Checking AI Translation Against the Source

The dangerous thing about automated translation is not that it produces gibberish. It is that it produces prose so fluent that no one thinks to check it. A Tamil sale deed passes through OCR, drops a negation during extraction, and comes back as clean English asserting the opposite of what the document says. The sentence reads well, so it survives review. On a matter where that clause decides liability, fluency has done more damage than a garbled output ever could.

Check the extraction before the translation

Two things happen between the scanned page and the English draft: the text is read off the image, and then it is translated. Most errors that matter are born in the first step. If OCR missed a line in Devanagari or misread a Kannada conjunct, the translation is faithfully rendering the wrong words. So begin by comparing the extracted text against the source image, script by script, before you form any view of the English. Correcting only the English sentence leaves the same defect in every later use of the file.

Build a deliberate test sample

Do not verify by skimming the whole document. Choose a short sample that concentrates the risk and read it closely:

• A negative or qualified statement, where a dropped "not" or "except" flips the meaning.
• A date, an amount and an identity number, which OCR distorts silently.
• A passage with clear legal effect, such as an operative direction or a covenant.
• A sentence that reads a little too smoothly to have come from a rough scan.

Watch the failure that reads well

The characteristic machine-translation error in legal work is a term of art rendered by its everyday sense. Nishedh becomes "ban" where the record means a specific prohibitory order; a word for "surety" arrives as "guarantee"; a procedural status is softened into ordinary description. Each reads naturally and each is wrong in a way that only a reader who knows both the language and the law will catch. Treat any confidently phrased legal conclusion as a prompt to return to the original phrase.

Trace the passages that carry weight

For the clauses the matter turns on, do not accept the translation on its own. Put the original phrase beside the English, and have a competent bilingual reader confirm that the rendering matches the source rather than a plausible paraphrase of it. Keeping source and translation visible together is what makes this check quick instead of a hunt through two separate files.

Record what you checked and what you did not

A verified translation is only useful if the next person knows how far the verification went. Note which passages were compared against the source, which corrections were made, and which sections remain a working translation that no one has confirmed. A reviewer should never have to assume the whole document was checked because part of it was.

Tools can support this rather than replace it. Juniorlawyer's legal translation keeps the source page and the translation aligned, so the comparison happens in view instead of across two windows, and corrections stay attached to the passage they fix.

The goal is not a translation that sounds authoritative. It is one whose every load-bearing sentence can be traced back to the words on the page, with the uncertain parts honestly marked. A fluent draft that no one can check against the source is a liability wearing the costume of finished work.