Which Content Types Should Never Go Through Machine Translation

A procurement manager ran a 120-page supplier contract through DeepL, thought the output looked clean, and sent it without review. One clause in the original read: “the buyer does not have the right to terminate except in cases of material breach.” The MT version: “the buyer has the right to terminate in cases of material breach.” One dropped negation, one completely flipped clause. Seven months of arbitration followed.

The machine got 99% of the contract right. In legal documents, that remaining 1% can be everything.

Why “98% accuracy” is a dangerous metric for certain content¶

Machine translation has gotten genuinely good at certain tasks. Translating product descriptions, summarizing a news article, getting the gist of an informal email - tools like DeepL or Google Translate handle these reasonably well.

But “good enough” means different things for different content categories. For a blog post, 95% accuracy is fine. For a drug dosage instruction, 95% accuracy means one in twenty patients gets the wrong dose.

The problem isn’t just that MT makes mistakes - all translators do. The problem is the type of mistakes MT makes:

• Dropped negations. “Shall not” silently becomes “shall” in a complex legal clause.
• Hallucinated specifics. LLM-based tools can add clause numbers, dates, or amounts that weren’t in the source. The text looks coherent and credible.
• Literal idioms. A figurative phrase gets translated word-for-word and becomes meaningless or offensive.
• Jurisdiction-blind terminology. Legal and medical terms often have jurisdiction-specific meanings that MT can’t infer from context alone.

The question to ask before any translation project: what happens if 2% of this text is wrong in an invisible way? If the answer is “nothing much,” MT is probably fine. If the answer is “someone gets hurt,” “the contract is void,” or “the brand campaign becomes a meme for the wrong reasons” - keep reading.

Legal documents: where a dropped “not” can cost everything¶

Legal documents are the clearest category where raw MT output is never acceptable as a final product.

The reason isn’t just that legal language is complex - it’s that legal language is precise by design. Every word in a contract, court filing, or patent exists to prevent exactly the kind of ambiguity that MT introduces.

Three specific failure modes in legal MT:

Terminology that doesn’t map directly. Legal concepts like “consideration,” “estoppel,” “force majeure,” or “tortious interference” have no exact equivalents in many legal systems. MT typically either borrows the English term unchanged, translates it literally (which is wrong), or substitutes a near-synonym that carries different legal weight. The German “Haftung” can mean “liability,” “responsibility,” or “obligation” depending on context - getting it wrong changes what a party is actually required to do.

Jurisdiction-blind phrasing. A clause written for German law doesn’t read the same way under US or Chinese law, even if “correctly” translated. MT has no concept of which legal system applies, so it can’t flag or adjust for these differences.

The negation problem. MT systems learn statistical patterns from training data. In most sentences, negations are handled consistently. But in complex legal clauses with multiple subconditions, “shall not” can quietly become “shall” - and this is essentially invisible to anyone who doesn’t compare the output sentence-by-sentence to the original.

As JK Translate’s analysis of legal MT risks puts it:

Mistranslating a single “not” or “shall” can completely invalidate a legal agreement.

In Germany, France, Argentina, and most EU countries, court-submitted documents must be translated by a certified sworn translator anyway - machine output is inadmissible by definition. But the risk of using MT for business contracts and internal legal review (where formal certification isn’t required) is just as real and often underestimated.

For legal content: sworn human translator for anything going to court or a notary. Some firms do use MTPE for high-volume legal documentation, but “post-editing” in that context means a qualified legal translator reviewing every sentence against the original - not a surface-level skim.

If you’re interested in how AI can hallucinate in legal translation specifically, we covered that in detail in a separate article on AI hallucinations in legal translation.

Medical and clinical content: patient safety doesn’t tolerate error rates¶

In medical translation, “close enough” isn’t a category. The translation is either accurate or it isn’t - and the consequences of the latter can be physical.

As documented in a Translators USA analysis of MT in healthcare:

A medication dosage instruction was machine-translated with “once daily” rendered as “once weekly.” The patient received a 700% overdose.

Specific areas where MT breaks down in medical content:

Dosage instructions and prescriptions. The difference between “once daily,” “twice daily,” and “once weekly” is obvious in plain language - but non-standard phrasing or unusual sentence structure can cause MT to conflate these. The consequences range from ineffective treatment to overdose.

Informed consent forms. These are legally required to accurately convey risks, benefits, and alternatives to a patient. If MT drops a conditional (“may cause” becomes “causes”) or a frequency qualifier (“rarely” becomes “commonly”), the patient’s ability to give genuine informed consent is compromised - and so is the legal validity of the form.

Clinical trial protocols. These are binding documents that define exactly how a study is conducted. Errors affect data integrity and regulatory compliance. Both the FDA and EMA require certified human translation for clinical trial documents submitted for regulatory approval.

Medical device manuals. “Clearance” in FDA regulatory language means approval. Translated literally into some languages, it means “empty space” or “physical removal.” A technician reading a mistranslated manual for a surgical device isn’t a theoretical risk.

The American Translators Association notes that machine translation “cannot pick up context” the way a human expert can, making it unsuitable for any communication where the consequences of misunderstanding are serious.

Marketing, advertising, and brand communications: MT can’t transcreate¶

Marketing content fails at machine translation for a different reason than legal and medical - it’s not about precision, it’s about resonance.

Marketing copy works because it connects with how people in a specific culture think, feel, and use language. That means idioms, cultural references, humor, wordplay, and emotional register - exactly what MT handles least reliably.

The classic examples are documented and real. As ALTA Language Services summarizes from its analysis of major brand failures:

KFC’s “Finger-lickin’ good” became “eat your fingers off” in an early Chinese campaign. Pepsi’s “Come alive with the Pepsi Generation” became “Pepsi brings your ancestors back from the grave” in Mandarin.

These aren’t just awkward - they signal to local audiences that the brand doesn’t understand their culture.

The structural problem: marketing translation done properly isn’t translation at all. It’s transcreation - starting from the intended emotional effect and recreating it in the target language, rather than converting source words. A skilled transcreator might write something completely different from the original and hit exactly the right emotional note. MT can only translate; it has no concept of intended effect.

This applies to: - Slogans and taglines - a three-word slogan might rely entirely on a phonetic pun or cultural reference with no equivalent in the target language. - Social media copy - tone, humor, and current cultural references don’t travel through statistical models. - Product naming - “Nova” (no va = “doesn’t go” in Spanish) is the textbook example of what happens when naming goes unreviewed. - Campaign creative - if your ad relies on a local celebrity, a cultural moment, or a specific type of humor, MT won’t catch what it doesn’t understand.

The minimum viable approach for marketing: a native-speaking copywriter or transcreator, not a translator using MT as a shortcut.

Literary and creative writing: the voice doesn’t survive¶

Literary translation is where the gap between MT and human ability is most visible - and most studied.

A novel’s voice is built from word choice, rhythm, sentence length variation, character-specific speech patterns, and layers of cultural context. These don’t survive statistical pattern-matching.

Poetry is the extreme case. A 2025 study in IEEE Transactions on Computational Linguistics found that state-of-the-art models misread enjambment - the continuation of a sentence across a line break - in 78% of test poems. MT treats line breaks as sentence boundaries, destroying the rhythm the poet built.

Beyond mechanics, poetry depends on: - Intentional ambiguity. A word chosen because it carries three meanings simultaneously - MT must pick one, usually the most statistically common. - Sound and rhythm. Rhyme schemes, meter, assonance - these don’t map across languages and MT makes no attempt to recreate them. - Cultural resonance. References to local landscapes, folk traditions, or historical events require cultural knowledge that MT doesn’t have.

A 2026 analysis from TechXplore frames the problem clearly:

“Only poems can translate poems” - translating poetry requires writing new poetry, not converting words.

The same logic extends to literary fiction where voice matters (which is most literary fiction), plays where dialogue rhythm is essential, screenplays, and song lyrics that need to fit a melody.

Religious and ceremonial texts: precision with centuries of context¶

Religious texts present a specific challenge: precise doctrinal language where the right word choice has been debated by scholars for centuries and often has no direct equivalent in the target language.

Biblical Greek has multiple words for “love” - eros, philia, storge, agape - that all collapse to one word in most modern languages. Choosing which word to use in a given passage carries theological weight. MT doesn’t know this distinction exists.

The same applies to: - Liturgical texts - highly formalized, often using archaic register that MT handles poorly. - Sacred legal texts - halakha, sharia, and canon law have terminological systems developed over centuries with cross-referenced definitions that MT can’t track. - Ceremony and ritual texts - wedding ceremonies, funeral prayers, initiation texts - the exact phrasing matters to the community in ways that vary by denomination, tradition, and region.

Religious communities translating texts into new languages typically use a team of theological and linguistic experts working together over extended periods. MT can’t substitute for that process.

Safety-critical technical content: warnings are not a statistical exercise¶

Technical documentation for hazardous equipment or chemicals sits in a middle ground - most of a large manual can be handled efficiently with MTPE, but safety warnings are a different category entirely.

“Do not operate while safety guard is removed” - dropped negation - someone loses a hand.

“Mix only with water” on a chemical label - MT substitutes a similar-sounding word - dangerous reaction.

The UN Globally Harmonized System (GHS) for chemical classification and labeling requires certified human translation for all hazard statements and precautionary phrases for exactly this reason: a 1% error rate is unacceptable when the consequence is physical harm.

The rest of the same manual - part numbers, maintenance schedules, specifications - can go through MTPE efficiently and cost-effectively. The safety section cannot.

What machine translation actually does well¶

None of the above means MT is the wrong tool across the board. Post-edited machine translation represented nearly 46% of professional translation workflows in 2024, according to CSA Research data cited by XTM Cloud - because it genuinely works for the right content.

MT works well, often with MTPE, for:

The pattern is clear: content where accuracy is largely binary (these specs are right or wrong), where cultural resonance isn’t required, and where an error is easy to spot on review - that’s where MT adds real value without adding risk.

Content where a single word changes meaning, where cultural fit matters, or where errors have physical, legal, or financial consequences - human translators remain the only reliable option. Understanding when to use MT vs. MTPE vs. human translation is one of the most practical decisions a translation buyer or project manager can make.

FAQ¶

Can machine translation be used as a first draft for legal documents?¶

Some firms use MT as a starting point for human post-editing. But “post-editing” for legal content means reviewing every sentence against the original - the time savings vs. translating from scratch are often smaller than expected. For any document going to court, a notary, or a government agency, certified human translation is legally required in most jurisdictions regardless of how good the MT output looks.

What is MTPE and when does it make sense?¶

MTPE (machine translation post-editing) is a workflow where a qualified translator reviews and corrects MT output before delivery. It works well for high-volume functional content (product descriptions, FAQ, internal docs) where the correction burden is manageable. For legal, medical, or literary content, the correction burden is so high that MTPE often takes nearly as long as a fresh human translation.

Is Google Translate safe for business documents?¶

For getting the gist of an informal internal email - yes. For anything that will be published, filed with an authority, sent to a client, or used in a legal or medical context - no. Free MT tools also send document content to their servers, which raises confidentiality concerns for sensitive business documents.

Which content types work best with machine translation?¶

Product catalogs, FAQ pages, non-safety-critical technical manual sections, internal memos, and standard factual news content work well with MT or MTPE. The criteria: predictable sentence structure, factual content where accuracy is largely binary, and a human review step before anything is published or filed.

How do I decide if my content is safe for MT?¶

Two questions: (1) What happens if 2% of this translation is wrong in a way I don’t catch on review? If nothing significant - MT is probably fine. (2) Does this content require legal precision, patient safety instructions, cultural resonance, or certified output? If yes to any of those - plan for human translation, with MT only as a possible first-draft tool with thorough human review.

Sources¶

1. The Risks of Machine Translation in High-Stakes Legal Documents - JK Translate
2. When Not to Use Google Translate: The Critical Risks of Machine Translation in 2026 - Translators USA
3. Using AI for Language Translation: Context is Everything - American Translators Association
4. Why AI Struggles to Translate Poetry - Archyde
5. The art of literary translation exposes the limits of AI - TechXplore
6. 10 Marketing Translation Fails That Impacted High-Profile Companies - ALTA Language Services
7. UN Globally Harmonized System (GHS) for chemical classification and labeling
8. Content Translation: The Crucial 2025 Guide - XTM Cloud