Translating with AI: when it's good enough, and how to check

Short answer

AI translation is good enough for the gist, and often for a short, plain message in a widely used language. It is not good enough on its own for medical, legal, money or safety text. You can check a translation in a language you do not read: translate it back in a fresh chat, check every number, date, name and "not" line by line, compare a second tool on the risky sentences, and ask a speaker to read the lines that matter. The routine below is our suggestion.

To write in a second language yourself, see Writing in a second language. This page is for text you cannot read.

What the studies show

It varies by language. In 2019, Google Translate rendered emergency department discharge instructions (647 sentences) accurately 92% of the time in Spanish and 81% in Chinese. Of the inaccurate sentences, 28% in Spanish and 40% in Chinese could have caused significant harm (Khoong et al., JAMA Internal Medicine, 2019). The authors said it "can supplement (not replace) written English instructions".

Newer tools are better, not safe. The same group retested with GPT-4 and Google Translate on 316 sentences, using tools as of late 2023. Sentence accuracy for GPT-4 was 97% in Spanish, 95% in Chinese and 89% in Russian; Google Translate scored 96%, 90% and 80%. Even so, 56% of GPT-4's Russian instruction sets had at least one error. Their advice: "Avoid machine translation (without professional translator review) for high-stakes communication in any language" (Kong et al., BMJ Quality & Safety, 2026).

In a second study, professional medical translators rated ChatGPT-4, Microsoft Copilot and Google Translate into Albanian, Brazilian Portuguese and Vietnamese. Scores were similar overall, Copilot and Google Translate made five potentially harmful errors between them, and the authors say "human oversight remains necessary" (Wu et al., WestJEM, 2026).

Fluent is not the same as correct. The 2024 machine translation benchmark (WMT) found the best English to Spanish systems "close to flawless", and titled its report "The LLM Era Is Here but MT Is Not Solved Yet". It warns that fluent output "can contain serious flaws" (Kocmi et al., WMT 2024).

Decide the stakes first

Officials say the same:

• US immigration. USCIS requires "a full English language translation which the translator has certified as complete and accurate", with the translator certifying they are competent (8 CFR 103.2(b)(3)).
• UK visas. "You must provide certified translations of any documents that are not in English or Welsh" (GOV.UK).
• US health care. Language help "must be provided free of charge". If a provider machine translates critical documents, a qualified human translator must review them (HHS Office for Civil Rights, December 2024).
• NHS. Interpreting and translation should be free at the point of delivery, a professional interpreter "should always be offered", and tools such as Google Translate "should be avoided in healthcare settings" (OHID migrant health guide, updated 2021).

So for a hospital, clinic or GP, ask for their interpreter. For a certified translator in the US, the ATA directory lists members and lets you filter by credential.

The routine

Remove names, addresses and account numbers you do not need before you paste (What not to paste).

Translate the text below from <SOURCE_LANGUAGE> to <TARGET_LANGUAGE>.
Reader: <READER>. Tone: <TONE>. Formality: <FORMALITY>.
Rules:
- Keep every name, number, date, time, amount and dose exactly as written.
- Keep every negation ("not", "do not", "never"). Do not soften or drop one.
- Use these fixed translations: <GLOSSARY>.
- Do not add or remove information.
After the translation, list every phrase that was ambiguous or hard to
translate, what you chose, and why.
---
<TEXT>

Replace the following:

• <SOURCE_LANGUAGE> and <TARGET_LANGUAGE>: such as English and Brazilian Portuguese. Name the variety where it matters.
• <READER>: such as my grandmother, 80, reads simple Portuguese.
• <TONE> and <FORMALITY>: such as warm and informal, family.
• <GLOSSARY>: terms with one agreed translation, or none.
• <TEXT>: your prepared text.

Then open a fresh chat, or a different tool, so it cannot see your original:

Translate this <TARGET_LANGUAGE> text into <SOURCE_LANGUAGE> as literally as
you can, sentence by sentence, numbered. Do not smooth or improve it.
Then list every number, date, name, amount, dose and negation you see.
---
<TRANSLATED_TEXT>

Put the back-translation next to your original. Mark any sentence where the meaning, a number or a "not" differs, and rewrite or simplify that sentence in the source and run it again.

Text you receive (a letter, a notice): ask for a literal translation first, then a plain summary, then a list of anything uncertain, and any deadline, amount or action it asks of you. If it asks you to act on money, housing, health or a legal matter, get a human translation before you act.

How to tell it worked

• The back-translation says the same thing as your original, sentence by sentence.
• Every number, date, name, dose and negation matches.
• Two tools agree on the risky sentences.
• For anything you send, a speaker read the key lines, or you kept the stakes low.
• The reader's reply, in their own words, shows they understood what you meant.

Sources

1. Khoong EC et al. Google Translate for Spanish and Chinese emergency department discharge instructions. JAMA Internal Medicine, 2019. Peer-reviewed study. Link
2. Kong M et al. Accuracy and safety of machine translation of patient-specific discharge instructions. BMJ Quality & Safety, 2026. Peer-reviewed study. Link
3. Wu E et al. AI-based translation tools for emergency department discharge instructions. Western Journal of Emergency Medicine, 2026. Journal article. Link
4. Kocmi T et al. Findings of the WMT24 General Machine Translation Shared Task, 2024. Benchmark (peer-reviewed workshop). Link
5. US Code of Federal Regulations, 8 CFR 103.2(b)(3). Regulation. Link
6. GOV.UK. Apply for a Standard Visitor visa. Government guidance. Link
7. HHS Office for Civil Rights. Dear colleague letter on Section 1557 language access, December 2024. Regulator guidance. Link
8. OHID. Language interpretation: migrant health guide, 2021. Government guidance. Link
9. American Translators Association. Language Services Directory. Professional body directory. Link

Changes

1. v1.010 Oct 2026first version