Can You Use AI to Translate a Website or App? What Works and What Needs a Human
Yes. AI translation, whether from a machine translation engine or a large language model (LLM), is good enough to publish for a large share of website and app text, including help articles, product descriptions and interface labels. It is not good enough on its own for legal pages, marketing headlines, medical or safety content, or anything regulated; those need a qualified human translator or reviewer. The setup that works is AI drafts, a glossary, human review ranked by risk, and a process that re-translates text whenever the source changes.
Translation is also an engineering job, not only a language job. Text has to be pulled out of code and templates, stored per language, and displayed in layouts that survive longer words and right-to-left scripts. That work is part of website, web application and mobile app development, and it decides how cheap each additional language is to add later.
Where is AI translation good enough on its own?
AI translation is good enough without a full human rewrite where the text is factual, the volume is high and a mistake is cheap to correct. That covers support articles and FAQs, product descriptions across a large catalog, internal tools and admin screens, and user-generated content such as reviews, ideally labeled as machine translated.
Quality is not uniform across languages, so have a native speaker spot-check a sample in each language before you rely on it.
Which content still needs a human translator or reviewer?
A human is needed wherever a wrong word creates legal, financial, safety or reputational cost.
Legal pages
Terms, privacy policies, consent wording and contracts should be translated or reviewed by a translator with legal experience, and checked by a local lawyer where the stakes justify it. This is not legal advice.
Marketing headlines and slogans
Wordplay, tone and cultural references rarely carry over. Headlines need rewriting by a native copywriter rather than translating.
Medical, safety and regulated content
Dosage, warnings, instructions and financial disclosures need a qualified reviewer. An LLM can drop a qualifier or change a meaning while the sentence still reads fluently, which makes the error hard to notice.
What does Google say about machine-translated content in Search?
Google's documentation does not ban machine-translated pages by name; it targets pages produced at scale that give users little value. As of October 6, 2026, Google's spam policies define scaled content abuse as generating many pages for the primary purpose of manipulating search rankings rather than helping users, "no matter how it's created". One listed example is scraping content and passing it through automated transformations, including translating, where little value is provided to users.
Google's guidance on generative AI content adds that generating many pages without adding value for users may violate that policy, and tells site owners to focus on accuracy, quality and relevance. Our reading, not Google's wording: translating your own useful pages for a market you actually serve is a different activity from mass-producing unreviewed pages in dozens of languages.
Google also states that it uses the visible content of a page to determine its language, not the lang attribute or the URL, and recommends a different URL for each language version. In one sentence, hreflang annotations tell Google which pages are localized versions of the same content, and each language version must list itself and all the others.
How do you keep brand terms consistent and translations up to date?
You need two things: a glossary, and a record of which source text each translation came from.
Glossary and brand terms
List product names, feature names and terms that must never be translated, plus the approved translation of key terms in each language and the tone to use, such as formal or informal address. Supply it with every translation request, as a glossary file where the engine supports one or as instructions in the LLM prompt.
Source changes
Store every piece of text under a key, together with the version of the source it was translated from. When the English changes, the system flags that translation as stale, re-translates only that item and queues it for review. Without this, versions drift apart: as an illustrative example, the English refund page says 30 days while the Arabic page still says 14.
What layout problems does translation cause?
Translated text is usually longer than English, and right-to-left languages mirror the whole layout. The W3C, citing IBM guidelines, gives an average expansion of 200 to 300 percent for English text of up to 10 characters, and about 130 percent for text over 70 characters. Buttons, menus and tabs therefore break first. Use flexible widths and test with your longest language.
Arabic, which matters for UAE audiences, is written right to left. On the web, the HTML dir attribute sets the base text direction, and MDN recommends it over CSS because direction belongs to the content. Navigation, alignment and directional icons such as back arrows must flip too, and English brand names or numbers inside Arabic sentences need testing. Mobile apps face the same issues whichever approach you pick from native, Flutter or React Native.
How do you set up an AI translation workflow with human review?
Sort content by risk, let AI draft everything, and send only the risky tiers to people.
- Move all text out of code and templates into language files or a translation system, each item with a key and a note on where it appears.
- Tier the content. Tier 1 (legal, regulated, safety, headlines): human translation or full expert review. Tier 2 (landing pages, checkout, onboarding, emails): AI draft plus native review. Tier 3 (help content, bulk catalog, internal tools): AI with sample checks.
- Build the glossary and style notes for each language.
- Generate AI drafts, passing the glossary, the context and any character limits.
- Have reviewers check the text on the real screen, not in a spreadsheet.
- Publish each language on its own URL with hreflang, then test layouts.
- Re-run the process on stale items when the source changes, and record who approved what.
What are the common mistakes when translating a website or app with AI?
The most common mistakes are treating translation as a one-off task and forgetting text that lives outside the page body.
- Leaving emails, error messages, PDFs, text inside images, page titles and the checkout in English.
- Sending short labels without context, so "Book" or "Save" gets the wrong meaning.
- Letting the AI translate brand names or damage placeholders such as {name}.
- Redirecting visitors automatically by assumed language, which Google's international guidance advises against.
- Keeping source-market currency, dates, units and address formats.
- Launching a language with nobody able to answer customers in it.
What should you do next?
Start with one language and an inventory. Choose the market from your analytics and sales inquiries, list every place text appears, assign tiers, and run a pilot on a small set of pages with a native reviewer. Count the reviewer's corrections: few means you can widen the AI-only tier, many means the glossary, the prompt or the engine needs work first.
Quick answers
Can AI translate an entire website or app?
AI can draft all of it, and the drafts are usually publishable for support content, product descriptions and internal tools after sample checks. Legal pages, marketing headlines and medical, safety or regulated content need a qualified human translator or reviewer before publishing.
Does Google penalize machine-translated content?
As of October 6, 2026, Google's spam policies do not ban machine translation by name. They treat as scaled content abuse the generation of many pages mainly to manipulate rankings, including automatically translated scraped content that gives users little value.
What is hreflang used for?
Hreflang annotations tell Google which pages are localized versions of the same content, so it can show the right language version in search results. Each language version must list itself and every other version.
Why does a layout break after translation?
Translated text is usually longer than English, especially short labels: the W3C cites average expansion of 200 to 300 percent for English text of up to 10 characters. Right-to-left languages such as Arabic also require the layout to be mirrored.
How do you keep translations in step with the original?
Store each piece of text under a key along with the source version it was translated from. When the source changes, flag the translation as stale, re-translate only that item and send it for review.
Conclusion
AI makes translating a website or app far less laborious, but it does not remove the two parts that decide the outcome: human review where errors are costly, and software built to hold several languages, text directions and changing source content. Decide the tiers first, then choose tools.
Entrant Technologies builds websites, web applications, mobile apps and custom software. If you are planning a multilingual version of yours, contact us and we will take a look.