Blog · Accuracy

9 Common Pitfalls of Machine Translation

person typing on typewriter

Machine translation gets most sentences right. Hallucinations, meaning translations that invent content the source never contained, run at 0.2% to 2% of output across published research on neural machine translation. A 2023 paper presented at the European Chapter of the Association for Computational Linguistics notes that stronger systems working from clean data should fall at the low end of that range.

Being right that often is what makes unverified machine translation risky. Nobody rereads a document that has been correct for 40 pages straight.

When the broken sentence does show up, it usually reads fluently in the target language. Catching it takes someone who can read both languages side by side, which is the one thing an unverified workflow leaves out.

Unverified machine translation means content goes straight from an engine to a reader, with no qualified human in between. That distinction matters, because the nine pitfalls below belong to the missing review step, not to the technology itself.

1. You can't tell which sentence is wrong

Machine translation errors don't announce themselves. The output is grammatical, confident, and fluent whether it's faithful to your source or not.

A human translator working in your language pair will flag a passage they're unsure about. An engine returns every sentence with the same certainty, so the one sentence that reversed your meaning looks exactly like the 500 that didn't.

For a blog post, that's an embarrassment. For a dosage instruction, a contract clause, or a letter telling a family what services their child will receive, the same error can injure a patient, change a contract's terms, or cost a child the services they need.

2. It loses the context around the words

Machines translate what's on the page. They don't know who's reading it, what was decided in the meeting last week, or which of a word's meanings you had in mind.

Puns, sarcasm, idioms, and industry shorthand all depend on context the engine can't see. A term like "positive result" means something very different in a lab report than it does in a customer survey, and nothing in the sentence itself tells the engine which one you meant.

A translator can ask. An engine picks the most statistically likely option and moves on.

3. It doesn't know your terminology

Every organization has words it uses in a particular way: product names, internal job titles, regulatory terms, the phrase your legal team insists on. A general-purpose engine has never seen any of them.

Left alone, it will translate the same term in three different ways across one document because it chooses sentence by sentence rather than working from an approved list. Inconsistent terminology is the error clients notice most, and it's the one that makes a document look unreviewed.

Fixing it requires two things a general-purpose engine doesn't have on its own: a glossary of your approved terms and a translation memory that stores your past approved translations. An engine trained on those inputs stops guessing and starts reusing language your organization has already signed off on.

4. It can't be creative

Marketing copy, taglines, and campaign concepts don't survive literal translation. They work because of rhythm, wordplay, and cultural reference, and all three break when you move them word for word into another language.

An engine will return something technically correct and completely flat. Worse, it can land badly in the target culture, because it has no way of knowing that the phrase it chose carries a meaning nobody intended. Creative work needs a person who can rebuild the line in the target language so it keeps its effect, rather than keeping its words.

5. You lose the time you thought you saved

Instant output feels like a schedule win. The time reappears later, when the translation is reviewed.

Someone still has to read the translation, decide whether it's accurate, and fix what isn't. When that work gets squeezed in between someone's actual deadlines, it moves at the speed of their calendar, not at the speed of the engine.

A planned review is faster than an unplanned one. Catching an error before publication costs an hour. Catching it after a customer does costs a reprint, a correction notice, or a conversation you'd rather not have.

6. Free translation costs more than it looks

The engine may be free, but the cleanup isn't.

That cleanup costs staff time and falls on people whose jobs are something else entirely. Pulling a bilingual colleague off their own work to fix an engine's output is a real expense that never shows up on an invoice, which is the same reason "free" employee translation costs more than you think.

It's also unpredictable. A clean document might need 20 minutes of review. A bad one might need a full retranslation, and you don't know which you have until someone qualified reads it.

7. Free tools can put your content somewhere you can't get it back

This pitfall has nothing to do with translation quality and everything to do with what happens to your file after you upload it.

Google's Terms of Service grant a license to "host, reproduce, distribute, communicate, and use your content," for purposes including "operating and improving the services" and "developing new technologies and services." Consumer tools are not covered by the confidentiality agreement a professional provider signs, and the free tier is not the enterprise product.

The exposure is already widespread. A TELUS Digital survey of 1,000 US employees at companies with 5,000 or more staff, fielded in January 2025, found that 57% of employees who use generative AI at work had entered sensitive or high-risk information into publicly available AI assistants. The information included personal data (31%), unreleased product and project details (29%), and customer information (21%). Separately, 68% of employees who use generative AI at work said they reach those assistants through personal accounts, which leaves their employer with no record of what was shared.

If your document contains patient information, student records, unreleased product specs, or anything under an NDA, the free translation box is a disclosure. Knowing how secure your translation data really is comes down to what a provider will tell you about storage, access, and retention.

8. You have no audit trail

An engine can't tell you who translated your document, what qualifications they held, or what changed between the first draft and the final file. It produces output, not evidence.

For regulated work, the evidence is part of the deliverable. Medical device documentation, clinical materials, legal filings, and special education records are all auditable, and "we ran it through a translation tool" is not an answer that holds up.

Certified providers are built around that requirement. ISO 17100:2015 sets the requirements for translation services, and ISO 18587:2017 covers post-editing of machine translation output. A provider certified to both can hand you a compliance report and a full audit trail showing who worked on your project and what qualified them to do it.

9. You may soon have to disclose AI-translated content

Disclosure rules for AI-generated content are arriving, and translation falls squarely within their scope.

Under Article 50 of the EU AI Act, organizations deploying AI systems that generate text "published with the purpose of informing the public on matters of public interest" must disclose that the text was artificially generated or manipulated. Those obligations apply from 2 August 2026. The regulation reaches any organization putting such a system to use in the EU market, so a US company with European customers, employees, or public-facing content may well be covered.

The exemption is the part worth noting. Disclosure isn't required where "the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content." A reviewed workflow answers the regulation. An unverified one doesn't.

Match the tool to the risk

Every pitfall above traces back to the same missing step: nobody qualified read the output before it reached the reader. So the question isn't whether to use an engine; it's how much review a given piece of content earns.

Three tiers cover most of what organizations translate:

  • Internal understanding. The gist of a supplier email or an inbound message. Raw machine translation is genuinely the right tool, and adding a human is a waste.
  • Operational content. Internal documentation and support articles, where meaning matters more than polish. This is where light post-editing fits: a linguist corrects what changes meaning and leaves the rest.
  • Published, regulated, or brand-facing content. Anything a customer, patient, regulator, or family will act on. Full post-editing under ISO 18587, or human translation, with the audit trail to prove it.

Which engine produces that first draft decides how much review the second and third tiers actually need. A public engine has never seen your terminology or your past work, so every project starts from zero and needs the same amount of correction every time. An engine trained on your glossary and translation memory starts from the language you've already approved, and the fixes your linguists make feed back in, so each project opens with a better draft than the one before it.

Most organizations already know which of their content sits in that third tier. The work is making sure it never quietly gets first-tier treatment because someone was in a hurry.

If you're weighing that for your own materials, it's worth seeing what a trained engine would change. Argo Translation builds custom AI translation engines around your glossary, your translation memory, and your industry, with a human linguist reviewing every project before it goes back to you.


Peter Argondizzo

Peter founded Argo Translation in 1995 after working as an Italian translator. He's spent 30 years helping businesses communicate across languages.