Why AI Can’t Translate Your Legal Documents

Certified translation has a legal requirement AI cannot meet

The UK Home Office, UKVI, HM Courts and Tribunals Service, UK universities and professional registration bodies all require certified translation for documents not originally in English or Welsh. This is not a quality preference. It is a defined set of requirements set by the receiving authority.

• A certified translation must be complete, not summarised. 
• It must be accompanied by a signed statement of accuracy from the translator. 
• It must include the translator’s full name, contact details and the date of translation. 
• It must be produced by a professional translator, not by the document’s subject or a family member.

The signed declaration is the part AI cannot provide. Even the most accurate translation produced by ChatGPT or any other AI tool cannot be certified because no human translator has put their name to it. The declaration is a legal undertaking, not a quality mark. It exists so the translator can be held professionally accountable for the work.

Which documents need to be certified, and to what standard, depends on the receiving authority and the country the application is being made in. Getting that wrong holds up the whole application.

AI can hallucinate content that isn’t in the source

This is the risk most people underestimate. AI translation tools sometimes generate text that reads as a fluent translation but includes content that was not in the source. Names, dates, identification numbers, clauses, even whole sentences can be fabricated, omitted or quietly altered. The output looks coherent. The error is invisible without a line-by-line comparison against the original.

In legal translation, this is not a stylistic concern. Some of these errors get caught. A passport number that doesn't match the passport scan, a date of birth that contradicts the birth certificate, a name spelled inconsistently across documents, a careful caseworker can spot any of these and the application stalls until they're fixed. That's costly enough on its own.

The errors that pass through are the ones the receiving authority has no way to verify. A clause rendered in plausible but wrong legal terminology. A statutory reference quietly altered. A condition or obligation worded so the meaning shifts. The authority is reading the translated document, not comparing it line by line against an original they cannot read. By the time the consequence surfaces, the document is already in use.

The cost of an undetected error scales fast. The financial version is a property deed translated with the wrong figures or terminology, leaving a homeowner paying the wrong level of tax, with the bill or fine surfacing later when an authority cross-checks the paperwork. The worst version is the one you never see. A divorce settlement with a clause translated subtly wrong can quietly remove rights or income one party was legally entitled to, with the loss only emerging when it is too late to challenge.

Human translators make mistakes too. The difference is that a qualified human translator working to professional standards reviews their own output against the source and catches their errors. AI does not review itself in this sense. It produces the next plausible token. When the plausible token matches the source, the translation is accurate. When it does not, the output is fluent and wrong.

And prompting the AI to check its own work is not a real safeguard. Asked to review a translation against the source, large language models will often confirm their own output even where it is wrong. Challenged on the rendering, they will sometimes defend it more confidently than they produced it in the first place. The same model that generated the error is being asked to spot it, and it tends to stick to its earlier answer rather than correct it. A second pass through the same tool is not a second opinion. 

For documents where legal weight rests on every detail being faithful to the original, this is the risk that ends the conversation.

Legal concepts don’t map cleanly between jurisdictions

The third problem is harder to spot but more consequential when it goes wrong. Legal terminology is jurisdiction-specific. A French civil registry document, a German corporate filing or a Brazilian court judgement uses terms tied to that country’s legal system. Some have direct English equivalents. Many do not.

Examples within English itself make the point. A solicitor in the UK becomes an attorney in the US and a lawyer in Australia. Escrow is a standard mechanism in US property transactions. The UK uses a different process entirely and has no equivalent term. An estate agent in the UK becomes a realtor in the US, a title that carries its own licensing framework.

AI tools handle the familiar cases reasonably well. They make systematic errors with the unfamiliar ones, substituting the nearest English-looking word rather than flagging that no equivalent exists. They cannot tell the difference. A human legal translator with knowledge of both the source and target legal systems can. For legal work, that knowledge is not an upgrade. It is the work.

Where AI does fit in legal work

None of this disqualifies AI for translation work generally. Internal knowledge bases, training content, FAQs and high-volume product material all benefit from AI-assisted workflow with qualified human review. The speed and cost saving is real and the residual error rate sits within tolerance for the use case. Brightlines offers that workflow as a standard service.

There is a category of legal-adjacent work where AI-assisted translation has a real place. Foreign-language correspondence read to decide whether the matter needs lawyers. Discovery documents scanned to identify the ones worth full translation. A foreign contract reviewed for internal understanding before external counsel is engaged. This is sometimes called gist translation or for-information-only translation. Here the translation is for orientation, not for legal effect, and the document is never relied on as the legal record. If anything in it matters, it gets translated properly before it does any legal work. AI-assisted translation with qualified human review fits this category too.

How Brightlines handles legal and certified translation

We produce certified translations of official documents for UKVI, the Home Office, HM Courts and Tribunals Service, universities, and professional bodies. Each translation is completed by a qualified human translator and certified by our agency, with the statement of accuracy and full contact details required for acceptance.

Legal translation is its own discipline. We match each brief to a translator with the right legal training and experience for the subject matter and the target jurisdiction. Work is checked clause by clause against the source, so nothing is added, omitted, or altered. Where a source concept has no clean equivalent, we explain the gap rather than guessing.

A rejected submission means lost time, and sometimes a lost opportunity. If you have an official document that has to be accepted first time, send it to us with your deadline and tell us what the receiving authority requires. We'll confirm the cost and process up front, then match it to one of our translators with the right legal training for that jurisdiction. You'll get a certified translation built to be accepted, not corrected.