Check an AI translation of a work document for numbers, units, names, and meaning

01Why a fluent translation can still be wrong

An AI translation can read smoothly while changing an important fact. A decimal point may move, a unit may disappear, a product name may be translated unnecessarily, or a negative instruction may become positive. These errors are especially risky in technical, purchasing, scheduling, compliance, and operational documents because a small wording change can alter what someone is expected to do.

The example below is synthetic. The goal is not to judge literary quality. It is to verify whether information that must remain stable across languages has actually remained stable.

02Start with a small synthetic work note

Assume the English source contains a short instruction for a fictional test team. The target language is not important for this exercise; the verification process is the same.

The critical values are therefore 12, 35 N·m, 80 °C, October 14, 2026, Mina Park, and Alpha-X. The phrase 'Do not exceed' is also critical because reversing or weakening the negation changes the operating instruction.

03Ask the AI to preserve protected details

When requesting the translation, specify which items must remain unchanged or must not be interpreted creatively.

Even with a careful prompt, verification is still required. The prompt reduces risk; it does not prove that every critical detail survived the translation.

04Check numbers, units, names, and negations one by one

1. Numbers: compare every quantity, decimal, percentage, date component, model number, and revision number with the source.
2. Units: confirm that 35 N·m remains 35 N·m and 80 °C remains 80 °C unless conversion was explicitly requested.
3. Names: verify personal names, company names, project names, and product names against the source spelling.
4. Negations: search for source expressions such as not, never, do not, cannot, except, and without, then confirm that the translated sentence preserves the same restriction.
5. Obligation strength: distinguish must, should, may, and can. These words are not interchangeable in instructions.
6. Dates: check both the numeric date and its role. A deadline must not become a meeting date or estimated date.

05Use back-translation as a spot check, not as proof

A useful second check is to take selected translated sentences and ask the AI to translate them back into the source language without seeing the original. This is especially useful for negative instructions, commitments, deadlines, and sentences where the wording feels unusually different.

Suppose the back-translation returns 'Keep the temperature at 80 °C' instead of 'Do not exceed 80 °C.' That difference is a warning that the translated instruction should be rechecked. However, a matching back-translation still does not prove correctness because the same AI may reproduce its earlier interpretation.

06Use Python to compare protected tokens

A small script can help detect obvious changes to protected details. The following standard-library example reads source.txt and translation.txt, extracts numbers and selected protected strings, and writes a report to outputs/translation_check_report.txt. It does not modify either original file and stops if the outputs folder already exists.

from pathlib import Path
import re
import sys

SOURCE_PATH = Path("source.txt")
TRANSLATION_PATH = Path("translation.txt")
OUTPUT_DIR = Path("outputs")
REPORT_PATH = OUTPUT_DIR / "translation_check_report.txt"

if not SOURCE_PATH.is_file():
    sys.exit("Missing input file: source.txt")

if not TRANSLATION_PATH.is_file():
    sys.exit("Missing input file: translation.txt")

if OUTPUT_DIR.exists():
    sys.exit("Stop: outputs folder already exists. Remove or rename it manually first.")

source = SOURCE_PATH.read_text(encoding="utf-8")
translation = TRANSLATION_PATH.read_text(encoding="utf-8")

number_pattern = re.compile(r"(?<!\d)\d+(?:\.\d+)?(?!\d)")
source_numbers = number_pattern.findall(source)
translation_numbers = number_pattern.findall(translation)

protected_strings = [
    "N·m",
    "°C",
    "Mina Park",
    "Alpha-X",
]

lines = []
lines.append("AI TRANSLATION CHECK")
lines.append("====================")
lines.append("")
lines.append(f"Source numbers: {source_numbers}")
lines.append(f"Translation numbers: {translation_numbers}")

if source_numbers == translation_numbers:
    lines.append("PASS: number sequence matches")
else:
    lines.append("CHECK: number sequence differs")

lines.append("")
lines.append("Protected strings:")

for item in protected_strings:
    in_source = item in source
    in_translation = item in translation
    if in_source and in_translation:
        lines.append(f"PASS: {item}")
    elif in_source:
        lines.append(f"CHECK: missing or changed in translation -> {item}")

lines.append("")
lines.append("Manual checks still required:")
lines.append("- Verify negations and obligation strength sentence by sentence.")
lines.append("- Confirm that each number still refers to the same item.")
lines.append("- Check dates, names, units, and product names in context.")
lines.append("- Spot-check high-risk sentences with back-translation.")

OUTPUT_DIR.mkdir()
REPORT_PATH.write_text("\n".join(lines) + "\n", encoding="utf-8")
print(f"Created: {REPORT_PATH}")

Run the script with python translation_check.py after saving the two texts. Matching number lists are useful but not sufficient. The translation could contain the same numbers in the wrong sentences, so every flagged or critical item still needs contextual review.

07Common mistakes when reviewing AI translations

• Checking only whether the translated text sounds natural.
• Missing a decimal, minus sign, percentage symbol, or unit change.
• Allowing a product or project name to be translated as ordinary vocabulary.
• Failing to notice that not, never, except, or without disappeared.
• Treating must, should, may, and can as equivalent.
• Trusting back-translation as definitive proof because it resembles the source.
• Comparing only token presence without checking whether each token remains attached to the correct statement.

Automated checks are strongest at finding visible mismatches. They are much weaker at verifying whether a sentence preserves intent, scope, exceptions, responsibility, or legal effect. Those require contextual reading and, for high-stakes documents, appropriate subject-matter review.

08Use a final translation review checklist

1. Compare every number, date, percentage, model number, and revision number.
2. Confirm every unit and ensure no unrequested conversion occurred.
3. Check personal, company, project, and product names against authoritative spelling.
4. Review every negation and exception directly against the source sentence.
5. Check obligation words such as must, should, may, and can.
6. Back-translate a few high-risk sentences without showing the original.
7. Read the final translation in context rather than relying only on automated token checks.

The practical rule is to treat an AI translation as a draft whose critical facts must be verified independently. For contracts, safety instructions, regulatory documents, medical content, or other high-consequence material, use the organization's required professional review process rather than relying on a lightweight AI and script check.

Execution and verification record

ChatGPT Python tool · Python 3.x · standard library pathlib, re, sys

• The original script executed successfully and created outputs/translation_check_report.txt.
• With a valid Korean translation, the original regex extracted source numbers ['12', '35', '80', '14', '2026'] but translation numbers ['35', '80', '14'], demonstrating a false mismatch caused by Unicode word-boundary behavior.
• All four protected strings N·m, °C, Mina Park, and Alpha-X passed.
• After replacing the regex with (?<!\d)\d+(?:\.\d+)?(?!\d), both files produced ['12', '35', '80', '14', '2026'] and the report returned PASS: number sequence matches.
• source.txt and translation.txt remained byte-for-byte unchanged.
• A second run stopped because outputs already existed, so the prior report was not overwritten.
• Local recheck on 2026-09-21 with Python 3.12.14: the corrected example passed; source preservation and overwrite protection also passed.

References

The explanations and examples were written for this site. See the official sources below for the related behavior and concepts.

• Python pathlib documentation ↗ (New tab)
• Python re documentation ↗ (New tab)