AI Legal Translation: What a Certified Translator Can Use and Disclose

A certified translator can use AI on legal documents only as a draft: the translator must review and correct every line, and the signed declaration is their personal statement that the result is complete and accurate. Whether AI must be disclosed depends on the receiving authority, the translator's professional code and the client contract. Tell clients in writing either way.

The reason is simple. "Certified" does not describe the translation, it describes who stands behind it. The tool is not a party to anything. A translator who signs over unchecked machine output has made a false statement in their own name, and the likeliest consequences are a rejected filing, a refund, and worse if the document was used in a dispute. The details differ between legal systems, as the table below shows, and none of this is legal advice.

Follow me on Instagram@sagnikteaches

Three systems, three meanings of "certified"

The word means different things in different places, so clients and translators talk past each other. The three below are the ones the question usually concerns, and the differences matter for how AI sits in the workflow.

Connect on LinkedInSagnik Bhattacharya

For the first row: US immigration filings are the usual example, and the rule in 8 CFR 103.2(b)(3) asks for a complete English translation with the translator's certification that the translation is complete and accurate and that they are competent to translate. It does not ask for ATA certification. The ATA's own page describes its certification as evidence of knowledge and skill, a credential obtained by a three-hour open-book proctored exam. The two ideas are routinely confused: a translator can give a "certified translation" without holding ATA certification, and holding it does not by itself certify any single document.

Subscribe on YouTube@sagnikteaches

In Germany, the civil procedure code (section 142(3) ZPO) lets a court require a translation made by a translator authorised or publicly appointed under the law of a federal state, and treats it as correct and complete once that translator certifies it, unless the contrary is shown. The provision says nothing about machine translation, and I could not find a published rule requiring AI use to be flagged on a sworn translation, so anyone working under that status should ask their authorising court or their professional association directly rather than assume silence means permission.

Always ask the receiving body for its requirements in writing before you start. Some accept a declaration on letterhead; some ask for a notary or an apostille on top, which is a separate step the translator cannot do. Check those before quoting a price and a turnaround.

What the declaration actually promises

The declaration is short and it is the thing at stake. A typical one reads like this (illustrative, adapt to the receiving body's wording):

I, [translator's name], confirm that I am competent to translate from [language] into [language], and that the attached translation of [document title, pages] is a complete and accurate translation of the original document.
Signed: [signature] | Date: [date] | Contact: [details] | Qualifications / membership: [if any]

Read it as five promises: you are competent in both languages, the translation is of this document, it is complete, it is accurate, and you will stand by it. AI can speed the typing and suggest terms. It cannot make any of the promises. That is the line between "AI-assisted" and "AI-translated", and a clean working rule follows from it: if you could not defend any sentence of the translation aloud to a judge or officer, do not sign.

Where AI helps and where it must stay out

Say a solo translator works in a legal niche with one or two agency clients. The sensible places to use AI are the ones that leave the translator in control of meaning.

• First drafts of routine, formulaic text. A vital record such as a birth or marriage certificate has a fixed structure. A machine draft followed by full verification can save time. Names, dates and places must be checked letter by letter against the original, because a swapped digit survives a fluent-looking sentence.
• Terminology lists. Ask the tool to extract the legal terms from a source contract and propose equivalents, then check each against a legal dictionary or a previous approved translation. This is where AI is least risky, because you review a list rather than trust prose. See how translators use AI to keep terminology consistent.
• A second reader. After you translate by hand, ask the tool to compare your version with the original and list possible omissions. The tool's list is a prompt to look again, and a false alarm costs you thirty seconds.
• Admin tasks. Clean a scanned table's layout, draft the cover email, convert dates to the target format, with your own review.

It should stay out of: documents where the client has prohibited it, texts that are in dispute or going to court where the exact wording could be contested, anything that falls under a confidentiality agreement you have not cleared, and any job where the receiving authority bans machine-assisted work.

A worked example with numbers

This is invented, to show how the time and the risk split. A 3,600-word share purchase agreement needs a certified translation. Translating from scratch at an assumed 450 words an hour is eight hours, plus 1.5 hours for review: 9.5 hours. Using an AI draft of the same document, the translator spends 0.5 hours preparing the text and glossary, then post-edits at an assumed 700 words an hour (just over 5 hours), then does a full second read (1.5 hours): about 7.1 hours in all. The saving, roughly 2.4 hours, is real but modest, and it disappears if the draft needs heavy rewriting. The real test of the saving is the error rate in a blind sample: before relying on the workflow, take 500 words, post-edit them, and count what you changed. If you changed more than roughly one word in three, the draft was not helping.

The figures are assumptions. Replace them with your own speeds from a trial, as in a translator's post-editing workflow.

What to disclose, and to whom

There are three audiences, and they need different things.

The client

Put it in your terms of engagement so it is agreed before the job. A clause could read:

Use of technology. I may use machine-translation or AI tools to produce a first draft or to check my work. Where a tool is used, I review and correct the whole text, and I am responsible for the final translation. I will not upload your documents to any online tool without your written permission. If you do not want any AI tool used on your documents, tell me before I start and I will work without it.

Many clients' own contracts forbid putting their information into external tools, so this clause protects you and gives them a say. A client who answers "no AI" has made the decision for you; keep a record of it.

The receiving authority

Some bodies ask whether machine translation was used, and some forbid it. Where there is a form or a stated requirement, answer it truthfully. Where there is none, the declaration covers the position as long as it is true. Do not add "no machine translation was used" if it was, even as a draft, unless the authority specifically defines that phrase to allow it.

Your professional body or insurer

Codes of conduct and professional indemnity terms may limit tools or confidentiality practices. Read the current code of any body you belong to, and ask your insurer whether AI-assisted work is covered as normal. Do not assume either, and keep the replies.

Confidentiality settings before any text is pasted

A legal document usually contains names, addresses, account numbers and sometimes health or family details. Before pasting any of it into an online tool, read the tool's current data page for how inputs are stored and whether they are used to improve the service, and check whether a business or API route has different terms from the free consumer one. If you cannot find a clear written answer, treat the tool as unsafe for client text. A safer approach for sensitive files is to remove or replace names and numbers in the draft stage and restore them yourself, or to work only in a tool your client has approved. For a wider look at the question, read whether it is safe to put client documents into AI translation tools.

Three realistic mistakes and how they show up

1. The silent number swap. A draft carries the source's number convention into the target text, so a figure written as 1.250 in a language that uses the full stop as a thousands separator (one thousand two hundred and fifty) reaches an English reader as one and a quarter, or a decimal point moves. It reads smoothly. Catch it by checking every figure, date and name in a second pass against the original, not against the draft.
2. The false-friend term. A legal term with a near-identical word in the other language but a different meaning (an "actual" that should be "current", a "company" that is a different legal form) is translated literally. Keep a glossary of approved renderings per client and run it as a find-and-check pass at the end.
3. The omitted stamp or marginal note. Machine tools extract text and ignore stamps, handwritten notes and headers. A certified translation must be complete, so list what the original contains that is not body text (seals, signatures, stamps, handwriting, illegible parts) and describe it in square brackets where your rules say to.

A short decision table by document type

This table is a starting point, not a rule. The agency or the authority may say otherwise, and the judge's expectations may differ from yours.

A marriage certificate, step by step

A short example shows how the pieces fit. A client sends a scan of a two-page civil marriage certificate and asks for a certified translation for an application to a public body.

1. Ask the receiving body first. The client checks and forwards the requirement: a full translation, the translator's signed statement, a copy of the original attached. No notary. You quote on that basis.
2. Read the terms line. The client has not banned AI tools, and your engagement clause says you may use them for a draft. The scan holds names, dates and registry numbers, so you decide not to upload it; you type the text yourself into your own document instead.
3. Draft. You translate it by hand, using the tool only to suggest equivalents for two registry terms you have not met, which you confirm in a dictionary of legal terms.
4. Verify. You read each name, date and number aloud against the scan, then once more in reverse order, last item first, which stops you reading what you expect. You note the circular registry stamp and a handwritten margin note and describe both in square brackets.
5. Declare. You attach the statement, sign, date and send the translation with a copy of the original.

Time: perhaps 45 minutes for a job like this, most of it checking. Nothing in the list needed AI to make it work, and the tool's contribution was two glossary suggestions. That is a normal result for short formulaic documents, and a reminder that the benefit grows with length and repetition.

A sample draft with the errors marked

Here is an invented clause and what a machine draft might do with it, to show what a careful reviewer looks for. Treat the draft as illustrative, not as any tool's output.

None of the four draft lines looks wrong on its own, which is why a translator should compare against the source line by line and not read the output for flow. If you do this on a sample of your own jobs, you will learn which kinds of error your tool makes most, and can aim your review there. A simple log of error types per job, kept for a month, is enough to show the pattern.

When a client brings their own AI translation

A common request is "I've already translated it with a tool, can you just certify it?" The answer is that you can offer a review-and-certify service, priced for a full review (see how translators price post-editing per word), but you cannot certify what you have not translated or checked to the standard of your declaration. Tell the client plainly: you will read every line against the original, correct it, and sign only the corrected version; if the draft is poor, you may translate afresh and the price will reflect that. Quote after you have looked at a page or two. Some authorities will not accept a certification of a client-produced text at all, so check that too.

A short reply you can reuse:

Hi [first name],
Thanks for sending the translation. I can check it against the original and certify it, but only after I have reviewed every line, and I will correct anything that is wrong. That is a full review, not a quick signature, so I will quote on the length of the document. If the draft needs a lot of correction I will tell you before I go further. Could you also confirm what the receiving body needs the statement to say?
[your name]

A job record that stands up later

If a translation is challenged months later, your record is your defence. Keep one line per job with the facts below, filled in here with an invented example:

Job: marriage certificate, 2 pages | Languages: [source] to [target] | Receiving body: [name], statement required, no notary
AI use: none on client text; tool used for 2 glossary suggestions (no client text uploaded)
Checks: names, dates, numbers read against original and in reverse; stamp and margin note described
Client told of tool policy: in engagement terms, accepted [date] | Signed by: [initials] on [date]

It takes two minutes. It also answers, without effort, the three questions anyone will ask: what did you use, what did you check, and what did the client agree to.

An approach for a small agency

If you run a team, the exposure rises because the person who signs may not be the person who used the tool. Write a one-page policy: which tools are approved and for which document types; no client text in unapproved tools; the signing translator reads every line of any AI-assisted job; the client is told in the contract; each job records whether AI was used. A job record could be one line in your project system: "AI draft: yes, approved tool; terminology list: yes; full second read by [initials]; signed by [initials]". The longer treatment is in does your translation business need an AI usage policy.

Finally, keep a short note of the date you last read each rule. Certification practice and AI terms both move, and a record that you checked in November is a better defence than a memory that you checked once.

Questions translators and clients ask

Does using AI make a translation uncertified?

Not automatically, because certification attaches to the translator's declaration, not to the tool. But the declaration says the translation is complete and accurate, and the translator carries that. If AI produced text the translator did not fully check, the declaration is not honest. Where an authority bans or restricts machine translation, that rule wins.

Do I have to tell clients I used AI?

Check your contract, your professional body's code and the receiving authority's rules, and err on the side of telling the client in your terms. Many clients also restrict it in their own contracts, so silence can be a breach even where no law requires disclosure.

Can I paste a client's contract into a free translation tool?

Usually you should not. Client documents are confidential, and consumer tools may retain or review inputs depending on settings and terms. Read the tool's current data terms, check what your client agreement allows, and get written permission before any client text leaves your own systems.

What should the translator's declaration say?

Typically: the translator's name and qualifications, the languages, a statement that the translation is a complete and accurate rendering of the attached original, the date and a signature, with contact details. Requirements vary by authority, so ask the receiving body for its wording before you prepare it.

Further reads

• How to Check AI Translations Before Customers See Them — A checking routine before any translation reaches a customer.
• How Small Translation Agencies Build an AI-Assisted Workflow — How small agencies build an AI-assisted workflow.
• DeepL vs ChatGPT for Translation: Which to Use for Each Client Document Type — Which tool for which type of client document.
• Is Machine Translation Good Enough for Client-Facing Documents? — When machine translation is good enough for clients.
• AI Translation vs Human Translators: What Documents Really Cost — What documents really cost by each route.
• AI-Powered Translation Management System: What a Small Agency Needs — A translation management system for a small agency.
• memoQ vs Trados for Small Translation Agencies: AI Features and Cost — memoQ and Trados priced for a small agency's seats, with their AI features, file compatibility and a two-week test before you commit.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ATA certification page; 8 CFR 103.2(b)(3) (Cornell LII copy of the regulation); GOV.UK guidance on certifying a document; section 142 of the German Code of Civil Procedure (ZPO) on gesetze-im-internet.de. Rules change and differ by receiving body; this is not legal advice.