An AI translation can read fluently in the target language and still be wrong: a missed negative, a swapped figure, or a term that means something else in your field.

Fluency is not the same as fidelity. The pass that matters is a check against the source, not a pass for whether the English (or Japanese, or Spanish) "sounds nice."

1. Give the tool the right job

Be specific about audience and tone: a help article, a contract clause, a product UI, a casual email. Paste the source in full, not a summary of it. If you have a short glossary — product names, feature names, words you never translate — include it.

Do not ask the tool to "improve" the source at the same time as translating it. That is how extra claims and softer wording appear. Translate first. Edit the source separately if it needs editing.

2. Check the places translations usually break

Read the output with the source open beside it. Slow down on:

• Negatives and conditionals. "Unless", "except", "only if", and "may not" are easy to drop or invert, and the sentence still sounds fine.
• Numbers, dates, names and units. Confirm each one. Watch for decimal and date formats that changed silently.
• Defined terms. If a word is a product name, a legal term, or a heading in your glossary, it should stay consistent, or stay untranslated if that is your rule.
• Scope. "Some users" is not "all users." "We plan to" is not "we will."
• Tone and address. A formal notice should not suddenly sound like a marketing email, or the reverse.

If you do not read both languages well enough to do this, have someone who does. A fluent wrong sentence is harder to catch than a clumsy one.

3. Know when a machine draft is the wrong starting point

Use a human translator, or at least a specialist review, when the text creates an obligation: contracts, terms, safety instructions, medical or financial information, or anything you would not be willing to explain if a word were wrong. An AI draft can still help you get a sense of a long document. It should not be the version you sign or ship in those cases.

Also pause if the source itself is ambiguous. Ask the author, then translate. Do not let the tool pick a meaning quietly.

Before you send

• The output was checked against the full source, not only read in the target language.
• Negatives, numbers, names and defined terms all match.
• No extra promise or softening was added in translation.
• Glossary and product names are consistent.
• If the text has legal, safety or financial weight, a qualified human has reviewed it.

If you are choosing a tool, compare them on terminology control and how they handle your kind of text, not only on speed. The language translation category on Toolbit lists the main options, and this comparison of business translation tools is a practical starting point. Whichever you pick, keep the source open.

Takeaway

Read an AI translation against the source, not only for fluency. Check negatives, numbers and terms, and do not send obligation-heavy text without a human review.