Paste a Spanish contract into DeepL or ChatGPT and you get a polished English draft in seconds. Most of it will be right, and that is exactly why these tools have become part of everyday legal work. If you already use them, or are thinking about it, that makes sense. Just make sure there is a safety net underneath, because in a contract one word can decide who pays, when, and on what terms.
The mistakes described below are not flaws of any particular product. They are patterns typical of machine translation and large language models in general, and they show up more or less often depending on the text, the settings and the version. Every example is invented for illustration. For the bigger picture, start with our guide to AI translation risks and who is responsible..
Why these errors are so easy to miss
Modern machine translation rarely produces clumsy sentences. The problem is the opposite: the output reads so naturally that nobody questions it. A text full of typos gets checked; a fluent one with a changed deadline does not. That is why the only reliable way to catch machine translation errors in contracts is to compare source and target sentence by sentence, rather than simply reading the English.
Nine common machine translation errors in contracts
1. Modal verbs that change legal force
A permission can turn into an obligation, or the other way round. In Spanish, podrá grants a right and deberá imposes a duty; the English needs to keep that distinction intact.
Illustrative example
- Source (ES): El Arrendatario podrá subarrendar la vivienda con autorización escrita del Arrendador.
- Machine output: The Tenant shall sublet the property with the Landlord's written authorisation.
- Correct: The Tenant may sublet the property with the Landlord's written consent.
How to spot it: list every modal in the source and check each one in the translation. The same check works in the other direction, where shall sometimes comes out as the much weaker podrá.
2. Negations that vanish
Exceptions and double conditions are slippery ground. Lose a single "not" and the clause says the opposite.
Illustrative example
- Source (ES): El Comprador no podrá reclamar daños y perjuicios salvo que acredite dolo del Vendedor.
- Machine output: The Buyer may claim damages unless the Seller's wilful misconduct is proven.
- Correct: The Buyer may not claim damages unless it proves wilful misconduct by the Seller.
How to spot it: highlight every negation and exception in the source (no, ni, salvo, sin, excepto) and confirm it survives with the same scope.
3. Missing sentences
This is the hardest one to see, because the translation still reads perfectly well. It is simply shorter than it should be.
Illustrative example
- Source (ES): La garantía tendrá una duración de dos años. Quedan excluidos los daños causados por un uso indebido.
- Machine output: The warranty shall last for two years.
- Correct: The warranty shall last for two years. Damage caused by misuse is excluded.
How to spot it: count the sentences in each paragraph on both sides, paying particular attention to the last one. A side-by-side view makes gaps obvious.
4. Amounts that do not match
Contracts often state sums in figures and in words. A translation can keep one and quietly alter the other, leaving you with an inconsistency in a signed document.
Illustrative example
- Source (ES): una penalización de 2.750 € (dos mil setecientos cincuenta euros)
- Machine output: a penalty of €2,750 (two thousand seven hundred euros)
- Correct: a penalty of €2,750 (two thousand seven hundred and fifty euros)
How to spot it: extract every figure, percentage and time limit from the source and tick them off one by one in the translation.
5. Number and date formats left unconverted
Spanish uses a full stop for thousands and a comma for decimals. Dates are day first, which a US reader will interpret differently.
Illustrative example
- Source (ES): : un anticipo de 1.250,50 € antes del 05/06/2027
- Machine output: an advance payment of €1.250,50 by 05/06/2027
- Correct: an advance payment of €1,250.50 by 5 June 2027
How to spot it: check every decimal figure and every numeric date. Writing the month out in full removes the ambiguity altogether.
6. Terminology that drifts
Engines work largely sentence by sentence and do not always remember what they chose three pages earlier. In a contract, each variant can look like a different party.
Illustrative example
- Source (ES): el Arrendatario (throughout)
- Machine output: "the Tenant" in clause 2, "the Lessee" in clause 5, "the Renter" in clause 9
- Correct: "the Tenant" throughout
How to spot it: list the defined terms and search for each one in the translation. Every term needs one equivalent, capitalised consistently.
7. Legal concepts translated word for word
Civil-law terminology does not map neatly onto English legal language. A literal rendering looks fine on the surface but carries no legal meaning.
Illustrative example
- Source (ES): El incumplimiento facultará a la otra parte para resolver el contrato y reclamar daños y perjuicios.
- Machine output: Breach shall entitle the other party to resolve the contract and claim damages and prejudices.
- Correct: Breach shall entitle the other party to terminate the contract and claim damages.
How to spot it: be suspicious of anything that sounds like a calque. If in doubt, check a bilingual legal glossary or ask a specialist legal translator.
8. Content nobody wrote
Language models tend to "complete" a text with what usually appears in similar contracts. This is often called hallucination, and in a contract it adds terms nobody negotiated.
Illustrative example
- Source (ES): Cualquiera de las partes podrá resolver el contrato con un preaviso de 30 días.
- Machine output: Either party may terminate the contract by giving 30 days' prior written notice, without penalty.
- Correct: Either party may terminate the contract by giving 30 days' notice.
How to spot it: check in reverse as well. Everything in the translation must be traceable to the source. Watch for extras such as "written", "reasonable" or "without penalty".
9. Broken cross-references
References to other clauses hold a contract together. Drop a sub-paragraph letter and the reference points somewhere else.
Illustrative example
- Source (ES): conforme a lo previsto en la cláusula 4.2, apartado b)
- Machine output: as provided in clause 4.2
- Correct: as provided in clause 4.2(b)
How to spot it: check each cross-reference against the actual numbering of the translated document, especially if it has been reformatted.
A practical review routine
You do not need to give up machine translation. You need a method. Here is the one we recommend:
- Be clear about the purpose. Understanding a draft is not the same as signing it or filing it with an authority. For official use, read first whether an AI translation is legally valid.
- Prepare your terminology. Before translating, list the defined terms and, if your tool allows it, load them as a glossary.
- Compare bilingually. Source and target side by side, sentence by sentence. Reading the English alone will not reveal omissions.
- Run targeted sweeps. Modals, negations, figures, dates and cross-references, each in a separate pass.
- Read the final text straight through. Check that it is consistent and reads like a contract drafted in English.
- If it will be signed, have a professional check it. A legal translator carrying out full post-editing (as described in ISO 18587) answers for every sentence, not just the overall impression.
If your team already translates contracts with machine translation, tell us how you review the output (info@tridiom.com). We would like to hear what works for you.
Frequently asked questions
Can DeepL or ChatGPT translate contracts accurately?
They usually produce fluent, good-quality drafts, which makes them genuinely useful for working quickly. But a contract demands complete accuracy, so the output should be checked sentence by sentence by someone who knows both languages and legal drafting.
Is reading the translation enough to catch the mistakes?
No. A read-through in the target language can catch awkward phrasing, but not omissions, lost negations or changed figures. For those you need to compare source and target, which the industry calls bilingual revision.
Will a glossary prevent these errors?
It helps a great deal with inconsistent terminology, but it will not fix modal verbs, omissions or added content. It is a useful first layer of control, not the only one.
What kind of review does an AI-translated contract need?
If the contract will be signed or have legal effect, full post-editing by a professional legal translator is advisable. If it must also be filed with an official body in Spain, you will probably need a sworn translation.
