Machine Translation for Official Documents: Where It Breaks

You have a birth certificate in front of you, your phone camera translates it in two seconds, and the result looks perfectly readable. Names, dates, the registrar's office, a parent's occupation — it all comes out in English. So the question is reasonable: if the machine already did the work, why pay someone to do it again?

The honest answer has two parts. Machine translation has genuinely gotten good, and on some kinds of text it is nearly indistinguishable from a careful human draft. Civil records are not that kind of text. They fail in specific, predictable places, and those places happen to be exactly what a reviewing officer looks at first. There is also a hard procedural wall that has nothing to do with quality. This guide covers both, and it ends with the cases where reaching for the app is the right call.

The procedural wall comes before the quality question

Set quality aside entirely for a moment. 8 CFR 103.2(b)(3) says that a document in a foreign language filed with USCIS must be accompanied by a full English translation, plus the translator's certification that the translation is complete and accurate and that they are competent to translate from the foreign language into English.

That certification is a statement made by a person. It is signed, it carries a name and a date, and it asserts competence. A piece of software has no competence to assert and no signature to give. Nobody at USCIS runs a detector to see whether text came from a machine — the barrier is simpler than that. There is no one to sign.

The rule does not ban machine translation. It requires a human certification, which amounts to the same thing in practice.

Worth noting what the rule does not say: it does not require a credentialed translator, it does not require an agency, and it does not require notarization. Those are separate questions, and the USCIS translation requirements page lays out what the regulation actually asks for. The one thing it does require is a named human taking responsibility, and that is the part no app provides.

Names get regularized into the wrong spelling

This is the failure that causes the most trouble, and it is built into how the technology works. A machine translation system is optimized to produce fluent, likely-looking output. When it encounters a personal name it does not recognize, it does not stop — it produces the spelling it has seen most often in training data.

So a surname written one way on the passport comes out another way on the translated birth certificate. The Arabic definite article in a name may be rendered as "Al", "al-", "El" or dropped. A Korean surname (이) might come back as Lee, Yi or Rhee depending on nothing the reader can control. Cyrillic transliteration has at least three competing systems in common use, so Georgiy and Heorhii can be the same man. Chinese given names get split, joined or reordered, sometimes within the same document.

A human translator handling civil records knows to do something a machine cannot: match the name to the identity document in the file, not to the most statistically probable English spelling. That judgment matters most in the languages where transliteration is unsettled, which is why Arabic to English translation and Russian records so often come back with names that do not line up against the passport.

The cost of getting it wrong is not stylistic. A name mismatch between a translated record and the passport reads as a discrepancy, and discrepancies invite questions.

Dates flip, and nothing warns you

Most of the world writes dates day-month-year. The United States writes month-day-year. For any day of the month above twelve, the ambiguity resolves itself. For days one through twelve, it does not, and the machine will pick a format without flagging that a choice was made.

A birth recorded as 05/03/2019 in Bogotá is March 5. Rendered as May 3 on a translated certificate, it contradicts every other document in the petition. Nothing in the output marks the moment where the guess happened.

Non-Gregorian calendars compound this. Records from Iran, Afghanistan, Ethiopia, Nepal, Thailand and Taiwan use their own era systems, and machine output tends to either copy the number straight through or convert it silently and sometimes wrongly. Chinese-language records write dates year-month-day (2019年3月5日), which at least cannot be misread — but the Chinese characters for year, month and day often survive into the English output untranslated.

A competent translator writes the date in a form that cannot be misread and, where the original uses a different calendar, shows the conversion. That is a judgment call, and judgment is the thing being purchased.

Seals, stamps and handwriting are simply missing

Point a translation app at a civil record and watch what it does with the parts that are not printed body text. It skips the round registrar's seal. It skips the handwritten marginal note added nine years after the birth was registered. It skips the fee stamp, the registrar's signature block, the security watermark and the reference number printed sideways in the margin.

Those elements are not decoration. On Latin American records, marginal notes (notas marginales) record later legal events — a legitimation, a correction, a recognition of paternity. On Eastern European records, the special marks field (особливі відмітки) carries exactly the kind of information a reviewer is looking for. A translation that omits them is not a complete translation, which is the specific word the regulation uses.

Handwriting is the second blind spot. Optical character recognition is built for printed type. Cursive in Spanish, Russian or Arabic, filled in by a clerk decades ago, produces either nothing or confident nonsense.

What a complete translation does instead:

• Describes each seal and stamp in a bracketed note, with any legible text translated
• Marks signatures as signatures rather than guessing at the name
• Renders handwritten entries, with an illegibility note where the original genuinely cannot be read
• Keeps the layout of the original so each element sits where it sits on the source
• Carries the reference and registration numbers across exactly, digit for digit

Institutional vocabulary has no machine equivalent

Civil records are full of terms that name an institution, an office or a legal status that has no American counterpart. Machine systems translate them literally, which produces English words in the right grammatical order that mean nothing to a US reader.

A Mexican civil registry office, a Brazilian notary registry (cartório), a Polish full copy of a record (odpis zupełny) as against the abridged copy (odpis skrócony), a Korean family relation certificate (가족관계증명서) as against the Basic Certificate (기본증명서) — each of these distinctions carries weight. The machine flattens them. It will give you "full extract" and "short extract" with no indication that one contains the information the filing needs and the other does not.

Academic records are worse. A grading scale, a credit hour convention and a degree title all have local meaning, and a literal rendering misleads rather than informs — which is why diploma and transcript translation is handled as a document type of its own rather than as plain text.

The pattern holds across document types. A machine translates words; a document translator translates a record, which means knowing what the form is, what fields it has, and which of them matter to the office reading it.

Where machine translation is genuinely the right tool

None of the above means you should never open the app. There are real uses, and pretending otherwise would be dishonest.

1. Triage. You have thirty pages from a foreign records office and need to know which three are the marriage certificate. Photograph them, skim the output, sort the pile. Nothing goes in a file.
2. Deciding whether a document is worth translating. Before you pay per page, it helps to know whether a document actually contains what you think it contains. Our pricing is per page of up to 250 words, so knowing which pages matter has direct value.
3. Reading correspondence. Emails from a relative, a records office reply, a hospital's instructions — anything that informs your decisions but never gets filed.
4. Understanding your own documents. If you want to know what your grandmother's death certificate says before anyone else reads it, read it. That is not a filing.
5. Draft work by a competent translator. Professionals do use machine output as a starting point and then rebuild it. The catch is that finding the errors takes the same language knowledge the machine was supposed to substitute for, so this is not a shortcut for someone who does not read the source language.

The line is clean: if the output goes into an official file, it needs a human behind it. If it only ever informs you, use whatever tool you like.

What this costs you to do properly

The gap people imagine between free and certified is usually larger than the real one. A certified translation is $24.95 per page, where a page means up to 250 words including numbers. Most civil records — a birth certificate, a marriage certificate, a single-page police record — are one page at that count. A one- to three-page document typically comes back within 24 hours, with same-day available when a deadline is tight.

What the price includes is the part the machine cannot supply: a human translator who reads the source language, formatting that mirrors the original so a reviewer can line the two up side by side, the bracketed treatment of seals and stamps, and a signed Certificate of Translation Accuracy naming the person who stands behind it. Revisions are included, which matters because the one thing you cannot do with machine output is go back and ask it to explain its reasoning.

If you want to check the scope before committing, you can upload the document and get the page count without paying.

Frequently asked questions

Can I use Google Translate to translate documents for USCIS?

You can use it to understand what a document says, but the output on its own does not meet 8 CFR 103.2(b)(3). The rule requires a complete English translation plus a signed certification from the translator stating the translation is complete and accurate and that they are competent to translate from that language into English. Software cannot sign that statement.

What if I run the document through a translation app and then fix the errors myself?

That is a legitimate way to work if you are a competent translator, and many professionals use machine output as a first draft. The problem is that catching the errors requires the language knowledge the machine was supposed to replace, and petitioners who are party to their own filing are not impartial reviewers of it.

Why does machine translation handle names and dates so badly on civil records?

Machine systems are trained to produce fluent text, so they regularize unfamiliar names into common spellings and reformat dates into the pattern they see most often. On a civil record, a name that does not match the passport or a date flipped from day-month to month-day is the kind of discrepancy a reviewing officer notices first.

Is there anything machine translation is genuinely good for in an immigration filing?

Yes. It is useful for triage — working out which page of a stack is the marriage record, checking whether a document is relevant before you pay to translate it, and reading correspondence that will never be filed. The line is whether the output goes into an official file.