Machine Translation on the Job: Where It Breaks
Why this matters
A lead who types "do not open the valve until the gauge holds" into a translation app and texts the result to a Spanish-speaking helper has done something reasonable, and has no way to know whether the helper received the instruction or its opposite. Machine translation is genuinely good now, good enough that crews use it every day for scheduling, parts lists and small talk. Its failures are rare, but they cluster in exactly the places a field-service shop cannot afford them: the word "not", the number with a unit, the difference between "must" and "should", and the trade term that means something else in ordinary speech. This card teaches where it breaks, how to see the break, and how to write so it breaks less.
What the tool is actually doing
A translation app or a general-purpose chat assistant does not understand the instruction and then express it in another language. It produces the sentence in the target language that most plausibly corresponds to your sentence, based on patterns from an enormous amount of text. Two properties follow, and they are the whole problem.
The output is fluent whether or not it is correct. A wrong translation does not look broken. It reads as smooth, grammatical Spanish or Vietnamese or Polish, so the person receiving it has no signal that anything went wrong.
Neither end can check it. The writer cannot read the output, and the reader cannot see the source. An error lives in the one place nobody involved can look. That is different from a bad interpreter, where at least the interpreter knows both languages.
Accuracy has improved sharply and keeps improving, and the specific failures below are less frequent than they were a few years ago. That is current as of 2026, and the honest way to say it is that you cannot predict which message will be the one that fails. The defense is to write and check as if any of them might.
The four ways it breaks on a job site
1. Negation gets dropped or moves. "Do not X until Y" is one of the hardest structures to carry between languages, because many languages build "not until" differently. In Spanish the correct form keeps the "no" ("no abra hasta que"); lose it and the sentence becomes "open it until", which is a complete, fluent instruction to do the dangerous thing. Long sentences with a negative in the middle and double negatives ("don't forget not to...") fail most.
- Cue: the back-translation (below) reads as permission where you wrote a prohibition.
- Source fix: lead with the safe action in positive form. "Keep the valve closed. Open it only after the gauge holds for 10 minutes."
2. Numbers, units and formats get mangled. Compound units collapse: foot-pounds of torque become "pounds", which is a weight. Inches and millimeters swap or vanish. Many countries write the decimal point as a comma and group thousands with a period, so "1.5" and "1,500" can trade meanings for a reader schooled elsewhere. Dates in month-first order read as day-first: 3/4 is March 4 in the US and April 3 in much of the world.
- Cue: a number arrives without its unit, or with a different one.
- Source fix: write every number with its unit spelled out, avoid decimals where a whole number in a smaller unit works ("15 millimeters", not "1.5 centimeters"), and write dates with the month as a word.
3. Safety words get softened. "Must" becomes "should". "Never" becomes "do not usually". "Danger" becomes "be careful". The tool is choosing the most common phrasing, and everyday language is less absolute than safety language.
- Cue: the back-translation sounds like advice where you wrote a rule.
- Source fix: put the rule word first and alone: "RULE: the breaker stays off." Short and flat survives better than polite and long.
4. Trade terms are translated as their everyday meaning. A flare fitting becomes a flame. Bleeding a line becomes medical bleeding. A trip (of a breaker) becomes a journey. A "head" is a body part, a "nipple" is anatomy, a "shot" compressor was fired at. "Breaker" often comes back as the word for a wall switch. The tool has seen the ordinary meaning far more often than the trade one.
- Cue: a word in the back-translation that makes no sense in context.
- Source fix: use the term your crew actually uses and keep it the same every time. A shop glossary is where that list lives; see related: Building a Two-Language Shop Glossary.
The artifact: one message, three wrong ways
An install lead types three instructions for a helper finishing a gas appliance connection, runs them through a translation tool into Spanish, and, before sending, asks a bilingual tech from another crew to read the output and say it back in English. The block below is constructed to show the four failure types side by side; it is not captured from any particular tool.
SOURCE (typed by the lead)
Do not open the gas valve until the test
gauge holds for 10 minutes.
Torque the flare nut to 15 ft-lb.
The breaker must stay off.
MACHINE OUTPUT (Spanish, accents omitted)
Abra la valvula de gas hasta que el
manometro se mantenga 10 minutos.
Apriete la tuerca de llamarada a 15 libras.
El interruptor deberia quedar apagado.
BACK-TRANSLATION (bilingual tech, aloud)
Open the gas valve until the gauge holds
10 minutes.
Tighten the flame nut to 15 pounds.
The switch should stay off.
Every one of the three instructions came back wrong, each in a different way:
- Instruction 1 lost its "not". It now tells the helper to open the gas valve during the test. This is the one that hurts someone.
- Instruction 2 carries two failures: "flare nut" became a flame nut, and the torque unit "ft-lb" collapsed to "pounds", which a helper could read as a weight or ignore.
- Instruction 3 carries two more: "must" softened to "should", and "breaker" became "switch", so a helper might turn off a wall switch and leave the circuit live.
The rewrite. The lead rewrites the source in short positive sentences: "Keep the gas valve closed. Watch the gauge for 10 minutes. If the gauge holds, open the valve." Then "Tighten the flare nut with the torque wrench to 15 foot-pounds." Then "RULE: the breaker in the panel stays off." Run again and read back by the same bilingual tech, all three come back with the meaning intact, including the unit and the word "panel", which pins "breaker" to the right device.
What decided it. The check cost the bilingual tech about two minutes. Without it, the only thing standing between the first instruction and an open gas valve during a leak test was the helper's own training, which is a control, but not one the message should be relying on.
Back-translation: what it catches and what it does not
Back-translation means taking the translated text and turning it back into English, by a different person or method than the one that produced it, then comparing that to the original. It is the cheapest check available, and it catches dropped negatives, lost units and softened rules reliably, because those changes survive the return trip.
Its limits matter as much as its strengths:
- Running the output back through the same tool proves little. The tool tends to undo its own error, so "open until" can come back as "do not open until" and look fine. Use a person, or at minimum a different tool, and treat a different tool as weaker than a person.
- It flags false alarms. Had the tool written "llave de gas", a common Spanish name for a gas shutoff, a literal back-translation could return "gas key", because llave means both key and valve. That is a back-translation artifact, not an error. The bilingual reader is what tells the two apart.
- It cannot judge register or region. A sentence can be accurate and still read as stiff, rude, or full of words a helper from a different country does not use. See related: Regional Variation in Trade Vocabulary.
Where it is acceptable and where it is not
| Use | Machine translation alone | With a bilingual read-back | Needs professional translation |
|---|---|---|---|
| Schedule, address, parts pickup | Acceptable | Better | No |
| Daily job instructions with no hazard | Risky | Acceptable | No |
| Any instruction touching gas, electricity, height, pressure or chemicals | No | Minimum | For anything reused |
| SOPs, checklists, training material | No | No | Yes |
| Contracts, estimates, wage notices, policies | No | No | Yes |
| Labels and signs | No | No | Yes, plus field test |
The bottom rows belong to siblings that own them. Written documents a shop reuses go through a real translation process; see related: Translating SOPs and Checklists for the Crew. Contracts carry their own legal layer; see related: Contracts in the Language the Deal Was Made In. Required safety training has to be delivered in a language the worker understands, which a pasted translation does not demonstrate; see related: Safety Training in a Language the Crew Understands.
Spoken translation in the field compounds the error
Voice translation adds a second machine step in front of the first: speech recognition turns what you said into text, then translation turns that text into another language. A running compressor, an accent, or a trade word the recognizer does not know corrupts the text before translation starts, and the translation then renders the corrupted text fluently. On a loud site, type rather than speak, keep it to one short sentence, and look at the recognized text on screen before you let it translate. The customer-side version of this, using photos and apps with homeowners, is covered elsewhere; see related: Using Photos, Gestures and Translation Tools Without Losing Accuracy.
References
- See related: Building a Two-Language Shop Glossary
- See related: Translating SOPs and Checklists for the Crew
- See related: Contracts in the Language the Deal Was Made In
- See related: Safety Training in a Language the Crew Understands
- See related: Regional Variation in Trade Vocabulary