Aiodic

Back to the table

35 Intermediate Tn Translation Applications

Translation

Converting text between languages.

What it is

Translation is turning words in one language into words in another while keeping what they mean. That sounds simple until you try it. A good translation doesn't swap words one for one. It carries over the meaning, the tone, the level of politeness, the jokes where possible, and the feeling a sentence gives a reader, and it does all that in a language with different grammar, different habits and a different culture behind it.

Machines have been attempting this for more than seventy years, and for most of that time they did it badly. Early systems followed hand-written rules: a dictionary plus instructions such as "in Russian this word comes before that one, so swap them in English". Later systems worked statistically, learning from millions of sentences that humans had already translated and stitching together the most likely phrases. Both approaches produced the stiff, occasionally absurd output that made machine translation a punchline.

The change came with neural networks, which read a whole sentence before writing the translation instead of working through it phrase by phrase. The transformer, the design behind today's chatbots, was invented in 2017 specifically for translation. Today there are two main kinds of AI translator:

That second point is what the popup text on this element is about. A chatbot can translate the same paragraph three ways: formally for a contract, warmly for a newsletter, and casually for a text message. Linguists call this register, meaning the level of formality you choose for an audience. Many languages build register into their grammar. French, German and Spanish have separate polite and familiar words for "you", and Japanese has whole systems of polite verb forms. A good AI translator asks itself which one fits, or lets you tell it.

How good is it now? In the 2024 Conference on Machine Translation (WMT), the main yearly contest in the field, professional human annotators marked the errors in translations from dozens of systems across 11 language pairs. The organisers named a general-purpose chatbot model, Claude 3.5 Sonnet, the best performing system overall: it was in the winning group for 9 of the 11 pairs, more than any system built only for translation. The organisers titled their report "The LLM Era is Here but MT is Not Solved Yet", and the second half of that title matters as much as the first. Quality is uneven. It's highest for widely used language pairs, like English and Spanish, which have huge amounts of translated text to learn from, and much lower for languages with little text online. Even in the best pairs, errors still slip through, and some of them change the meaning completely.

Why it matters

It opens up readers and customers you couldn't reach before. For most of history, translation was something only publishers and large companies could afford. A book translation by a professional literary translator is a serious investment, often priced by the word. AI makes a usable first draft almost free, which is why self-published authors, small online shops and course creators are now asking whether they can finally reach readers in German, Spanish or Japanese. Amazon has said that historically fewer than 5 percent of titles on Amazon.com were available in more than one language. That gap is the opportunity.

It's the everyday version of a hard problem. Translation touches almost everything else on this table. It depends on tokens, which is why the same text costs more to process in some languages than others. It depends on datasets, which is why languages with little written material online are served worse. It inherits bias: if a language has no gendered pronouns and English does, the model has to guess, and its guesses tend to follow stereotypes about who is a doctor and who is a nurse. And it can hallucinate, adding a sentence that wasn't there or smoothly dropping one that was.

Fluent isn't the same as correct. This is the most important thing to understand about AI translation. Older systems failed loudly: the output was clumsy, so you knew not to trust it. Modern systems fail quietly. The translation reads beautifully, and one sentence in fifty says something the original didn't. If you can't read the target language, you have no way to notice. The more polished the tool, the more a human check matters, and the more it matters who does the checking.

The stakes vary enormously. A menu, a friendly email or a product description with a small slip is embarrassing at worst. A medical instruction, a legal statement, a contract or a safety warning with a small slip can hurt someone. The same tool can be perfectly fine for one and irresponsible for the other. Deciding which is which is now part of the job.

It changes the work of human translators rather than simply removing it. In a January 2024 survey of its members by the Society of Authors, the UK trade union for writers and translators, about a third of the translators who replied said they had already lost work to generative AI. At the same time, a growing share of paid translation work is post-editing, where a professional corrects and polishes a machine draft instead of translating from scratch. For authors, that creates a middle option between "do it all with AI" and "pay for a full human translation".

Real-world examples

The Georgetown-IBM demonstration (January 1954)

On 7 January 1954, at IBM's headquarters in New York, linguists from Georgetown University and engineers from IBM gave the first public demonstration of a computer translating one language into another. An operator who didn't speak Russian punched Russian sentences onto cards, and IBM's 701 computer printed English translations, such as "We transmit thoughts by means of speech", at about two and a half lines a second. More than sixty sentences were translated, on subjects from chemistry to politics.

IBM's press release the next day called the result "an unqualified success", and Leon Dostert, the Georgetown linguist behind the project, predicted that within five, perhaps three, years electronic translation "in important functional areas of several languages may well be an accomplished fact". What the audience wasn't shown so clearly was how small the system was. The historian of machine translation John Hutchins later documented that it knew just 250 words and six grammar rules, and that the sentences had been carefully chosen to fit them.

The prediction missed by decades. In 1966 a US government advisory committee, known as ALPAC, concluded that machine translation was slower, less accurate and more expensive than human translation, and US funding for it fell away for years afterwards. The lesson has aged well: a hand-picked demonstration tells you very little about how a translation system handles your own text.

Google switches to neural translation (September 2016)

On 27 September 2016, Google announced that it had moved its Chinese-to-English translations, about 18 million a day at the time, to a new neural system called Google Neural Machine Translation. The old system translated a sentence in pieces. The new one, in the words of the Google Brain researchers Quoc Le and Mike Schuster, treated "the entire input sentence as a unit for translation". Bilingual human raters compared the old and new output side by side, and Google reported that the new system cut translation errors by between 55 and 85 percent on several major language pairs.

Others followed quickly. In August 2017 the German company behind the Linguee dictionary launched DeepL Translator, a neural system that the company said beat its bigger rivals in blind tests. On 6 July 2022, Meta announced NLLB-200 ("No Language Left Behind"), a single model that translates between 200 languages, including many African and Asian languages that commercial tools had largely ignored. Meta released the model and its test data freely for others to use. Within six years, machine translation had gone from a joke to something billions of people used without thinking about it, which is exactly why its remaining errors became a serious problem.

Hospital discharge instructions (2019)

When patients leave a hospital emergency department, they're given written instructions: which medicine to take, which to stop, and when to come back. Researchers at the University of California, San Francisco, led by Elaine Khoong, wanted to know whether Google Translate could safely translate these for patients who don't read English. They took 100 real, anonymised sets of instructions, 647 sentences in all, and translated them into Spanish and Chinese. Bilingual translators then translated the results back into English, and two doctors judged each sentence for accuracy and for the chance of harm.

The study, published in JAMA Internal Medicine in February 2019, found that 92 percent of sentences were accurate in Spanish and 81 percent in Chinese. Most errors were harmless, but 2 percent of the Spanish sentences and 8 percent of the Chinese ones contained mistakes that could have hurt a patient. In one, "Hold the kidney medicine until you have a chance to speak with your kidney doctor" came back from the Chinese as, in effect, "keep taking kidney medicine". The English word "hold" means pause in a medical instruction and keep in everyday speech, and the machine picked the wrong one. The researchers also found that spelling and grammar mistakes in the English original made errors more likely.

The authors concluded that machine translation could supplement the English instructions but not replace them, and recommended giving patients the English version too, with a warning that the translation might be inaccurate. Translation tools have improved a lot since 2019, but the lesson hasn't changed: a 92 percent accurate translation still means some sentences are wrong, and you only find them if someone checks.

Afghan asylum claims (2023)

In April 2023, the tech news site Rest of World reported on a problem raised by Respond Crisis Translation, a volunteer network of translators who help refugees and migrants. Afghan refugees in the United States were filing asylum applications that had passed through machine translation from Pashto and Dari, two languages with far less online text than English or Spanish. In one case the article described, an automated tool had changed a woman's "I" into "we" throughout her written statement. Because her written application no longer matched the story she had told in her first interviews, a judge rejected her claim.

The translators also described tools stumbling over things a human would handle easily: the words for some relatives in Dari dialects, military ranks that can be vital evidence for former soldiers who worked with US forces, regional dialects and everyday expressions translated word for word. Asylum decisions depend on a person's account being consistent, so a small pronoun error can look like a lie. It's a sharp example of two problems meeting: weaker AI quality in low-resource languages, and a high-stakes setting where nobody fluent checked the output.

Kindle Translate (November 2025)

On 6 November 2025, Amazon announced Kindle Translate, an AI translation service for authors who publish through Kindle Direct Publishing (KDP). It launched in beta, translating English to and from Spanish and German to English, and Amazon said a translated ebook could be ready within about 72 hours. By an update on 30 June 2026, it had added French, Italian and Portuguese. At the time of writing the service is free to authors, and Amazon says translations are automatically checked for accuracy before publication. An Amazon spokesperson told the industry site Slator that this check is itself done by AI, developed with language specialists. Authors can preview the translation or let it publish automatically, and the titles are labelled in the store so readers know what they're buying.

Reaction was split. Some self-published authors welcomed a way to reach readers in languages they could never have afforded before. Publishers and translators raised concerns about quality and about books reaching readers with no fluent human having read them. For an independent author, Kindle Translate is a real option, and its biggest risk is the option to publish automatically. It's still in beta, so languages, eligibility and terms may have changed; check Amazon's current KDP pages before relying on it.

An everyday scenario: a mystery goes to Madrid

This is an illustrative scenario, not a documented case. Imagine Priya, who writes cosy mysteries set in a Yorkshire village. Her series sells steadily in English, and she wants a Spanish edition.

She runs the first chapter through a chatbot with a careful prompt and reads the result with a Spanish-speaking friend. Most of it is good. But the vicar's dry understatement has turned into plain statements, and the humour has gone. A local phrase, "she's not backwards in coming forwards", has been translated literally and makes no sense. Her detective now addresses her own sister with the formal usted, which makes the family sound like strangers. And one clue, a pun on a character's name, can't survive translation at all, which means a small plot change.

Priya decides the AI draft is a good start, not a finished book. She hires a professional translator who works from Spain to post-edit it, giving them her style notes, a glossary of recurring names and places, and a list of every pun and clue. It costs less than a full translation, and it takes two readers instead of none to catch what a fluent reader would notice.

Try it yourself

This exercise tests an AI translation using back-translation, the same method the hospital study used. You translate a paragraph into another language, then translate it back in a fresh chat and compare. You don't need to speak the other language. You need a free chatbot and about twenty minutes.

  1. Choose a paragraph of your own writing. Pick something with personality: a joke, an idiom, a bit of dialogue or a phrase that's very local. A bland paragraph will translate well and teach you nothing.
  2. Translate it with a proper brief. Open a chatbot and paste this prompt, filling in the brackets:
Translate the text below from English into [Spanish].

Audience: [adult readers of light mystery novels in Spain].
Tone: keep the author's voice, which is [warm, dry and funny].
Register: [use the familiar "tĂș" between family members and "usted" with strangers].

Rules:
- Translate the meaning and the effect, not word for word.
- If an idiom or joke has no natural equivalent, choose the closest natural one and list it under "Notes".
- Don't add or leave out any information.
- After the translation, list every sentence you were unsure about and explain why.

Text:
[paste your paragraph]
  1. Read the notes, not just the translation. The list of uncertain sentences is often the most useful part. It shows you where the meaning is at risk.
  2. Back-translate in a new chat. Open a fresh conversation, so the AI can't remember your original, and paste:
Translate the text below into English as literally as possible. Don't smooth it out or improve it. I want to see exactly what it says.

[paste the translation]
  1. Compare the two English versions side by side. Look for changed meanings, missing sentences, added details, a joke that became a plain statement, and "I" turned into "we" or "he" into "she". Mark anything that matters.
  2. Try a second tool. Put the same paragraph through a dedicated translator such as Google Translate or DeepL, both free to try at the time of writing, and back-translate that too. Where two tools disagree is a good place to look for trouble.
  3. Ask the AI to check itself. Back in the first chat, paste:
Here is a literal back-translation of your translation. Compare it with my original, list every difference in meaning, tone or register, and suggest a fix for each.

[paste the back-translation]
  1. Decide what the paragraph needs. If it's an email to a supplier, what you have may be fine. If it's a chapter of your book, treat what you have as a draft for a fluent human to finish.

One warning: back-translation is a smoke detector, not a proof. It catches meaning errors well, but a translation can back-translate perfectly and still sound unnatural to a native reader. Only a fluent reader can tell you whether it reads well.

Common mistakes and misunderstandings

Publishing a book translation nobody fluent has read. This is the big one for authors. An AI translation can be consistent for a hundred pages and then garble a crucial clue or a character's name. Readers in that language will notice, and reviews are public. At minimum, have a fluent reader go through the whole book. For anything you care about, hire a professional post-editor.

Forgetting the rules of the platform. At the time of writing, Amazon's KDP content guidelines define AI-generated content as text, images or translations created by an AI tool, and they say it counts as AI-generated "even if you applied substantial edits afterwards". You must disclose that when you publish. Translations must also credit the translator and the original author. Check the current guidelines before you publish, and check your contracts too: if a publisher already holds the translation rights to your book, you may not be free to publish a translation yourself.

Assuming every language is equally good. Quality is highest where there's the most translated text to learn from, such as English with Spanish, French, German or Chinese. It drops for languages with less material online. A tool that's excellent from English to Spanish can be unreliable into Pashto or Igbo.

Translating words instead of meaning. Idioms, puns, sarcasm, rhymes and cultural references rarely survive a literal translation. "It's raining cats and dogs" needs the other language's own expression for heavy rain, not a sentence about pets. Ask the AI to translate the effect, and to tell you when it had to change something.

Ignoring register. Whether characters use the formal or familiar "you" tells a reader how they relate to each other. An AI that switches between them from chapter to chapter makes a story feel wrong even when every sentence is accurate. Decide the rules in advance and put them in the prompt, or in a system prompt if you're translating many chapters.

Translating chapter by chapter with no memory. Every new chat starts fresh, so a character's nickname, a place name or a made-up word can be translated differently each time. Keep a glossary of names and recurring terms and paste it in with every chapter.

Trusting fluent output. A smooth translation isn't necessarily a correct one. Modern tools are good enough that their mistakes read naturally. That's why the check matters more now, not less.

Using AI where a certified human is required. Many official settings, including immigration applications, court documents and some medical and legal paperwork, require a certified or sworn translator. Even where they don't, if a wrong word could hurt someone or cost them a case, a machine translation should never be the final version.

Pasting private material without thinking. Unpublished manuscripts, contracts and personal documents are all confidential. Check how the tool you're using handles what you paste in before you use it for anything sensitive. See Privacy.

Related elements