sentence as a unit”. It reported that the switch cut translation errors by 55% to 85% on several major language pairs.2
A year later, the transformer, the design behind today’s chatbots, was first tested on translation.3
What does machine translation get wrong?
The plain point: the mistakes are rare words, names, missing words and hidden assumptions, and they read just as smoothly as the right parts.
Google said of its own neural system that it “can still make significant errors that a human translator would never make, like dropping words and mistranslating proper names or rare terms”.2
It can also pick up stereotypes. Some languages, such as Turkish, do not mark whether a person is male or female, so a translation into English has to guess. In 2018 Google described how its system had leaned on stereotypes, such as making doctors male, and began offering both a feminine and a masculine translation in some cases.4 AI bias meaning explains where habits like that come from.
Tone and jokes are hard too. IBM’s history notes that human translators use experience to carry over “context, idioms, humor and structure”.1
Machine translation is great for understanding a menu, an email or a web page. For anything where a wrong word costs money, health or legal trouble, have a person who knows both languages check it.
Often confused with
- Interpreting: translating speech live, out loud. Translation apps that listen first turn speech into text. See speech-to-text meaning.
- Localisation: adapting a product for another country, including prices, dates, pictures and jokes, not just the words.
- A chatbot: chatbots can translate too, using the same kind of language model. They can also rewrite, which means they may smooth over or change details.
Questions people ask
Is Google Translate accurate?
Usually accurate enough to understand something, but not guaranteed. Google itself has said its neural system can drop words and mistranslate names and rare terms, so check anything important with a person.
How does AI translation work?
An AI model learns from huge numbers of sentences already translated by people, then reads your whole sentence and writes it in the other language. It is predicting a likely translation, not looking each word up.
Learn more on this site
- Language and AI: natural language processing meaning.
- The engine inside: neural network meaning.
- Confident mistakes: hallucination meaning.
- The words: the AI glossary.
- Related: What Is Sentiment Analysis? Meaning in Plain English.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. How does modern machine translation work?
- It swaps each word using a pocket dictionary
- A person types every translation by hand
- An AI model trained on existing translations reads the whole sentence
Show the answer
Answer: An AI model trained on existing translations reads the whole sentence. Today’s neural systems learned from huge numbers of translations and take in the whole sentence at once.
2. Which mistake can machine translation still make?
- Dropping words or mistranslating names and rare terms
- Refusing to translate anything short
- Translating only into English
Show the answer
Answer: Dropping words or mistranslating names and rare terms. Google said its own neural system can drop words and mistranslate proper names or rare terms, and the errors read smoothly.
3. When should a person who knows both languages check a translation?
- Never, the app is always right
- When a wrong word could cost money, health or legal trouble
- Only when translating a menu
Show the answer
Answer: When a wrong word could cost money, health or legal trouble. Machine translation is great for the gist, but anything high-stakes should be checked by a person.
Related terms: Grounding · Knowledge cutoff · Latency · Multimodal · Prompt chaining · Prompt engineering
- 1IBM, Machine-aided translation, read at source 5 Oct 2026“In 1954, IBM introduced a translation solution that lessened the reliance on humans. The 701 Electronic Data Processing Machine, the company’s first commercial scientific computer, ran an experimental software program to automatically translate Russian into English.” … “The results were thrilling but far from perfect.” ↑a“Human interpreters use experience and instincts to reflect context, idioms, humor and structure in their translations.” ↑b
- 2Google Research, A Neural Network for Machine Translation, at Production Scale, 27 Sep 2016, read at source 5 Oct 2026“Phrase-Based Machine Translation (PBMT) breaks an input sentence into words and phrases to be translated largely independently” … “Neural Machine Translation (NMT) considers the entire input sentence as a unit for translation” … “GNMT reduces translation errors by more than 55%-85% on several major language pairs” ↑a“GNMT can still make significant errors that a human translator would never make, like dropping words and mistranslating proper names or rare terms” ↑b
- 3Vaswani et al., Attention Is All You Need, arXiv:1706.03762, submitted 12 Jun 2017, read at source 1 Oct 2026 (as cited on this site’s natural language processing page): “Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train.”↑
- 4Google Research, Providing Gender-Specific Translations in Google Translate, 10 Dec 2018, read at source 5 Oct 2026, which describes gender-neutral Turkish phrases about doctors coming out masculine and about nurses feminine, and the fix of providing “both feminine and masculine translations when translating single-word queries from English to four different languages (French, Italian, Portuguese, and Spanish)”.↑
Vaswani et al., Attention Is All You Need, arXiv:1706.03762, submitted 12 Jun 2017, read at source 1 Oct 2026.
The comparisons on this page (a phrasebook, a well-read friend, gluing pieces together), the descriptions of interpreting and localisation and the advice on when to use a person are this site’s own ways of explaining, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK